AI & Mental Health — A Curated Collection

Last updated: (Sydney time)

This is a curated and openly non-neutral collection of resources gathered by Mark Anns. It reflects the view that AI is already reshaping the world, that attempts to slow its overall trajectory are unlikely to succeed, and that fragmented geopolitical responses will continue to produce uneven approaches to mental health, safety, and regulation.

Some of the links here are papers, others are talks, interviews, or audio discussions. They are shared to support thinking, not as neutral or complete answers. Items marked “Preprint by Mark Anns” are his own preprints. Each card is labelled by kind of document, how it was checked and, where relevant, who is behind it, so evidence, policy, advocacy and opinion can be told apart.

My position on AI and mental health

Public health, harm reduction, and global access: a personal statement by Mark Anns.

I write this as a clinician who works at the intersection of lifestyle medicine, digital health and AI governance, and who takes the World Health Organization’s definition of mental health seriously. From a public health point of view, every intervention carries potential benefits and harms, including the harm of leaving people with no support at all. Mental health, in WHO’s terms, is about our capacity to cope with life’s stresses, realise our abilities, participate in work and learning, and contribute to our communities, within the social and economic conditions we actually live in. Many people simply do not have access to formal mental health services. Their support comes from family, community leaders, religious figures, peers, or, increasingly, from digital tools. In that landscape, it is reasonable to ask whether carefully governed AI systems can offer a net benefit: whether they help some people cope a little better, understand their situation, practise skills, or stay connected to daily functioning more than they expose them to new risks.

In reality, most people will not seek out well designed, specialist digital mental health tools. They will use broad, general purpose AI chat engines trained on vast, largely unvetted datasets, each shaped by different commercial priorities, regulatory settings and cultural assumptions, with often hidden and unaccountable biases. How a chat engine built and governed in one context, for example under emerging Chinese rules that explicitly treat psychological harm and emotional dependency as first-order risks, frames mental health and acceptable responses may differ from how systems built in other regions do so, yet these tools are globally accessible and often indistinguishable to users. There is now an intense race to ensure these engines “do no harm,” particularly around suicidality, psychosis risk, emotional dependence, discrimination and misinformation, but realistically we can only aim to minimise foreseeable harms, not eliminate them entirely. In that context, the harm-reduction position that AI can be better than nothing, for some people in some conditions, is a defensible one. It recognises that people will use these systems anyway, often without safe local alternatives, and that our ethical task is to move them toward less harmful, more supportive patterns of use: encouraging evidence-informed tools where possible, strengthening guardrails and human-in-the-loop oversight in general engines, and being transparent about cultural bias, data limits and the need to check information with trusted human sources.

My position is deliberately shaped by a social-determinants and capabilities view of mental health. I am deliberately resisting models that relocate complex social harms into the individual and then offer narrow clinical fixes. This is closely related to harm-reduction debates in drug policy, where we accept that certain risks and behaviours will occur and develop interventions, such as needle and syringe programs or drug checking services, to reduce avoidable harm rather than pretend we can eliminate the behaviour. In the same way, we know people will seek help, information and “someone to talk to” about distress even when services are thin or absent. They will turn to unregulated apps, forums, influencers, or no one at all. In that setting, a carefully governed AI tool that offers evidence-informed information, practical coping strategies and space to reflect does not constitute full care, but it may offer something where there was previously nothing, and may modestly support everyday functioning for some users on the mental health spectrum, while we continue to push for stronger human, community and structural responses.

Much of the current professional debate about AI and mental health is shaped by clinical risk aversion. That posture is understandable, and it has a job to do. It also narrows the conversation. It tends to ask only whether AI should be used inside a therapy room, and to answer with caution about liability, boundaries and consent. Those are real questions, but they are not the only ones, and treating them as the only ones leaves out most of what is actually happening. A public-health lens asks different questions. Who is already using these systems, and why. What happens to people who have no access to formal services. How commercial and jurisdictional differences shape what these tools say to users in distress. What substitution looks like when funding decisions get made. Those questions are underrepresented in mainstream professional discussion, and one of the reasons I have written this statement is that I think they should not be.

At the same time, a public-health and harm-reduction frame insists that harms are real and must be taken seriously. AI systems can misinterpret distress or miss acute risk. They can reinforce stigma, over-individualise problems that are actually responses to poverty or injustice, reproduce cultural bias, and be used to justify cutting back on human services and social change. For people in crisis, those with severe or complex difficulties, and communities already harmed by technocratic “solutions,” these are not abstract concerns. I agree with colleagues and advocates who highlight these risks and I share their worry about substitution: using AI as a reason not to invest in human, relational care or action on social determinants. When I say the harm-reduction position that AI can be better than nothing is defensible, I mean it as a risk-managed claim, not an unqualified endorsement. Any use of AI in mental health must come with clear limits on what it is for, strong governance, transparent communication of risks and biases, explicit pathways to human support and escalation, and a firm commitment that AI tools will add to, not replace, community-based support and serious work on the social determinants of mental health.

This is also the lens that has shaped what is and is not in this collection. The resources below deliberately reach outside a purely clinical frame. They include global governance work, harm-reduction thinking, cultural and jurisdictional analyses, consumer and community voices, and critical work that questions the profession’s response as much as the technology. My own preprints are marked as such. None of this is a recommendation to use any specific tool. It is an argument that the debate we are currently having is too narrow, and that a wider one is the one worth having.

How to read this collection

The public health frame, the questions applied to every debate, and what the labels on each card mean.

This collection uses a public health frame. Mental health is understood as the World Health Organization defines it: a state of mental well-being that enables people to cope with the stresses of life, realise their abilities, learn and work well, and contribute to their community, existing on a complex continuum. Every use of AI carries potential benefits and potential harms, and both must be weighed against the realistic alternative, which for many people is no support at all.

The hub is organised in five parts: potential benefits, potential risks, the conditions that shape both, broad context that is not mental health specific, and the curator's preprints. Benefits come first on purpose. Public debate leans heavily towards harm, and benefits such as health literacy, planning, triage, training and resource development are rarely discussed.

Questions applied to every debate: 1. What is the actual question? 2. Where on the mental health continuum does it apply? 3. Where does the benefit or risk sit: one conversation, a product, a platform, a service system, or society? 4. Compared with what: usual care, a waitlist, or no care? 5. Who benefits, and who carries the risk? 6. What evidence counts, and has lived experience shaped it? 7. What response fits, and who has the power to act on it? 8. Whose cultural framing is built into the answer?

Each card has one home, where it is shown in full, and chips for every other debate it informs. Select a chip, or use the filters, to see every card for that debate wherever it lives. Continuum labels show where on the WHO continuum a resource applies: wellbeing and promotion; early distress and prevention; mild to moderate conditions; severe and complex conditions; and population and system (whole-population or service-level uses).

Because this field changes faster than formal research can keep up, the collection includes preprints, policy, position statements, advocacy, opinion and early signals alongside peer-reviewed evidence. Every card carries labels so you can tell them apart. What kind of document it is: research, evidence review, policy and regulation, guidance and standards, position statement, analysis and advocacy, opinion and commentary, practice signal, or tool and dataset. How it was checked: peer reviewed, preprint (not yet reviewed), or not reviewed. Who is behind it, shown only when it matters: developer research, industry funded, or advocacy organisation. A lived experience label marks work written by or with people with lived experience. Some cards add a short note on why the item matters now.

A verified status means the resource and its citation details were checked. It does not mean the claims in the resource are proven. Labels on older cards were assigned in October 2026 and are being checked over time.

What's been happening since the last update

Update of 27 October 2026: the main developments from late April to October 2026, and what was added to the hub.

This update, made on 27 October 2026, adds 118 cards covering research, policy, court cases and public debate from late April to October 2026. Six of them are from earlier in 2026 but were missing from the hub. There is also a new category, C2.1 Privacy and commercial use of personal data, because privacy kept coming up in every other debate and had no home of its own. The harm, benefit, feed, privacy and sovereignty debates moved closer together over these six months, and the paragraphs below summarise what happened.

Harm has moved into the courts and into state law. In the United States, eight state attorneys general sued OpenAI on 14 October 2026. The Garcia case against Character.AI, brought after the death of a 14-year-old, settled in mid-October. In Character.AI's bankruptcy, states are now asking for protection of children's chat histories, so acute harm and privacy have become one legal question. Common Sense Media tested ChatGPT's teen accounts with more than 4,000 prompts. It found no parental alerts for suicidal ideation and rated the product an unacceptable risk.

The line between chatbots and social media feeds is blurring. ChatGPT now shows advertising to free users in 89 countries, which moves the most widely used chatbot towards the same attention-and-advertising model that drives feeds. Rules for social media are also shifting from banning access to banning design features. California now restricts infinite scroll and autoplay for under-16s, and the EU's proposed KIDS Act pairs an under-15 account limit with safety-by-design duties. Australia's Digital Duty of Care bill reached exposure draft on 8 September. eSafety's three-month data shows under-16 account holding fell only from about 52% to 42%. Six months in, 70% of parents in one survey said their children still had active accounts.

Evidence on relationships with chatbots is becoming longer term, and regulators are starting to name dependency. A one-year study of Character.AI users linked sustained companion use with lower wellbeing, partly through less time with people in person. China's rules for human-like AI services came into force on 15 July 2026 and prohibit designs that foster emotional dependency.

Unclear regulation has its own costs. When the therapy chatbot Ash was withdrawn from the UK, 60% of former users surveyed said their access to support got worse, and the most common replacement was a general-purpose chatbot with no mental health safeguards. In Australia, the TGA is consulting on whether some digital mental health tools should stay outside its regulation, with submissions closing on 7 December 2026. The FDA has released a discussion paper on regulating generative AI in medical devices, and the UK's MHRA is setting up a community of practice for digital mental health developers.

Privacy and data sovereignty moved too. Australia released the second-stage Privacy Act exposure draft on 31 August 2026, and the Children's Online Privacy Code must be registered by 10 December 2026. A study of 25 popular mental health apps found that every one contained at least one tracker its privacy policy did not name. The Lowitja Institute's July review identifies gaps in Indigenous leadership and data governance across Australian AI and health policy, and new work from Aotearoa looks back on ten years of Māori data sovereignty.

In the profession, a 30-country survey found that just over half of psychotherapists now use generative AI in practice, mostly for treatment planning, administration and psychoeducation. The Australian College of Mental Health Nurses has set out consent expectations for AI scribes. In the broader context, early-career workers aged 22 to 25 in AI-exposed jobs show the first signs of reduced hiring. Data centre electricity use rose 17% in 2025. Australia is consulting on energy, water and community conditions for large data centres, and middle powers are debating how to keep access to AI without depending on one country.

New cards from this update are marked in their notes as added from the 27 October 2026 scan. Every card was checked against its source page for title, publisher and date. Preprints and sources with a developer, industry or advocacy interest are labelled.

Start with a question

Research questions

Questions people often ask about AI and mental health, each with a short overview and a curated set of cards from this hub. Six core questions come first; more are below. The overviews set out the debate rather than argue a position. My own view is in the position statement above.

How could AI be used for mental health and wellbeing?AI may support wellbeing, service planning and care, but its value depends on what it improves and for whom.

AI can be used across the mental health continuum, from making wellbeing information easier to understand to supporting services for people with significant illness. Reviews describe personalised learning and conversational tools as possible ways to make health education more engaging, while a trial tested education designed to help people question AI health advice critically (Tbaishat & Elfadel, 2025; Ma et al., 2025). Lifestyle programmes may also use chatbots to deliver support. These are ways to extend or tailor existing approaches, not evidence that AI itself improves mental health.

For populations and services, machine learning can help examine neighbourhood patterns and identify factors associated with poorer mental health (Du et al., 2023). Participatory modelling can help planners compare service scenarios, but projections are not the same as measured improvements in people’s lives (Vacher et al., 2024). At the individual service-entry point, an AI self-referral chatbot was associated with a larger rise in referrals than in control services (15% against 6%), and an independent assessment confirmed the increase in referrals but found no relevant evidence on quality and accuracy, with cost-effectiveness only explored (Habicht et al., 2024; Fleeman et al., 2026).

AI can help draft or adapt information, including psychoeducation for people with more complex conditions. But quality varies: one comparison found that AI-generated public mental health answers did not meet a sixth-grade reading standard, while a study of psychosis materials found inclusivity rated lower than accuracy and clinical usefulness, with text written at a high reading level (Wang et al., 2026; Yilanli et al., 2026). Translation technology may help when language support is scarce, but people interviewed preferred human-supported use for emotionally complex conversations (Kreienbrinck et al., 2026).

Phone and wearable data could help identify changes over time and prompt a check-in, including for people already receiving care. A review found moderate-to-good accuracy for relapse prediction, but most studies had a high risk of bias and none showed that acting on predictions improved outcomes (Fang & Chen, 2026). Frequent self-reports have also shown promise for predicting near-term suicidal thinking, although accuracy varies between studies (Melia et al., 2025). Consent, governance and data stewardship remain part of making this research usable in practice (O’Dea et al., 2026).

Compared with what? Judge AI against the support people would otherwise receive, including existing information and services, waiting or no support, and whether it adds benefit without shifting unacceptable risks onto users.

Wellbeing and health literacy

Read the education evidence alongside the broader lifestyle context.

Population planning

Some tools inform planning; modelled scenarios should not be mistaken for evaluated outcomes.

Triage and access

Read the promising referral findings alongside the independent assessment of evidence gaps.

Developing resources

These studies cover public information, severe illness, language access and co-design.

Personal and device data

Read accuracy findings alongside implementation limits and data governance questions.

All cards for

Overview last reviewed: 27 October 2026

Is AI support better than nothing, and for whom?For much of the world the alternative is no care at all. The question is whether AI support is safe, locally relevant and connected to human help.

Start with the size of the gap. WHO's Mental Health Atlas 2024, drawing on 144 countries, puts the global median at 13 mental health workers per 100,000 people. In low-income countries there is little more than one mental health worker of any kind per 100,000, compared with more than 60 in high-income countries, and two-thirds of reporting countries have one psychiatrist for 200,000 people or more. In low-income countries fewer than 10 per cent of people with a mental health condition receive care (World Health Organization, 2025a, 2025b). For much of the world, nothing is the current alternative.

WHO's position has two sides. Its AI for health programme says the potential of digital technologies remains largely untapped, especially in low- and middle-income countries, and names skilled workforce gaps as one of the problems AI could help address (World Health Organization, n.d.). Its mental health report sees digital tools as key to scaling care for common conditions, but warns that relying on them too heavily can exclude the most vulnerable (World Health Organization, 2025b). And a WHO-supported expert workshop in 2026 found that everyday use of general-purpose AI for emotional support has far outstripped investment in understanding its effects (World Health Organization, 2026). Read together, the message is less "use more AI" than "use it deliberately, inside health systems, with evidence".

Countries with the thinnest services are framing the issue in their own terms. India, with about 0.75 psychiatrists per 100,000 people against a benchmark of three, has built Tele-MANAS, a free 24-hour helpline in more than 20 languages that handled about 3.43 million calls by March 2026. Its Economic Survey 2024-25 proposed integrating AI-based solutions, and authors from NIMHANS, which supports the service, suggest carefully adding AI as a supervised layer for triage and counsellor support, with human oversight, rather than a stand-alone app (Suhas et al., 2026; Ministry of Health and Family Welfare, 2025; Sriram, 2026). In Indonesia, a parliamentary brief counts 3,136 active psychologists against a need of 9,947 and one psychiatrist per 200,000 people, and concludes chatbots should support professionals within the national health system (Winurini, 2025). Young Indonesians are already confiding in chatbots as a safe, cheaper space, while local researchers are adapting a system similar to Australia's THIS WAY UP programme (The Jakarta Post, 2026; Napitupulu, 2026). Nigeria has fewer than 300 psychiatrists for more than 240 million people, and chatbots are being offered there with psychiatrists cautioning about their limits (Beeler, 2026). In the Middle East and North Africa, a review describes treatment gaps of 80 to 95 per cent and sees promise in adapted, human-supervised tools (El Hajj et al., 2026).

People who may gain most are those facing distance, cost, language barriers or long waits. In England, an AI self-referral chatbot was associated with more referrals to NHS Talking Therapies, with particularly large increases for non-binary people and ethnic minority groups (Habicht et al., 2024). Interviews across five countries found translation tools could help in urgent or under-resourced settings, but fell short in emotionally complex conversations (Kreienbrinck et al., 2026). In Jordan, a WHO trial found its scripted STARS chatbot, with brief helper calls, reduced anxiety and depression more than enhanced usual care among distressed young people, including Syrian refugees (Bryant et al., 2026). Rural Australian research also points to care navigation as a way to make tele-mental health easier to use (Fisher et al., 2024).

The better-than-nothing case has limits. A tool is not a workforce (Bruckner et al., 2011), and a Wellcome editorial warns of a two-tier pattern in which wealthier people see therapists and poorer people only get chatbots (Brown & Herz, 2026). A review of Global South grey literature found SMS, voice and community health worker platforms reached far more people than smartphone apps, so smartphone-first AI may miss those with least access (Onyango, 2026). Evidence on outcomes is still thin: a review of trials found modest symptom reductions alongside high risk of bias (Sohn et al., 2026). Survey accounts, such as former Ash users saying their access to support worsened when the service closed, show need rather than effectiveness (Stamatis et al., 2026).

For support to be governed well, communities need a say in what is built, what data it uses, and how people reach human help. Interviews with AI researchers and innovators across 13 low- and middle-income countries highlight local readiness, language, participation, oversight and capacity-building (Baka et al., 2026), and a Kenyan team has co-designed a platform with young people and is now testing it, while a Global South review looks at cultural fit (Macharia et al., 2026; Lebese & Tsibolane, 2026). Reviews also stress privacy, Indigenous data sovereignty and locally led governance (Khan et al., 2026; van der Boor & van der Boor, 2026). Because people already use general chatbots as de facto care, some researchers argue that regulation has to start from that fact (Refoua et al., 2026).

Compared with what? In many low- and middle-income countries the realistic alternative is no care at all. Elsewhere it may be a long wait, or a service that is costly, distant or not available in someone's language.

The size of the gap

WHO's latest figures, read alongside the workforce analysis.

What WHO says about AI

Read these together: untapped potential, and use that has outrun the evidence.

India

AI as a layer inside public infrastructure.

Indonesia

A parliamentary brief, public commentary and local research.

Africa and the Middle East

Scarce services, early tools and calls for cultural adaptation.

Who gains access?

Examples across referral, language and rural access.

Limits of better than nothing

Equity risks, reach and the state of the evidence.

Governed for context

Read these together.

All cards for

Overview last reviewed: 27 October 2026

Will AI replace psychologists, or change what the work is?AI is taking on some tasks and creating new ways to practise, but evidence does not show it replacing psychologists as a whole.

AI is already being explored for the work around therapy: documentation, screening and measurement. In simulated psychiatric consultations, clinicians and simulated patients saw possible benefits from ambient scribes, alongside concerns about trust and privacy (Bokhari et al., 2026). A review of precision psychiatry describes some measurement approaches as useful in selected settings, while AI-assisted tools remain promising rather than ready for routine use (International Review of Psychiatry, 2026). Practice guidance stresses consent, secure data handling and professional accountability (Australian Psychological Society, 2026).

Training is one area where AI may extend what human clinicians can do. A review of virtual-patient studies found reported gains in interviewing and communication, but whether these transfer to real clinical work is uncertain (Gilbert et al., 2026). In a randomised study, simulated practice improved novice counsellors’ reflections and questions when paired with AI feedback, but practice alone did not (Louie et al., 2026). Automated supervision feedback also has limits: experts judged 89.2% of feedback points justified, yet 11 of 21 feedback instances included at least one unjustified point (Montesano et al., 2026).

Surveys suggest clinicians are already using AI, although what counts as use and who is surveyed varies. In a 30-country survey, 54.6% of psychotherapy professionals reported purposeful use of generative AI in clinical practice; common tasks included treatment planning and administration (Blease et al., 2026). An Australian survey found professionals using AI mostly for research and report writing (Cross et al., 2024). Practitioners have also described which chatbot tasks they would accept help with and where they draw the line (Goldie et al., 2025). Psychologists were more positive about AI virtual patients for training than AI acting on their behalf (Hlynsson et al., 2026).

The work is not just a set of tasks. A World Psychiatric Association guideline says digital and AI tools should support, not replace, care involving providers (Hilty et al., 2026). A review focused on general practice likewise keeps the clinician-patient relationship central while weighing access against risks such as bias, privacy and unreliable outputs (Blashki et al., 2026). The curator papers widen the question beyond the clinic, asking how psychology’s work across communities, education and organisations may change too. The evidence so far describes tools and early use, not a settled answer about the future of the profession.

Compared with what? For many people the realistic comparison is not AI versus a psychologist, but AI-supported work versus delayed, limited or no support at all.

Work around therapy

Read the practice evidence alongside the consent and governance guidance.

Training and supervision

Promising results, with important questions about realism and feedback quality.

Clinicians using AI

These surveys describe different settings and kinds of use, so their findings are not directly interchangeable.

What stays human?

Read the wider professional and relational questions together.

All cards for

Overview last reviewed: 27 October 2026

The challenges of guardrails: what are they, and can they work?Guardrails are trained into a model and built around it, not written as rules, so they have to be taught and tested again for every version and setting.

Many people picture a guardrail as a rule written into the software, such as "if someone mentions suicide, show a helpline number". Large language models don't work like that. They are not programmed line by line. They learn patterns of response from huge amounts of text, and further training then shapes how they behave. In practice, guardrails come in layers: training the model itself to respond safely, written instructions about how it should behave, separate classifiers that check conversations and block responses that break the rules, and product features such as helpline referrals, break reminders, routing sensitive conversations to a safer model, and human review (OpenAI, 2025b, 2025c). Age checks and referrals are part of this too, not just model behaviour (eSafety Commissioner, 2026).

Writing the rule is only the first step. To change the model, developers turn their policy into training: people write examples of ideal responses, the model's answers are scored, and the model is updated so that higher-scored answers become more likely (OpenAI, 2025a). For a 2025 mental health update, OpenAI says more than 170 clinicians wrote ideal responses and rated the model's answers (OpenAI, 2025c).

Because the behaviour is learned rather than written in, it is never applied exactly as intended. Training goals can pull against each other: when an April 2025 update added users' thumbs-up ratings as a training signal, the model became more sycophantic and was withdrawn, even though OpenAI's tests had looked good (OpenAI, 2025a). Safety training can also be shallow, mostly shaping the first few words of a reply (Qi et al., 2024), and it can wear thin as a conversation goes on. OpenAI says its safeguards work more reliably in short exchanges (OpenAI, 2025b). Safety training also shapes tone: after provocation, models with heavy safety training showed almost no irritability, while models with light safety training scored far higher (Teferra et al., 2026).

Every new model, and every significant update to an existing one, changes what the model has learned, so its guardrails have to be taught and measured again. The same product name can hide very different behaviour: GPT-4 became less willing to answer sensitive questions between its March and June 2023 versions (Chen et al., 2023). OpenAI estimates that an October 2025 update cut responses falling short of its mental health standards by 65% to 80% compared with the GPT-5 version released two months earlier, and says it runs mental health evaluations on new models before release (OpenAI, 2025c). Changes made by others matter too. Fine-tuning GPT-3.5 Turbo on just 10 harmful examples, for under US$0.20, removed its guardrails, and even fine-tuning on ordinary data weakened them (Qi et al., 2023). A test of 21 downloadable models found that their defences largely failed against tampering (Hossain et al., 2026). The same model can also behave differently through a public app and through the API that other services build on (Cantor et al., 2026). A safety result therefore describes one version, in one setting, at one time.

A key difficulty is that risk can emerge gradually. A model may sound supportive while reinforcing a user’s belief, or miss an indirect sign of distress; a response that looks acceptable in a short prompt may be different after many exchanges. In one study, harmful responses to crisis inputs were uncommon overall, but some models struggled with indirect or ambiguous signals (Arnaiz-Rodriguez et al., 2026). Another paper argues that safety tests should track whole conversational trajectories, not only final answers (Morrin et al., 2026).

Tests and benchmarks offer evidence about specific models, prompts and safeguards, not a guarantee of future performance. A suicide-risk benchmark found that models often did not ask direct follow-up questions or guide users adequately towards human care (Bentley et al., 2026). Developers also publish safety updates and benchmarks. OpenAI describes work to recognise risk across conversations and a benchmark developed with licensed mental-health experts (OpenAI, 2026). These claims are useful to examine (Kivira × University of Roehampton, 2026), but any benchmark score should be read alongside independent testing, because its design, the scenarios tested and what counts as a good response all affect what the score can show.

The hub's starting papers ask whether safeguards are designed around real mental health settings and community needs, rather than added after a general-purpose system has been built. The hub’s sociocultural question is who defines safe, helpful and harmful. Language, family relationships and ways of expressing distress vary, so a safeguard based on narrow assumptions may misread a person or community. Community participation offers one response: a multilingual study involved people with lived experience in its advisory work, while a Hong Kong project co-designed a chatbot with people living with depression and clinical professionals (Ghadigaonkar et al., 2026; Lo et al., 2026). These approaches do not prove that a system is safe, but they widen whose knowledge shapes it. The public-health comparison is not simply AI versus ideal care: it is also AI versus the support someone can realistically access.

Compared with what? Compare a guarded system with the realistic alternatives available to its users, including no support, human services, and unguarded or differently configured AI.

Start here: the hub's frame

The curator's two preprints: why the debate is about the wrong question, and why people at different points on the mental health spectrum talk past each other.

Techniques and limits

Start with the OpenAI explainers for how guardrails are built, then the research on how easily they shift.

Independent tests

These studies test different prompts and settings, so their results are not directly interchangeable.

Developer claims, critically

Compare stated intentions and benchmarks with independent evaluations.

Sociocultural approaches

Read these for examples of lived-experience input, co-design and local adaptation.

All cards for

Overview last reviewed: 28 October 2026

Is AI going to destroy us?A useful answer depends on what “destroy us” means, how plausible the pathway is, and what risks we face along the way.

“Is AI going to destroy us?” is a question about a possible extreme future, not a settled prediction about what will happen. It helps to separate existential risk, which could threaten humanity’s future, from serious harms that may affect people sooner. The International AI Safety Report reviews the capabilities and risks of general-purpose AI (Bengio et al., 2026), while other research connects nearer harms with possible longer-term existential scenarios (Subasri et al., 2025). Those are different kinds of claims, and evidence about present harms does not by itself establish how likely an existential outcome is.

The case for serious concern often focuses on what could happen if AI systems become much more capable, difficult to oversee, or usable for harmful purposes. AI 2027 presents one detailed, branching scenario, explicitly as a scenario rather than a prediction. Its tracker compares the scenario’s claims with real-world evidence, but tracking a scenario does not establish that its future events will occur. Interviews and arguments from figures such as Hinton and Suleyman add perspectives on possible risks, rather than a shared forecast.

There is also disagreement about whether existential risk is a well-supported or useful focus for policy. A Brookings article lays out arguments on both sides, and a separate collection of views from 19 Brookings scholars shows that experts differ about the scale of the threat and appropriate responses (Klein et al., 2026). These differences matter: the sources describe substantial uncertainty, and they do not provide a settled probability that AI will cause human extinction. Nearer harms and severe but non-existential outcomes still belong in the discussion.

For planning, the practical question is how to reduce plausible harms without treating a worst-case scenario as certain. Australia’s AI Safety Institute is intended to test models and applications, support regulators and agencies, and contribute to international governance (Department of Industry, Science and Resources, 2026). A commissioned report examines how capable AI systems might influence human overseers and how to evaluate that risk (Australia’s AI Safety Institute, 2026). Other proposals include binding rules and independent oversight, as well as evaluation, reporting and liability (Diallo, 2026; Klein et al., 2026).

Compared with what? Compare the possibility of an AI-driven existential catastrophe with nearer AI harms and with the realistic alternative of less capable or less widely deployed AI, not with a risk-free future.

Start here

Read these together for an introduction to the risks and the link between nearer harms and longer-term possibilities.

The case for concern

Scenarios and commentary can help map arguments, but they are not all forecasts or research findings.

Sceptics and critics

These sources foreground disagreement and questions about how to frame the risk.

What to do now

Read these for examples of oversight, evaluation and governance responses.

Watch and listen

Bengio and Harris make the case for concern; Luccioni argues the real harms are already here. Two ways into the AI 2027 scenario: a 34-minute explainer and the authors' own three-hour discussion.

All cards for

Overview last reviewed: 27 October 2026

How is AI reshaping the conditions that cause distress?AI may change the social conditions that shape mental health, from job security and environmental pressures to who controls the systems.

Mental health is understood here as a continuum, shaped not only by individual experiences but also by social conditions. AI may change the social and economic conditions that shape mental health, and it may also help advance care (Ettman & Galea, 2023). For example, machine learning has been used to examine links between neighbourhood conditions and poor mental health across US census tracts (Du et al., 2023). But social factors beyond age and gender are rarely used in mental-health AI research, a gap that limits what it can reveal about populations (Scherbakov et al., 2025).

Work is one route through which AI could affect distress: it may change job security and work demands, and the hidden digital labour behind many AI systems often comes with low pay, no benefits and the risk of being dropped at any time (Gray & Suri, 2019). Recent US studies offer different but not necessarily conflicting signals. One found no statistically significant rise in unemployment among recent graduates, while another reported lower employment among 22 to 25-year-olds in AI-exposed occupations, describing this as an early, non-causal indicator (Fairlie & Wu, 2026; Brynjolfsson et al., 2026). A separate survey found rising concern about AI-related job loss (Bracha et al., 2026). Reviews also describe technostress, while youth employment remains a wider policy concern (Wankhede & Khandelwal, 2026; Elder & Tobin, 2026).

AI may help researchers map environmental exposures, but the evidence and the technology’s own footprint both matter. A scoping review found only seven studies using machine learning to predict population health risks from climate-sensitive extreme weather, including post-traumatic stress (Ssebyala et al., 2024). Street-view studies offer ways to assess surroundings, but one estimates perceived stress from images using a model trained on ratings from 20 people, rather than asking residents (Han et al., 2022); a review of greenery and mental health found low or very low evidence quality (Bardhan et al., 2024). Meanwhile, estimates and projections raise questions about data centres’ electricity and water use (Aczel et al., 2026; IEA, 2026).

Who sets AI priorities, owns infrastructure and can shape the systems may influence who benefits and who bears costs. A planetary-health review argues for governing AI as a global commons (Creutzig et al., 2026), while the African Union strategy calls for Africa-centred, equitable AI development (African Union, 2024). Other policy analyses describe countries balancing domestic capacity with reliance on foreign systems, and middle powers seeking to preserve choice through cooperation or open infrastructure (Chapman et al., 2026; Rediker, 2026; Tony Blair Institute for Global Change, 2026). These are governance arguments, not evidence that a particular approach improves mental-health outcomes.

Compared with what? Compare AI-shaped work, environments and services with the realistic alternative of existing insecurity, environmental pressures and, for many people, little or no mental-health support.

Start here: population health

These readings frame AI as part of the wider conditions shaping mental health.

Work and security

Read the employment findings as early signals, not settled causal estimates.

Environment and resources

Read these together to consider both environmental exposure and AI infrastructure’s footprint.

Control and sovereignty

These policy and analysis pieces explore whose priorities shape AI systems and infrastructure.

Watch and listen

Geopolitics (Cadwalladr, Bremmer, Toews, Wang, Graylin) and the resource costs (Ren, Luccioni).

All cards for

Overview last reviewed: 27 October 2026

More questions (9)
What are the real risks, and why do the debates get mixed up?The risks depend on what the AI does, how people use it, and what support it replaces or adds.

Acute harm is the question of whether a chatbot responds safely to a crisis, such as suicidal thoughts, or instead misses cues or reinforces dangerous beliefs. In one evaluation, harmful crisis responses varied by model, and some models struggled with indirect or ambiguous signs of risk (Arnaiz-Rodriguez et al., 2026). A review found that none of 29 commercial agents met its criteria for an adequate crisis response (Elbarbary et al., 2026). But early reports of delusion-related harms are self-selected and retrospective, so they cannot show how common the harms are or whether chatbots caused them (Morrin et al., 2026).

A different debate is about relationships with chatbot companions: whether feeling understood or attached helps with loneliness, or can displace human connection and deepen isolation. A 12-month study found that feeling less socially connected predicted later chatbot use, and more chatbot use predicted greater emotional isolation, while it did not significantly change social connection (Folk & Dunn, 2026). A review describes short-term studies reporting reduced loneliness, alongside longer-term concerns about loneliness and socialising (Pi & Hunter, 2026). The evidence is still developing, and a government review notes that much of it comes from short-term studies or company disclosures (Greene-Sanders, 2026).

Algorithmic feeds raise another question: how recommendation systems and commercial design shape what people see and how they use platforms. This is not the same as a chatbot giving a particular answer. A working paper studying Instagram’s move to an algorithmic feed found worse mental health among teenagers, with social isolation and negative social comparison as possible pathways (Mandile, 2026). Young people interviewed about recommended content described repeated exposure and limited control, while also pointing to the value of support and connection in feeds (Winstone et al., 2026; Milton et al., 2026). Researchers caution that evidence isolating algorithms’ effects remains thin (Metzler & Garcia, 2024).

Therapeutic claims and medical devices bring questions about whether a specific tool improves mental health, what evidence supports its claims, and whether it should be regulated as a health product. In a review of depression-chatbot trials, the overall effect on symptoms and adherence was not statistically significant (Huang et al., 2026). A separate trial found benefits on some wellbeing measures, but not on depression, anxiety or stress (Cachia et al., 2026). The TGA says device regulation depends on a tool’s intended purpose (Therapeutic Goods Administration, 2026). These debates get merged because the same broad label, “AI and mental health,” covers crisis conversations, companionship, feeds and clinical products. Their risks and evidence are not interchangeable: a feed study does not establish a chatbot’s crisis safety, and a trial of one app does not prove that general-purpose chatbots work as therapy.

Compared with what? The realistic comparison may be no support at all, human connection or care, or a different digital service, depending on the question.

Start here: Acute harm

Read the crisis evaluations alongside the limits of real-world harm reports.

Start here: Relationships

Evidence differs by study length and by what counts as connection or loneliness.

Start here: Feeds and design

These sources cover recommender systems, commercial design and users’ mixed experiences.

Start here: Therapy and devices

Read product-specific trial findings alongside the question of regulatory purpose.

Why the debates get mixed up

The curator's preprint on clinicians and public health practitioners describing different parts of the same field.

Watch and listen

Turkle on artificial intimacy, Bloom on loneliness, Perel in the therapy room, Odgers against the teen tech panic, and two Australian Black Dog Institute discussions.

All cards for

Overview last reviewed: 27 October 2026

Is the problem the chatbot or the business model?The question is not only what a chatbot says, but what the service is designed to keep users doing and what it does with their data.

There is evidence and concern about how engagement-focused design can shape people’s experiences, but it does not show that every feed or recommendation system harms everyone. A Dutch panel study reports worse mental health among teenagers after Instagram moved to an algorithmic feed, while young people interviewed in the UK described repeated exposure, emotional feedback loops and limited control (Mandile, 2026; Winstone et al., 2026). People with lived experience have also described feeds as sources of support and connection, and suggested more explicit controls (Milton et al., 2026).

Chatbots raise a related business-model question: what happens when a conversational service also sells advertising? OpenAI says its ads are labelled and separate from answers, and that conversations are private from advertisers (OpenAI, 2026). A medical ethics blog warns that sensitive disclosures could be used to profile people and target advertising, while noting that EU rules do not directly address these risks (Gross & van Kolfschooten, 2026). That is a warning, not evidence that every chatbot uses disclosures this way. Consumer Federation of America testimony advocates limits on using chat inputs for advertising (Winters, 2026).

Some proposed rules target how services are designed; others focus on who can access them. Australia’s draft feed proposal would let people over 16 choose whether recommendations shape their default feed, and its regulator has called for safer recommender design and user controls (Albanese & Wells, 2026; eSafety Commissioner, 2026). The European Commission has preliminarily found TikTok’s design features inadequate under the Digital Services Act, a finding that is not yet final (European Commission, 2026). By contrast, the proposed GUARD Act centres on age checks and restricting minors’ access to AI companions. The proposed CHATBOT Act includes limits on manipulative features and targeted advertising (GUARD Act, 2026; U.S. Senate Committee on Commerce, Science, and Transportation, 2026).

Privacy is part of the business-model question because personal disclosures may be commercially valuable, but the available evidence has limits. A study of mental health apps, not chatbots as a whole, found undisclosed trackers and incomplete disclosure of third-party AI processing (Georgiou et al., 2026). Authors writing about mental health AI argue that agreeing to terms of service is not adequate consent for using therapy transcripts to train AI (Larrauri et al., 2026). Australia’s draft privacy reforms propose stronger consent and other data-handling duties, while a draft children’s privacy code covers specified online services (Attorney-General's Department, 2026; Office of the Australian Information Commissioner, 2026). These proposals do not establish how any particular service handles data.

Compared with what? Where services are scarce or costly, the realistic alternative for some people is no support at all. Elsewhere it may be a service with less engagement-driven design, less personalised advertising and stronger privacy protections.

Engagement and feeds

Read the reported harms alongside accounts of support and user control.

What rules target

These sources distinguish design controls from access restrictions.

Chatbots and ads

Company statements, policy analysis and advocacy make different kinds of claims.

Privacy and data

The app study is relevant context, but does not establish practices across all chatbots.

All cards for

Overview last reviewed: 28 October 2026

Should we talk to AI versions of people who have died?Griefbots and generative ghosts are already for sale. A few small studies find comfort and closure in supervised, short use; ethicists worry about dependence, consent and a business that profits from not letting go. No one has tested what happens over months.

Chatbots that speak as a person who has died go by several names: griefbots, deadbots, generative ghosts, digital resurrection. They are built from a dead person's messages, voice and photos, and the newer ones generate new things the person never said rather than replaying old material (Morris & Brubaker, 2025). The market is real: one health journal report puts the digital legacy industry at about US$22.46 billion in 2024 and growing fast, with evidence and regulation trailing availability (Rebernik, 2026). Most of what is written is ethics and commentary. The empirical base is a handful of small studies, and no trial has tested whether talking to a simulation of the dead helps or harms grief over time.

What small studies show is more mixed than either the hopeful or the alarmed commentary. Ten mourners who used chatbots in grief, some as a simulation of the dead person, described a safe place to express feelings and a way to finish unfinished conversations, and most wanted the bot as a supplement to human help rather than a replacement (Xygkou et al., 2023). In a supervised study, 16 bereaved adults each spent about 20 minutes with two AI versions of someone they had lost; they preferred the version that spoke in the first person, forgave invented facts but not a wrong tone or an endearment the person never used, and every one said they would use it again while most feared what it might do to other grieving relatives (Manning et al., 2026). In Guangdong, 25 adults judged griefbots acceptable only conditionally, depending on timing, purpose and how raw the grief was, and saw commercial uses as more troubling than commemorative ones (Lu et al., 2026). A Japanese survey found most bereaved people did not want a digital bond, but nearly one in five did, and that wish predicted complicated grief five months later (Kawashima et al., 2023). That is the closest thing to a longitudinal warning sign, and it measured the wish, not the use.

The clinical argument runs through continuing bonds. Grief theory no longer treats ongoing connection with the dead as a problem in itself; the question is whether the bond helps a person adjust to a world without them. Griefbots sit awkwardly here. Unlike a memorial page, where mourners speak about the dead together, a griefbot is a private two-way conversation with a simulation, available at any hour (Jiménez-Alonso & Brescó de Luna, 2023). Ethicists worry this can keep a person in an in-between state, foster dependence, and blur the line that grieving is supposed to help them accept (Lindemann, 2022; Marceau, 2026). Others argue the bots cannot carry emotional weight at all and should be framed as a temporary ritual rather than a relationship (Bao & Zeng, 2024). The more hopeful reading is that, used for a time and with support, they might help some people say what was left unsaid.

The design and governance questions are clearer than the clinical ones. Cambridge ethicists propose standards for providers: a way to retire a bot with dignity, constant reminders that it is an AI, adult-only access, and consent from both the person being recreated and the person talking to them, with a warning about product placement through a dead relative's voice (Hollanek & Nowaczyk-Basińska, 2024). Unauthorised recreations, such as the avatar of a Chinese singer made without his parents' consent, show how thin the rules are (Spitale & Germani, 2026). China is the one jurisdiction with substantial law on this, covering the dead person's portrait rights, relatives' consent and labelling of AI content (Cheng, 2025). One early proposal was to regulate deathbots as medical devices, so that benefit would have to be shown before sale (Lindemann, 2022). The business model questions elsewhere in this hub apply with extra force: a service that profits from a bereaved person's attention has every incentive not to let them stop.

For practice, the issue has already arrived. Grief Australia ran practitioner training in 2026 on why clients seek AI contact with the dead, what the risks are, and how to assess, set boundaries and refer (Grief Australia, 2026). A reasonable stance from the cards below: ask clients about it rather than wait to be told; treat supervised, time-limited use as different in kind from a subscription that runs indefinitely; be most cautious in acute grief and with children; and remember that the strongest evidence of benefit comes from people who chose to use these tools, under supervision, for twenty minutes. The curator's position statement at the top of this collection sets out the broader view on where responsibility sits.

Compared with what? Compare a griefbot with the other ways people keep bonds with the dead: photos, letters never sent, visiting a grave, a memorial page, talking to them aloud. Then ask what the bot adds (a reply) and what that reply costs (someone else writes it, and someone else profits from it). Distinguish a therapist-guided conversation from a subscription.

Start here: what these things are

Definitions, the market and the design space.

What the small studies show

Interviews, user studies and surveys. None is a trial, and none follows people over time while using a bot.

The ethics of talking to the dead

Arguments for caution, temporary use and medical device status.

Design standards, law and governance

What providers should do, and the one jurisdiction with real rules.

In the consulting room

Practitioner guidance, plus the wider companion chatbot evidence this sits inside.

Watch and listen

Talks, a documentary and interviews. Personal accounts from Kurzweil and Berreby, fieldwork from Cearns, ethics from Voinea and the Cambridge authors, and a founder's pitch from 2019 for contrast.

All cards for

Overview last reviewed: 28 October 2026

Whose mind is AI built for?AI may reflect the culture it was built on, or be adapted with the people it is meant to serve.

The question is not simply whether AI can offer mental health information, but whose experiences it recognises and whose it may misread. Across the mental health continuum, needs range from general information to assessment and support. Research has identified gender and other biases in language models used for mental health, and an audit found diagnostic patterns differed across gender, race and ethnicity even when symptoms were held constant (Straw & Callison-Burch, 2020; Pendse et al., 2026). These findings raise concerns, but do not show that every model or use has the same risks.

Culture and language can change how a model responds. In experiments, changing the prompt language and cultural context changed models’ assessments of depression and recommendations for therapy (Jia et al., 2026). In a study with young Saudi women, a culturally sensitive chatbot aligned more closely with a Saudi therapist than generic GPT did, and participants preferred it more often (Aldaweesh et al., 2026). Rural South African adolescents also used AI in co-design workshops, while many noticed cultural bias in generated content (Dallison et al., 2025). These examples suggest adaptation and participation can matter, but do not establish broad clinical benefit.

Adaptation can mean more than translating words. Proposed approaches include addressing culture at every stage, from data and design to evaluation, deployment and governance, and involving communities beyond simple translation (Rai et al., 2026). Reviews focused on Africa report that mental health chatbots are rarely adapted to African cultures and languages, while Indigenous AI governance work highlights co-design, cultural relevance and data sovereignty (Lebese & Tsibolane, 2026; Khan et al., 2026). Evidence from First Nations digital mental health services points to better access and reduced distress where services are culturally responsive, but this is not evidence that AI itself produces those outcomes (Zhai et al., 2026).

Compared with what? Where culturally competent care is scarce or slow to reach, the realistic alternatives for some people are no support or a generic model.

Evidence of bias

Read the audit and prompt studies alongside the community co-design findings.

Cultural adaptation in practice

Examples range from tested adaptations to frameworks and proposals.

Non-Western and Indigenous contexts

Evidence types vary; some resources are reviews or proposals rather than evaluations of AI outcomes.

All cards for

Overview last reviewed: 27 October 2026

Who owns the data, and who decides?Data decisions happen at different scales: communities, individuals and governments may each have distinct interests and authority.

Indigenous data sovereignty concerns the authority of Indigenous peoples to govern data about their communities, lands and cultures. It is not simply another name for personal privacy: it includes collective priorities, culturally relevant ways of interpreting information, and decisions about use and benefit. The CARE Principles offer a way to put Indigenous governance alongside the FAIR principles for making data findable and reusable (Carroll et al., 2021). Work on digital health also emphasises community participation and Indigenous authority over AI and data governance (Cordes et al., 2024; Veras et al., 2026).

For mental health data, the question can be more individual and immediate: what information is collected, who receives it, and whether a person has a meaningful choice. An analysis of 25 apps found undisclosed tracker software across the apps studied, while many apps did not name specific AI providers (Georgiou et al., 2026). Authors writing about training AI with therapy transcripts argue that general terms of service are not enough and call for separate, explicit consent (Larrauri et al., 2026). A simulated consultation study raised privacy and trust concerns, but called for evaluation with real patients (Bokhari et al., 2026).

National or sovereign AI is a different question again. It is about a country’s ability to shape, access or build AI systems and infrastructure, rather than an individual’s control over a record or a community’s authority over its data. A report tracking government-backed projects describes a surge in these projects, while noting continued reliance on overseas model foundations (Chavez et al., 2026). Another analysis describes countries balancing access to foreign AI with domestic capability, and notes that some projects may fall short of their ambitions (Chapman et al., 2026). These are policy and strategy discussions, not evidence of better mental health outcomes.

These questions can overlap without becoming interchangeable. A locally built system might support national capability, but that alone would not establish Indigenous authority or protect a person’s privacy. Conversely, individual consent may not settle a community’s concerns about collective data use. For any proposed use, ask who sets the rules, who can say no or change them, who benefits, and who carries the risks. Compare the proposal with the actual alternative available, including existing services and data arrangements, rather than an idealised option with no trade-offs.

Compared with what? Compare each proposal with the available care, data-governance and AI-infrastructure alternatives, not with an idealised risk-free system.

Indigenous data sovereignty

Read governance principles alongside community-led examples and policy gaps.

Mental health privacy

These sources cover apps, therapy data and consent in clinical practice.

National sovereign AI

Read strategy claims alongside policy direction and a language-focused model example.

Watch and listen

Indigenous data sovereignty explained by its leading voices, Australia's national health AI infrastructure, and Bell on decolonising AI. Plus the BBC explainer on the OpenAI agent's breach of the Medicare statistics portal.

All cards for

Overview last reviewed: 27 October 2026

What counts as good evidence, and who gets to say?Good evidence depends on the question, the comparison, who was included, and whether people affected helped define success.

Evidence is strongest when it answers the question people actually need answered. A trial can test whether a particular tool changes a particular outcome for a particular group, against a stated alternative. Reviews can show how consistent findings are across studies, but combining studies does not remove their limitations. One review of 29 trials found symptom improvements, but rated the evidence low to very low certainty; another review of 11 trials found no statistically significant overall effect on depression or adherence (Gong et al., 2026; Huang et al., 2026).

Results can differ by tool, outcome and comparison. In a six-week trial with undergraduates, an AI app improved some measures of resilience and social wellbeing, but not depression, anxiety or stress compared with usual care (Cachia et al., 2026). A small trial comparing ChatGPT with human coaching found no significant difference in procrastination outcomes, and some participants reported unwanted effects (Hennemann et al., 2026). Reviews also find gaps: studies of youth apps did not systematically assess harms, and relatively few large language model studies used real patient data and clinician assessments (Høgsdal et al., 2026; Lokadjaja et al., 2026).

Benchmarks and evaluation frameworks test different things from trials. A benchmark may check whether a model responds safely to set scenarios; a clinical framework can ask whether evaluation covers technical performance, safety and real-world outcomes. For example, a working paper found models were best at detecting potential suicide risk but often failed to ask direct follow-up questions or guide users to human care, while a comparison found models answered very high-risk prompts directly more often through APIs than through public interfaces (Bentley et al., 2026; Cantor et al., 2026). These tests can expose weaknesses, but performance on prepared prompts alone does not establish benefit in care.

People with lived experience can help define which outcomes matter, which situations are missing, and what feels safe or acceptable. Co-design work has shaped a chatbot prototype and a survey, while workshops with service users and peer specialists explored both possible benefits and risks (Lo et al., 2026; Lawn et al., 2026; Peng et al., 2026). Developer-led benchmarks can also make methods visible, but their results answer the questions their designers chose to test. OpenAI describes a benchmark developed with mental health experts; reading it critically means checking its scope and whether independent, real-world evidence supports its claims (OpenAI, 2026).

Compared with what? For many people the realistic alternative is no support or usual care, while some studies compare AI with human support.

Trials and reviews

Read the trial results alongside the reviews and their cautions about certainty and harms.

Benchmarks and evaluation

These assess different aspects of performance; none alone establishes real-world benefit.

Lived experience and co-design

Read these for how lived experience can shape research questions and design.

Reading developer research

Consider who designed the evaluation, what it measures, and what it cannot show.

Which part of the spectrum is the evidence about?

The curator's preprint on why clinical and population evidence answer different questions.

All cards for

Overview last reviewed: 27 October 2026

How should AI be governed, and is regulation arriving too late?AI is already entering mental health, while rules differ on what should be allowed, tested and monitored.

AI governance is not simply a choice between letting systems develop and banning them. It asks what uses are acceptable, what evidence and safeguards are needed, who is accountable, and how affected people can take part. WHO health guidance sets principles around autonomy, safety, transparency, accountability, inclusion, equity and sustainability, while the OECD offers a practical checklist for adoption. A snapshot of 240 health AI policies found advisory transparency tools common and binding obligations rare, but a policy count cannot tell us how rules work in practice (WHO, 2021; OECD, 2026; Moss et al., 2026).

In Australia, the debate may feel late because people are already using general-purpose AI for mental health. A Black Dog Institute roundtable reported this use and identified gaps in knowledge, regulation and infrastructure. A Black Dog Institute explainer says many consumer-facing tools lack AI-specific oversight. The TGA says whether a digital mental health tool is regulated as a medical device depends on its intended purpose, and its consultation is considering whether to keep, remove or change an existing exclusion. These sources describe current gaps and proposals, not proof that a particular approach will work (Maidment et al., 2026; Maidment, 2026; Therapeutic Goods Administration, 2026).

The EU and UK illustrate different approaches and timelines. An analysis of the EU AI Act discusses how health AI is classified, while a 2026 agreement would delay some high-risk rules. In the UK, an independent commission recommended staged authorisations and ongoing monitoring for healthcare AI. In the United States, states have become policy laboratories, but a 2026 map describes a fragmented landscape and a viewpoint argues for a federal floor. An FDA advisory committee has also identified risks such as hallucinated advice, missed crises and model drift. Recommendations, proposals and enacted rules are not the same thing (van Kolfschooten & van Oirschot, 2024; Niestadt, 2026; Medicines and Healthcare products Regulatory Agency, 2026; Shumate et al., 2025; Gluck, 2026; Aldhalimi, 2026; U.S. Food and Drug Administration, 2025).

China’s rules for public-facing generative AI services combine support for innovation with requirements on data, content, discrimination and user protection. International health guidance adds a different lens: a WHO expert workshop recommends mental health impact assessments and monitoring effects across whole populations, not just clinical outcomes, while a World Psychiatric Association guideline says AI should support, rather than replace, care and should be evidence-based and supervised. Read together, these approaches suggest governance that reaches beyond device approval alone. But guidance does not by itself show that harms are prevented, and the available policy comparisons do not settle whether regulation is arriving too late (Cyberspace Administration of China et al., 2023; World Health Organization, 2026; Hilty et al., 2026).

Compared with what? Compare governed AI with today’s patchwork of rules and, for people seeking mental health support, the realistic alternative may be no support at all.

Australia

Read the current landscape alongside the TGA’s open consultation.

EU and UK

Separate legislation, regulatory guidance and recommendations.

United States

Read the state-law mapping alongside federal oversight discussions.

China and elsewhere

Different legal and policy settings shape what governance prioritises.

International and health-specific guidance

Read principles and practical guidance with the policy landscape in mind.

Watch and listen

WHO and UN voices (Labrique, Pujari, the WHO/Europe conference and webinar series, the ADB and WHO Western Pacific session on who gains), then government plans in action: England's 10 Year Health Plan from three independent institutes, India's national strategy launch, US policy staff training, and CSIRO on Australian adoption.

All cards for

Overview last reviewed: 27 October 2026

What changes when AI moves onto our bodies and our own computers?Wearables and local models can keep data closer to home, but they move recording and responsibility to places current safeguards don't reach.

Most of the evidence in this hub is about chatbots that people open on a phone or computer and that run on a company's servers. Two shifts are changing that picture. AI is moving into devices worn all day, such as companion pendants and camera glasses, and capable models can now run on a person's own phone or computer. The best open-weight models, which anyone can download and run, are estimated to lag the leading closed models by less than a year (Bengio et al., 2026).

Running AI locally can keep sensitive information off company servers. Researchers have built mental health tools that run entirely on a phone, including an offline support app tested with students and a psychiatric decision-support platform designed so that patient data never leaves the device (Vimaleswar A et al., 2025; Bandara et al., 2026). Small models on phones have predicted stress with low delay and predictable resource use (Soyebo et al., 2026). In clinical research, a self-hosted open model rated patient engagement across 1,131 therapy transcripts without the data leaving institutional servers, and its ratings were linked with alliance, motivation and later symptoms (Eberhardt et al., 2025). Two psychologists who reviewed a local documentation assistant both rated on-device processing as decisive for their trust (Pico et al., 2026). Apart from the transcript study, these are preprints or early demonstrations, not outcome trials.

The privacy case is real. An audit of 179,057 ChatGPT conversations from India, Nigeria, Brazil and Pakistan found 21.31% contained personal health data, and over 95% of memory entries were extracted without the user explicitly asking (Zaman & Hoque, 2026). Popular mental health apps also send data to trackers their policies don't name (Georgiou et al., 2026). Small multilingual models have also shown potential for privacy-preserving clinical work in low-resource languages, although psychological complaints remained hard to extract (Ghaffarzadeh-Esfahani et al., 2026).

The trade-off is that local models sit outside the safeguards discussed elsewhere in this hub. Open-weight models' safeguards are easier to remove, their use is harder to monitor, and once released they cannot be recalled (Bengio et al., 2026). A test of 21 open-weight models found current alignment-stage defences largely failed to withstand tampering (Hossain et al., 2026), and another study found 3,471 'uncensored' models on Hugging Face, with copies persisting across platforms after removal (Garcia et al., 2026). Neither study looked at mental health content specifically. Cloud tools are not a high bar either: across 21 studies of mental health chatbots, crisis protocols were mostly underdeveloped and only two reported formal adverse-event monitoring (Olisaeloka et al., 2026).

Wearable devices raise different questions. Some companion pendants listen through the day and send messages back to the wearer, so companionship is built into daily life rather than opened on demand (Frances & Noorily, 2026). Evidence on app-based companions is already mixed: one study found more grief expression and focus on relationships, alongside more language about loneliness, depression and suicidal thoughts (Yuan et al., 2026). Wearable data can also be turned into personal health insights by AI agents (Merrill et al., 2026), and passive sensing for mental health is a growing research field (Shen et al., 2025).

Camera glasses also affect people who never chose to use them. In a study of 525 people in China, bystanders wanted more transparency and protection than wearers were willing to give, and 65 to 90% would take defensive action in sensitive settings (Wang et al., 2026). Concern varies by culture, with South Korean respondents reporting greater privacy concerns than US respondents (Chock et al., 2026). Hamburg's data protection authority concluded that recording people outside the wearer's close circle will generally not be permissible (Hamburg Commissioner for Data Protection and Freedom of Information, 2026), and a report for the European Data Protection Board describes glasses becoming tools for observing others (Bruna, 2026). UNESCO's neurotechnology recommendation extends to data such as voice analysis, heart rate variability and sleep monitoring when used to infer mental states (UNESCO, 2025). Proposed design responses include processing on the device where possible and making sensing visible to people nearby (Axelrod, 2026).

Taken together, devices and local AI shift both control and responsibility toward individuals, clinicians and communities. That can suit privacy and data sovereignty, but it means safeguards have to be built into the tools people choose and into training and community norms, rather than assumed from a provider.

Compared with what? Compare a local or on-device tool with the cloud service it would replace, which may offer more safety features but less privacy, and with no support at all. For wearables, also count the people around the wearer, who did not choose the device.

Start here: Local and on-device models

Mostly preprints and demonstrations; the therapy transcript study is the strongest evidence.

Start here: Why keep data local

Read the cloud privacy evidence alongside what local tools give up.

Start here: When safeguards can be removed

The tampering studies are general, not mental health specific; the review sets the baseline for cloud tools.

Start here: Always-on companions and wearables

Companion evidence comes from apps; devices are newer than the research.

Start here: Bystanders and mental privacy

Regulators, researchers and UNESCO approach the same problem from different angles.

All cards for

Overview last reviewed: 27 October 2026

Where is this heading, and what might become possible?Four directions are visible: AI in the background of care, continuous personal measurement, models that run on hardware people control, and planning at population scale. Some of it is here now. Most of it is still promise.

Prediction in this field has a poor record, so this question maps directions rather than dates. Smartphone apps and virtual reality were both expected to transform care and arrived slowly, held back by weak evidence, poor engagement and slow implementation rather than by the technology itself (Torous et al., 2025). The same authors now forecast generative AI moving from wellness products towards health care proper as it gains access to clinical data, with the need for regulation and evidence growing as it does (Torous & Cipriani, 2025). Health technology assessment bodies already list 24-hour mental health chatbots used alongside a treatment plan among the technologies expected to affect care (Canada's Drug Agency, 2025). This question looks at four directions: AI in the background of care, personal measurement, models that run on hardware people control, and planning at population scale. For the risks of always-on devices and open models in particular, see the devices question.

The nearest change is AI in the background of ordinary consultations. Australian GP authors describe scribes, chatbots, digital therapies, just-in-time interventions, predictive analytics and agentic AI as the coming toolkit, while keeping the GP-patient relationship central (Blashki et al., 2026). Psychiatrists and simulated patients in one study saw ambient scribes reducing documentation burden but wanted transparent consent and clinician oversight (Bokhari et al., 2026), and the Australian College of Mental Health Nurses has already set out that care must not depend on agreeing to a scribe (ACMHN, 2026). NHS England plans £10 billion over three years on technology and data, with an AI triage tool due to reach all NHS App users by April 2028 (NHS England, 2026). In training and supervision, a multi-agent system rated therapist fidelity with very high agreement with human raters, although on synthetic sessions (Kamaleddin et al., 2026), and a simulated-patient environment showed models overestimating trainee competence (Rizzi et al., 2026).

A second direction is continuous, personal measurement. A roadmap in The Lancet Digital Health argued that measurement-based care built on phone and wearable data could be the foundation for AI in mental health (Koutsouleris et al., 2022). Since then an agent reading a person's own wearable data answered health questions with 84 percent accuracy (Merrill et al., 2026), a personal health model gave sleep coaching rated close to human experts (Khasentino et al., 2025), and one study collected up to 12 months of iPhone and Apple Watch data from more than 4,000 people (Douglas et al., 2026). Across 52 studies, passive sensing predicted relapse one day to four weeks ahead with moderate to good accuracy, but three quarters were at high risk of bias and none showed that acting on a prediction improved outcomes (Fang & Chen, 2026). Proposals for mental health digital twins (Natarajan et al., 2025), reinforcement learning that times support to the person (Park et al., 2026) and monitoring built on within-person change rather than population norms (Shaywitz et al., 2026) show where the field wants to go. The two personal health studies come from Google and should be read as developer research.

The third direction is models that run on hardware people and services control, often called local LLMs. Open-weight models anyone can download are estimated to lag the best closed models by less than a year (Bengio et al., 2026). A JMIR viewpoint argues that fully open models, deployed locally, offer the strongest privacy protection and the most practical route to AI health services in low-resource regions, with distilled models from 1.5 billion parameters able to run on a phone (Wu et al., 2026). NVIDIA researchers argue that small models are powerful enough and far cheaper for the narrow, repetitive tasks that make up most agent work, with large models kept for open conversation (Belcak et al., 2025). In health care the choice between a closed cloud model that is quick to deploy and an open model on local hardware that gives control but demands skills and money is now a recognised design decision, often resolved by hybrid systems (Dennstädt et al., 2025). Across 68 studies, locally deployed small models scored highest for privacy in healthcare and legal settings, though the reviewers stress these are architectural advantages rather than guarantees, and local models still face device theft, weight extraction and tampering (Dikici & Bilgin, 2026).

In mental health this is already visible in early form: a psychiatric decision-support app running three small models entirely on a phone (Bandara et al., 2026), an offline support app built on a one billion parameter model (Vimaleswar A et al., 2025), a documentation assistant for psychologists that processes sessions on the device (Pico et al., 2026) and a self-hosted model that rated 1,131 therapy transcripts without the data leaving institutional servers (Eberhardt et al., 2025). The same capability makes language-specific and community-controlled models possible: Arabic-centric open models (Sengupta et al., 2023), Te Hiku's Kaitiakitanga licence for Māori language technology (Te Hiku Tech, 2022), and middle-power strategies that weigh open, interoperable infrastructure against securing access to frontier systems (Tony Blair Institute, 2026; Leicht & Winter-Levy, 2026). Whether locally and regionally adapted models could serve cultural fit better than centralised systems, and what governance and quality risks that creates, is still an open question. Most of these are demonstrations and prototypes. None has been tested against standard care, and open models are also easier to strip of safeguards, which the devices question covers.

A fourth direction is planning rather than treatment. Systems modelling already helps Australian planners handle uncertain demand and choose investments at regional level (Whiteford et al., 2023), hybrid forecasting is being applied to long-term mental health workforce planning (Hewage & Rostami-Tabar, 2024), machine learning has predicted suicide risk in 32,661 adults in care with high accuracy in retrospective data (Peña-Gómez et al., 2026), and digital twins are proposed for emergency planning (Sun et al., 2025). At scale, India's Tele-MANAS handled about 3.43 million calls by March 2026 and its authors suggest adding triage assistants and multilingual speech-to-text for counsellors while keeping humans at the core (Suhas et al., 2026), and clustering 102,684 chatbot users found that more than half dropped out early (Wolfe et al., 2026). Retrospective accuracy and call volume are not the same as better outcomes, and none of the planning tools has yet been shown to change them.

What actually arrives will be shaped by forces outside the clinic. Data centre electricity demand rose 17 percent in 2025 and is projected to double by 2030 (IEA, 2026), governments backed 184 sovereign AI projects across 67 countries (Chavez et al., 2026), the business model of conversational services is shifting towards advertising (see the business model question), and scenarios such as AI 2027 sketch much faster capability growth than most health planning assumes (AI Futures Project, 2025). A reasonable reading of the cards below: expect scribes, triage and local tools soon; expect personal sensing and digital twins to stay at research stage until someone shows that acting on them helps; and expect the choice between cloud and local to become a routine clinical governance question rather than a technical one. The curator's own view on what should come first is set out in the position statement at the top of this collection.

Compared with what? Read each forecast against the record of mental health apps and virtual reality, which were also expected to transform care and arrived slowly. For each trend ask what it would replace, who would run it, whether it has been tested outside a demonstration, and whether a local model, a cloud service or no tool at all is the right comparison.

Start here: where the field says it is going

Reviews and forecasts from clinical researchers, plus horizon scans.

AI in the background of care

Scribes, triage, training and supervision tools: the changes most likely to reach ordinary practice first.

Personal measurement, prediction and digital twins

Wearables, passive sensing, just-in-time support and digital twins. Strong demonstrations, little outcome evidence.

Local and small models: what might become possible

Open-weight and small models that run on phones, laptops and institutional servers, and the trade-offs against cloud services.

Community, language and sovereign models

What local models make possible for languages, communities and middle-sized countries.

Planning and scale

Systems modelling, workforce forecasting, risk prediction and very large services.

What will shape the trajectory

Energy, geopolitics, speed of capability growth and scenarios.

Watch and listen

Open-weight and small models from a lawyer, a Hugging Face co-founder and a climate-focused researcher.

All cards for

Overview last reviewed: 28 October 2026

Part 1: Potential benefits

Most public debate about AI and mental health asks what could go wrong. This part asks what could go right, and for whom. Benefits are organised by function, because a tool that improves health literacy, one that helps plan services, and one that delivers support where no clinician exists raise very different questions and need different evidence.

B1. Health literacy and promotion

Can AI help people understand mental health, find credible information and act on it?

READ
Evidence reviewPeer reviewed

Artificial intelligence (AI) for social innovation in health education: promoting health literacy through personalized ai-driven learning tools – a systematic review

Review of 75 studies examining AI-supported health education and personalised learning, including comprehension, engagement and health literacy. Relevant to understandable public information and community education, while community-level implementation remains less developed.

Tbaishat, D. M., & Elfadel, M. W. (2025). Artificial intelligence (AI) for social innovation in health education: promoting health literacy through personalized ai-driven learning tools – a systematic review. BMC Medical Education, 26(1), 123. https://doi.org/10.1186/s12909-025-08462-3

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Can co-designed educational interventions help consumers think critically about asking ChatGPT health questions?

Australian randomised trial of co-designed animation and image-carousel education for adults without university-level education who had used ChatGPT. Supports practical public education about understanding AI health advice.

Ma, et al. (2025). Can co-designed educational interventions help consumers think critically about asking ChatGPT health questions? Results from a randomised-controlled trial. npj Digital Medicine. https://www.nature.com/articles/s41746-025-02056-5

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Evolving Health Information-Seeking Behavior in the Age of Artificial Intelligence and Voice-Assisted Technologies: Qualitative Study

Think-aloud study with 27 adults aged 19 to 80 searching Google, ChatGPT and Alexa for health questions, including a mental health scenario. People folded AI into existing search routines rather than replacing them: 85 percent valued ChatGPT's formatting and speed, usefulness often outweighed trust, and 56 percent used AI answers as a springboard for further checking. Relevant to how health literacy and AI literacy now overlap.

Wardle, C., Urbani, S., & Wang, E. (2025). Evolving health information-seeking behavior in the age of artificial intelligence and voice-assisted technologies: Qualitative study. Journal of Medical Internet Research, 27, e79961. https://doi.org/10.2196/79961

Open resource
Continuum
READ
ResearchPreprint, not yet reviewed

PEARL: a personalised exercise assistant using reinforcement learning

Large four-arm trial comparing no nudges, random nudges, fixed nudges and adaptive reinforcement-learning nudges. The adaptive system selects message content and timing; the primary analysis included 7,711 participants.

Lee, A. A., et al. (2025). A Personalized Exercise Assistant using Reinforcement Learning (PEARL): Results from a four-arm Randomized-controlled Trial. arXiv:2508.10060. https://arxiv.org/abs/2508.10060

Open resource
Continuum
READ
ResearchPreprint, not yet reviewed

Co-Pilot for Health: personalised algorithmic AI nudging

Twelve-week Singapore study of 84,764 participants using wearable data and a graph-based recommendation system. Reports higher step counts and moderate-to-vigorous activity in a matched comparison.

Chiam, J., Lim, A., Nott, C., Mark, N., Teredesai, A., & Shinde, S. (2024). Co-Pilot for Health: Personalized Algorithmic AI Nudging to Improve Health Outcomes. arXiv:2401.10816. https://arxiv.org/abs/2401.10816

Open resource
Continuum
READ
Evidence reviewPeer reviewed

Systematic review and meta-analysis of chatbots for lifestyle behaviours

Synthesises 19 studies with 3,567 participants on physical activity, diet and sleep. Useful evidence for conversational delivery of lifestyle programmes, with varied study designs and chatbot technologies.

Systematic review and meta-analysis of the effectiveness of chatbots on lifestyle behaviours. (2023). npj Digital Medicine, 6, 118. https://www.nature.com/articles/s41746-023-00856-1

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Physical activity interventions for depression, anxiety and distress

Overview of 97 reviews covering 1,039 trials and 128,119 participants. Provides the mental-health rationale for including physical activity in health promotion and AI-supported lifestyle programmes.

Singh, B., et al. (2023). Effectiveness of physical activity interventions for improving depression, anxiety and distress: an overview of systematic reviews. British Journal of Sports Medicine, 57, 1203-1209. https://pubmed.ncbi.nlm.nih.gov/36796860/

Open resource
Continuum
READ
Opinion and commentaryPeer reviewed

Is it time to better harness AI for improving lifestyle behaviours?

Explores adaptive wearable prompts, reinforcement learning and generative AI across lifestyle behaviours. Useful for designing health-promotion programmes that combine personalised support with broader public-health action.

Chaput, J.-P., Lopes, M. V. V., LeBlanc, A. G., & Goldfield, G. S. (2026). Is it Time to Better Harness Artificial Intelligence for Improving Lifestyle Behaviours? International Journal of Public Health. https://pmc.ncbi.nlm.nih.gov/articles/PMC12886051/

Open resource
Continuum
READ
ResearchPeer reviewed

Motivational interviewing-oriented AI assistant for vaccine communication

Randomised study of 177 adults comparing an AI assistant plus education with government web information over five weeks. Illustrates interactive delivery of a public-health campaign rather than one-way information alone.

Li, Y., et al. (2025). Effects of a Theory- and Evidence-Based, Motivational Interviewing-Oriented Artificial Intelligence Digital Assistant on COVID-19 Vaccine Attitudes Among Adults in Hong Kong. Journal of Medical Internet Research, 27, e72637. https://www.jmir.org/2025/1/e72637

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Adaptive bandit experiments to improve engagement in mental health

Eight-week real-world study involving 1,100 participants, alongside simulations, testing adaptive selection of message features. Shows how delivery systems can learn from responses rather than send the same prompt to everyone.

Kumar, H., et al. (2024). Using Adaptive Bandit Experiments to Increase and Investigate Engagement in Mental Health. Proceedings of the AAAI Conference on Artificial Intelligence, 38(21), 22906-22912. https://pmc.ncbi.nlm.nih.gov/articles/PMC11044947/

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

AI-Enabled Digital Health Promotion and Prevention: Computational Literature Review

Topic-modelling review of 6,328 Scopus articles on AI in digital health promotion, sorting the field into three contexts: technologies embedded in everyday devices, clinical and data-driven promotion, and population-level, societal and policy-oriented tools. Helpful map of where the promotion and prevention literature clusters, and where population-level work is still thin.

Carrilho, M. G., Pinto, D. C., Wagner, R., Rohden, S. F., Arriaga, M. T., & Pinto, L. Q. (2026). AI-enabled digital health promotion and prevention: Computational literature review. JMIR AI, 5, e84492. https://doi.org/10.2196/84492

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

AI Can Be a Powerful Social Innovation for Public Health if Community Engagement Is at the Core

Viewpoint arguing that AI can be a social innovation for public health (surveillance, community health worker support, health education, behaviour change) only if community engagement is built in from the start. Not mental health specific, but directly relevant to the community co-design question.

Bazzano, A. N., Mantsios, A., Mattei, N., Kosorok, M. R., & Culotta, A. (2025). AI Can Be a Powerful Social Innovation for Public Health if Community Engagement Is at the Core. Journal of Medical Internet Research, 27, e68198. https://doi.org/10.2196/68198

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Use of ChatGPT to obtain health information in Australia, 2024: insights from a nationally representative survey

Nationally representative survey finding that 9.9 percent of Australian adults, about 1.9 million people, had asked ChatGPT health questions in the six months to June 2024, with only moderate trust in the answers. Not mental health specific, but the best Australian population baseline for health-related AI use.

Ayre, J., Cvejic, E., & McCaffery, K. J. (2025). Use of ChatGPT to obtain health information in Australia, 2024: insights from a nationally representative survey. Medical Journal of Australia, 222(4), 210-212. https://doi.org/10.5694/mja2.52598

Open resource
READ
Opinion and commentaryPeer reviewed

AI Health Literacy: a reflective framework for LLM-based generative AI in public health education and promotion

Conceptualises AI health literacy as a judgement competence with three dimensions: task appropriateness, context of use and critical verifiability.

Scherenberg, V. (2026). AI Health Literacy: a reflective framework for LLM-based generative AI in public health education and promotion. Frontiers in Public Health, 14, 1891658. https://doi.org/10.3389/fpubh.2026.1891658

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Evaluating the Impact of Digital Health Literacy on the Adoption of Preventive Health Measures in Socioeconomically Vulnerable Communities: A Narrative Review

Narrative review of digital health literacy and uptake of preventive health measures in socioeconomically vulnerable communities, including validated DHL instruments.

Gupta, P. K., Jogdand, K., Yerpude, P., Haqqani, M. H., Suheb, M. A., & Sen, N. (2025). Evaluating the Impact of Digital Health Literacy on the Adoption of Preventive Health Measures in Socioeconomically Vulnerable Communities: A Narrative Review. Cureus. https://doi.org/10.7759/cureus.94276

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Developing a Digital Mental Health Platform for the Arab World: From Research to Action

Conceptual framework for the Mental Health Assistant (MeHA) platform, proposing mental-health information, educational resources, conversational interaction and connections to service providers. Relevant to health literacy and service navigation for Arabic-speaking communities.

Househ, M., Alam, T., Al-Thani, D., Schneider, J., Siddig, M. A., Fernandez-Luque, L., Qaraqe, M., Alfuquha, A., & Saxena, S. (2019). Developing a Digital Mental Health Platform for the Arab World: From Research to Action. Studies in Health Technology and Informatics, 262, 392-395. https://doi.org/10.3233/SHTI190101

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

APA guide to navigating AI-generated advice thoughtfully and safely

This American Psychological Association guide advises people to use AI-generated advice cautiously and verify health information with a healthcare practitioner. It says AI is not a safe or effective replacement for licensed mental health care and advises seeking immediate human help in a crisis.

American Psychological Association. (2026). APA guide to navigating AI-generated advice thoughtfully and safely. https://www.apa.org/topics/artificial-intelligence-machine-learning/guide-navigating-ai-advice.pdf

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

Using AI for mental health support? Here's what you need to know

This Australian Psychological Society article explains how AI might be used as a complementary tool, such as for wellbeing exercises or preparing questions. It warns that AI advice can be generic, inaccurate or inappropriate, and should not replace personalised, evidence-based care from a psychologist.

Australian Psychological Society. (2026, October 6). Using AI for mental health support? Here's what you need to know. https://psychology.org.au/insights/using-ai-for-mental-health-support-safely

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

National Health Literacy Framework

The Framework sets out a national approach with 7 guiding principles and 4 aims to strengthen health literacy, and reports that up to 60% of people in Australia have lower health literacy. A 2022 survey found fewer than 1 in 10 people fully understood what they read online, more than a quarter lacked enough information to manage their health, and half had trouble identifying credible online health information; the Framework estimates that a 25% improvement could save AUD 7.3 billion a year.

Why it matters now: New national framework, relevant now that AI chat is a common source of health information.

Australian Government Department of Health, Disability and Ageing (2026). National Health Literacy Framework. Australian Government Department of Health, Disability and Ageing. https://www.health.gov.au/sites/default/files/2026-08/national-health-literacy-framework.pdf

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

A scoping review of digital mental health literacy interventions for the general population

This scoping review covered 30 publications from 16 countries, involving more than 8,550 research subjects, on digital mental health literacy interventions. It found potential to improve mental health literacy outcomes, with AI tools such as chatbots supporting personalisation and engagement, and non-AI approaches offering accessible, cost-effective options.

Muhammad Khir, S., Bakar, A. N. A., Rashid, A. H. A. et al. (2026). A scoping review of digital mental health literacy interventions for the general population. Discover Psychology. https://doi.org/10.1007/s44202-026-00826-0

Open resource
Continuum
READ
Guidance and standardsNot reviewed

Counseling Patients and Families on Using AI Chatbots

The guidance recommends open conversations about chatbot use, AI literacy, privacy settings and family rules rather than simply telling families to avoid chatbots. It reports that 67% of teens aged 13-17 use chatbots, around three in ten daily, and 13.1% of U.S. youth aged 12-21 have used generative AI for mental health advice, while noting risks including misinformation, privacy concerns and emotional overreliance.

American Academy of Pediatrics (2026). Counseling Patients and Families on Using AI Chatbots. American Academy of Pediatrics, Q&A Portal Library. https://www.aap.org/en/patient-care/media-and-children/center-of-excellence-on-social-media-and-youth-mental-health/qa-portal/qa-portal-library/qa-portal-library-questions/counseling-patients-and-families-on-using-ai-chatbots/

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewedAdvocacy organisation

Supporting Children and Young People in an AI-Enabled World

Child Helpline International introduces three AI guidelines developed with child helpline members and expert partners, for helplines, children and young people, and parents and caregivers. The guidelines support informed conversations and safer AI use while emphasising that they do not replace human connection or cover every AI question or use in helpline operations.

Child Helpline International (2026). Supporting Children and Young People in an AI-Enabled World. Child Helpline International website. https://childhelplineinternational.org/supporting-children-and-young-people-in-an-ai-enabled-world/

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Beyond Chatbots: Generative AI as Public Health Infrastructure and a New Digital Social Determinant of Health

The authors argue that generative AI could become public health infrastructure and a digital social determinant of health by influencing access to health information, health literacy and care navigation. They present a conceptual argument, not empirical findings, and call for equitable access, governance, trust and human oversight.

Suraj P. G., Clemence Jenifer S., Joice Swarnalatha R. et al. (2026). Beyond Chatbots: Generative AI as Public Health Infrastructure and a New Digital Social Determinant of Health. Frontiers in Public Health, Digital Public Health. https://doi.org/10.3389/fpubh.2026.1910815

Open resource
Continuum
Also relevant to

B2. Planning and population intelligence

Can AI help identify need, understand people's circumstances and places, map services and target resources?

READ
ResearchPeer reviewed

Social Drivers of Mental Health: A U.S. Study Using Machine Learning

Machine learning analysis of 38,379 US census tracts identifying which behavioural, environmental and social drivers best explain neighbourhood-level depression and poor mental health. An example of AI used to study collective exposure rather than individual pathology.

Du, S., Yao, J., Shen, G. C., Lin, B., Udo, T., Hastings, J., Wang, F., Wang, F., Zhang, Z., Ye, X., & Zhang, K. (2023). Social Drivers of Mental Health: A U.S. Study Using Machine Learning. American Journal of Preventive Medicine, 65(5), 827-834. https://doi.org/10.1016/j.amepre.2023.05.022

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Is demography destiny? Machine learning and population health outcomes

Uses demographic data to estimate six population-health outcomes or risk factors across US states, with separate training and validation states. Supports comparing observed outcomes with demographic expectations.

Luo, W., et al. (2015). Is Demography Destiny? Application of Machine Learning Techniques to Accurately Predict Population Health Outcomes from a Minimal Demographic Dataset. PLOS ONE, 10(5), e0125602. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0125602

Open resource
Continuum
READ
Guidance and standardsPeer reviewed

Machine learning in population and public health: development and implementation guidelines

Guidance informed by literature reviews and a modified Delphi process with 17 experts. Supports community participation, representative data, subgroup comparison, testing and practical implementation.

Pinto, A. D., et al. (2025). Machine Learning Applications in Population and Public Health: Guidelines for Development, Testing, and Implementation. JMIR Public Health and Surveillance, 11, e68952. https://pmc.ncbi.nlm.nih.gov/articles/PMC12551935/

Open resource
Continuum
Also relevant to
READ
Tool or datasetNot reviewed

WorldPop Global2: high-resolution population estimates for 2015-2030

Population mapping infrastructure combining machine-learning spatial disaggregation with demographic estimation and projection, including age and sex. A potential foundation for catchment analysis, access mapping and changing service needs.

Tatem, A. (2025, September 3). WorldPop 'Global2': Global high-resolution population estimates for 2015-2030. WorldPop. https://www.worldpop.org/blog/worldpop-global2-global-high-resolution-population-estimates-for-2015-2030/

Open resource
Continuum
Also relevant to
READ
Evidence reviewNot reviewed

Approaches to projecting future healthcare demand

Reviews macro, cell-based, microsimulation and machine-learning approaches to future demand, including population growth, ageing, fertility, mortality and migration. Helps planners select methods suited to their question and data.

Matias, M. A., Santos, R., Kasteridis, P., Grasic, K., Mason, A., & Rice, N. (2022). Approaches to Projecting Future Healthcare Demand. CHE Research Paper 186. University of York. https://www.york.ac.uk/media/che/documents/papers/researchpapers/CHERP186_projecting_healthcare_demand.pdf

Open resource
Continuum
READ
ResearchPreprint, not yet reviewed

Predictive and prescriptive modelling for long-term mental-health workforce planning

Combines statistical and machine-learning demand forecasts with stock-and-flow modelling and regional nursing-workforce scenarios. Directly relevant to anticipating mental-health staffing requirements.

Hewage, H. C., & Rostami-Tabar, B. (2024, revised 2025). A hybrid predictive and prescriptive modelling framework for long-term mental healthcare workforce planning. arXiv:2406.17463. https://arxiv.org/abs/2406.17463

Open resource
Continuum
READ
ResearchPeer reviewed

Machine-learning forecasts of mental-health bed occupancy

Uses 866 weekly observations from 2008-2024 to compare forecasting methods for mental-health bed occupancy. Illustrates a concrete operational use for anticipating capacity pressures.

Avinash. (2025). Time series forecasting of bed occupancy in mental health facilities in India using machine learning. Scientific Reports. doi:10.1038/s41598-025-86418-9. https://www.nature.com/articles/s41598-025-86418-9

Open resource
Continuum
READ
ResearchPeer reviewed

Reducing mental-health emergency visits through participatory population modelling

Australian Capital Territory study involving 33 stakeholders and population-level scenarios for 2023-2032. Examines combinations of service capacity, community crisis care and coordination under changing population needs.

Vacher, C., et al. (2024). Reducing mental health emergency visits: population-level strategies from participatory modelling. BMC Psychiatry. doi:10.1186/s12888-024-06066-7. https://bmcpsychiatry.biomedcentral.com/articles/10.1186/s12888-024-06066-7

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

CDC AI Strategy FY 2026 to 2030 and guidance for public health agencies (March 2026)

CDC's first AI strategy, published 12 March 2026 with a new AI website. Four pillars: accelerated adoption, governance and public trust, AI across CDC data platforms, and an AI-ready workforce. Companion guides for state, tribal, local and territorial agencies cover generative AI (CDC estimates upwards of 41,000 staff hours saved a year and 103 AI use cases) and agentic deep research tools, with human review and expert validation required throughout. Useful as a model of what an agency-level AI strategy looks like; mental health is not mentioned.

Centers for Disease Control and Prevention. (2026, March 12). CDC's AI strategy (FY 2026 - FY 2030). Office of Public Health Data, Surveillance, and Technology. https://www.cdc.gov/ai/strategy/index.html

Open resource
Continuum
Also relevant to
Related
READ CDC's Vision for AI in Public Health
READ Considerations for Generative AI in Public Health
READ Considerations for Agentic Research in Public Health
READ FedScoop: CDC releases AI strategy with eye toward agentic uses
READ
Evidence reviewPeer reviewed

Natural Language Processing and Social Determinants of Health in Mental Health Research: AI-Assisted Scoping Review

AI-assisted scoping review of 1,768 NLP studies in mental health research, asking how often demographic and social determinants of health data are actually used or reported. Finds clinical notes and social media dominate, with social determinants often missing, which limits what these methods can say about populations.

Scherbakov, D. A., Hubig, N. C., Lenert, L. A., Alekseyenko, A. V., & Obeid, J. S. (2025). Natural Language Processing and Social Determinants of Health in Mental Health Research: AI-Assisted Scoping Review. JMIR Mental Health, 12, e67192. https://doi.org/10.2196/67192

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Promises and Pitfalls of Internet Search Data in Mental Health: Critical Review

Critical review of individual- and population-level internet search data for mental health research: associations with depression, anxiety, suicide and seasonal patterns, and the limits of these signals. Balanced on both promise and pitfalls.

Loch, A. A., & Kotov, R. (2025). Promises and Pitfalls of Internet Search Data in Mental Health: Critical Review. JMIR Mental Health, 12, e60754. https://doi.org/10.2196/60754

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Utilizing Big Data From Google Trends to Map Population Depression in the United States: Exploratory Infodemiology Study

Exploratory infodemiology study using Google Trends to map population depression across US states, proposed as a low-cost complement to survey epidemiology for service planning.

Wang, A., McCarron, R., Azzam, D., Stehli, A., Xiong, G., & DeMartini, J. (2022). Utilizing Big Data From Google Trends to Map Population Depression in the United States: Exploratory Infodemiology Study. JMIR Mental Health, 9(3), e35253. https://doi.org/10.2196/35253

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Methodologies for Monitoring Mental Health on Twitter: Systematic Review

Systematic review of 164 papers using Twitter data and machine learning to predict mental health outcomes. Finds only 16 of 119 datasets had externally validated ground truth, a sobering check on the population surveillance claims often made for social media AI.

Di Cara, N. H., Maggio, V., Davis, O. S. P., & Haworth, C. M. A. (2023). Methodologies for Monitoring Mental Health on Twitter: Systematic Review. Journal of Medical Internet Research, 25, e42734. https://doi.org/10.2196/42734

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Mental Health Risk Detection From Social Media Text Data: A Scoping Review of the Machine Learning Research Landscape

Scoping review mapping the recent machine learning landscape for mental health risk detection from social media text: target outcomes, data sources, labelling and evaluation practices. Useful for understanding how varied and unstandardised this field remains.

He, Y., He, Y., & Liu, D. (2026). Mental Health Risk Detection From Social Media Text Data: A Scoping Review of the Machine Learning Research Landscape. PsyCh Journal, 15(3), e70100. https://doi.org/10.1002/pchj.70100

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

UNICEF: Children are adopting AI technologies more than three times faster than adults (2026)

UNICEF analysis of household survey data from ten countries estimating that at least 20 million children have used AI, about 2 million of them for advice about things that worry them. Frames children's AI use as a global experiment with unknown effects on development and emotional wellbeing and calls for protections, research and child-centred design.

UNICEF. (2026, June 30). Children are adopting AI technologies more than three times faster than adults [Press release]. https://www.unicef.org/eca/press-releases/children-are-adopting-ai-technologies-more-three-times-faster-adults

Open resource
Continuum
Related
READ UNICEF Innocenti: Snapshot of AI usage and concerns among children and parents (June 2026)
READ UNICEF: Guidance on AI and children (2025 edition)
READ
Guidance and standardsNot reviewed

PAHO and IDB: Artificial Intelligence in Public Health, Readiness Assessment Toolkit (2024)

Toolkit for health ministries to assess and build preparedness for integrating AI into public health functions, produced by the Pan American Health Organization with the Inter-American Development Bank. Not mental health specific, but the closest real document to a regional public health AI guideline for the Americas.

Pan American Health Organization, & Inter-American Development Bank. (2024). Artificial intelligence in public health: Readiness assessment toolkit (Version 2.0). PAHO. https://www.paho.org/sites/default/files/2024-08/ai-ra-tool-rev-final-eng-aug-1_0.pdf

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Artificial intelligence in public health: promises, challenges, and an agenda for policy makers and public health institutions

Lancet Public Health viewpoint from Dimitra Panteli, Martin McKee and colleagues at the European Observatory and WHO Europe. Sets out where AI can help public health (surveillance, risk-factor monitoring, environmental health, resource allocation, tailored communication, misinformation detection, routine automation) and what stands in the way: equity, bias, explainability, privacy, cybersecurity, human oversight, infrastructure, workforce skills and governance. Notes that in a 2023 WHO Europe survey only a third of countries had a policy on big data and advanced analytics. The most cited agenda-setting piece on the topic.

Panteli, D., Adib, K., Buttigieg, S., Goiana-da-Silva, F., Ladewig, K., Azzopardi-Muscat, N., Figueras, J., Novillo-Ortiz, D., & McKee, M. (2025). Artificial intelligence in public health: promises, challenges, and an agenda for policy makers and public health institutions. The Lancet Public Health, 10(5), e428-e432. https://doi.org/10.1016/S2468-2667(25)00036-2

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Opportunities and challenges of artificial intelligence in public health: a systematic review on technological efficacy, ethical dilemmas, and governance pathways

PRISMA systematic review of 136 studies (2019 to January 2025) across English and Chinese databases. Finds AI improving epidemic surveillance, emergency response, health communication, decision support and drug and vaccine development, against risks of algorithmic bias, privacy, widening inequities and loss of public trust. About 70 percent of the evidence comes from high-income countries. Recommends explainable AI, strong data governance and participatory, adaptive oversight.

Gao, Q., Chen, L., & Huang, Z. (2026). Opportunities and challenges of artificial intelligence in public health: a systematic review on technological efficacy, ethical dilemmas, and governance pathways. Frontiers in Public Health, 13, 1748797. https://doi.org/10.3389/fpubh.2025.1748797

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Impact, use, and implications of artificial intelligence in public health decision making by elected officials: a scoping review

Scoping review asking how AI has actually supported public health decisions by officials and elected representatives. From 13,239 records only 7 studies qualified, all about COVID-19 (vaccination decision support, contact tracing, quarantine and movement restrictions), and all descriptive with no outcome evaluation. A sober reminder of how thin the evidence for AI in real public health governance still is.

Campbell, E. A., Goodtree, H., Gillani, S., Oyefolu, O., Kelly, A., Rivers, C., & Watson, C. (2026). Impact, use, and implications of artificial intelligence in public health decision making by elected officials: a scoping review. Frontiers in Public Health, 14, 1745684. https://doi.org/10.3389/fpubh.2026.1745684

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewed

Why We Need More Investment in AI for Prevention and Public Health (Community Impact Policy Institute, June 2026)

White paper by Jay Varma, former New York City public health lead, arguing that AI for prevention (surveillance, outbreak response, chronic disease prevention, vaccination campaigns, environmental monitoring, communications) is the largest under-funded health opportunity of a generation. Advocacy rather than evidence, but a clear statement of the case for using AI upstream of the clinic.

Varma, J. K. (2026, June 15). Why we need more investment in AI for prevention and public health: The largest health opportunity in a generation is under-funded [White paper]. Community Impact Policy Institute, Fedcap. https://fedcapgroup.org/news/new-report-presents-a-strategic-framework-for-the-future-of-ai-in-prevention-and-public-health

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

Conversational, Longitudinal, Ecological Assessment (CLEA): Exploring a new AI-driven method for qualitative data collection in a behavioural health context

Preprint introducing CLEA, an AI-driven conversational method for longitudinal qualitative data collection in behavioural health.

Downes, S., Krys, T., O’Hara, K., Western, M., Thompson, L., & Brigden, A. (2026). Conversational, Longitudinal, Ecological Assessment (CLEA): Exploring a new AI-driven method for qualitative data collection in a behavioural health context. . https://doi.org/10.64898/2026.01.20.26344494

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Innovation of Wearable Devices for Public Health Monitoring

Describes decentralisation of health monitoring from institutions to individuals and communities via IoT, AI and sensors.

Aini, I. N. (2026). Innovation of Wearable Devices for Public Health Monitoring. Journal of Public Health Indonesian, 2(5), 99-106. https://doi.org/10.62872/d9j65n53

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Collective Well-Being to Improve Population Health Outcomes: An Actionable Conceptual Model and Review of the Literature

Actionable conceptual model of collective wellbeing for improving population health outcomes, spanning built and social environment, access to services and social determinants.

Roy, B., Riley, C., Sears, L., & Rula, E. Y. (2018). Collective Well-Being to Improve Population Health Outcomes: An Actionable Conceptual Model and Review of the Literature. American Journal of Health Promotion, 32(8), 1800-1813. https://doi.org/10.1177/0890117118791993

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Advancing positive social determinants of health through collective impact and the 100% New Mexico Model

Describes the 100% New Mexico model: collective impact to assess access to ten service sectors, disaggregate disparities and build a shared agenda on social determinants.

McCrae, J. S., & Spain, A. K. (2023). Advancing positive social determinants of health through collective impact and the 100% New Mexico Model. Archives of Public Health, 81(1), 109. https://doi.org/10.1186/s13690-023-01120-4

Open resource
Continuum
Also relevant to
Related
READ Preprint version (Research Square)
READ
Opinion and commentaryPeer reviewed

Mental Health Priorities in the Arab World: Insights From Jordan

Policy column on mental-health needs and services in Jordan, identifying digital health technology, integrated care and youth-focused interventions as areas for investment. Relevant background for planning services in their social and regional context.

Ben-Zeev, D., Larsen, A., Beaulieu, A., Guler, J., & Bawaneh, A. (2024). Mental Health Priorities in the Arab World: Insights From Jordan. Psychiatric Services. https://doi.org/10.1176/appi.ps.20230365

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Digital Twins for Public-Health Emergency Planning: A Conceptual Framework

Perspective proposing digital twins for prevention, preparedness, response and recovery, including simulation, early warning and resource allocation. A potential population-health planning approach, not an evaluated model of urban design's effects on stress or mental health.

Sun, T., Lu, J., Li, Q., & Kou, Y. (2025). Can digital twin technology revolutionize public health emergency management? Insights from the PPRR framework. Frontiers in Public Health, 13, 1631339. https://doi.org/10.3389/fpubh.2025.1631339

Open resource
Continuum
READ
ResearchPreprint, not yet reviewed

Do VLMs See What Sensors Feel? A Scalable Expert-Guided Design for Wheelchair Accessibility Assessment from Street View

A technical preprint testing an expert-guided vision-language-model framework for assessing wheelchair accessibility from Google Street View. At 407 University of Florida locations, model ratings showed partial alignment with GPS-derived wheelchair dwell behaviour. The study suggests a possible way to screen environmental access barriers, while identifying limitations with subtle surfaces, temporary obstacles and viewpoints.

Wang, D., Hagen, A., Gatmaitan, I., Zhou, H., Dong, Y., Valipoor, S., Wong, V. W. H., & Li, L. (2026). Do VLMs See What Sensors Feel? A Scalable Expert-Guided Design for Wheelchair Accessibility Assessment from Street View. arXiv:2606.07642, version 1. https://arxiv.org/abs/2606.07642

Open resource
Continuum
Also relevant to
READ
Tool or datasetNot reviewed

Urban HEART: Urban Health Equity Assessment and Response Tool

WHO's urban health-equity assessment and response framework supports identifying, prioritising and tracking inequities in health and its social determinants. It provides a population-planning foundation for comparing places and directing action. It is not an AI tool or evidence that WHO used AI satellite analysis to identify mental-health deserts.

World Health Organization & WHO Centre for Health Development (Kobe, Japan). (2010). Urban HEART: Urban Health Equity Assessment and Response Tool. World Health Organization. ISBN 9789241500142. https://iris.who.int/items/9560b26b-5c44-483b-8f8c-ff7f79ccf61e

Open resource
Continuum
Also relevant to
Related
READ WHO expert consultation on Urban HEART, 2013
READ WHO Health Inequality Data Repository and HEAT update, 2026
READ
ResearchNot reviewed

The role of virtual services and AI in mental health care across Canada

This Mental Health Commission of Canada report examines virtual services and AI in mental health care. It reports that about one in seven people in Canada used AI mental health tools and that, although 76% of users found them effective, only 14% trusted them; respondents emphasised human involvement, privacy protections and oversight.

Abbott, M., Kalenteridis, K., Thraya, Z., Cooper, M., Hlavacek, C., Yakhnis, I., & Martin, L. (2026). The role of virtual services and AI in mental health care across Canada. Mental Health Commission of Canada. https://mentalhealthcommission.ca/wp-content/uploads/2026/06/The-Role-of-Virtual-Services-and-AI-in-Mental-Health-Care-Across-Canada-1.pdf

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

Around 1 in 5 young people use AI chatbots for mental health advice, survey finds

This NBC News article reports on a RAND survey of young people aged 12 to 21. It says 19% had used AI chatbots for mental health advice and reports the study lead author's call for regulation.

Bendix, A., & Yang, A. (2026, June 1). Around 1 in 5 young people use AI chatbots for mental health advice, survey finds. NBC News. https://www.nbcnews.com/health/mental-health/ai-chatbots-mental-health-advice-young-people-rcna347758

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Exploring real-world use of AI chatbots for mental health support: Cross-sectional survey study

This Belgian survey examined 349 people who used AI chatbots for mental health support, including 73 students recruited specifically because they used AI for this purpose. Users often chose freely available chatbots such as ChatGPT for their accessibility and constant availability, while students also valued anonymity and people in the general-public group reported feeling supported.

Bernaerts, S., Buelens, F., De Witte, N. A. J., Balaskas, A., Colman, T., & Van Daele, T. (2026). Exploring real-world use of AI chatbots for mental health support: Cross-sectional survey study. JMIR Mental Health, 13, e104316. https://doi.org/10.2196/104316

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

RAND: Definitions drive estimates of chatbot mental health use

This news article reports an analysis of publicly available ChatGPT conversations using an AI-based classifier. It says estimates of mental health use vary depending on how that use is defined.

Canady, V. A. (2026). RAND: Definitions drive estimates of chatbot mental health use. Mental Health Weekly, 36(34), 3–5. https://doi.org/10.1002/mhw.34992

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Artificial intelligence in mental healthcare: A review of health economics, sustainability, and global health perspectives

This narrative review examines AI in mental health, including digital interventions, prediction tools and telepsychiatry. It finds that AI may support earlier detection, tailored care and access to services, but warns of privacy risks, bias, limited transparency and the potential to worsen inequalities.

Chyczij, F., Paixão, M., & Sara, D. G. (2026). Artificial intelligence in mental healthcare: A review of health economics, sustainability, and global health perspectives. Global Health Economics and Sustainability, 4(3), 026260038. https://doi.org/10.36922/GHES026260038

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

Neighbourhood health framework

England's framework for neighbourhood care, including risk stratification and data for proactive care, with AI-assisted triage pilots planned. Shows population data and AI being written into place-based service design. No results reported.

Department of Health and Social Care & NHS England. (2026, March 17). Neighbourhood health framework. GOV.UK. https://www.gov.uk/government/publications/neighbourhood-health-framework/neighbourhood-health-framework

Open resource
Continuum
Also relevant to
READ
ResearchNot reviewed

Americans and AI 2026: Chatbots, smart devices and views on impact

This Pew Research Center report presents findings on Americans’ use of and views about AI. It reports that 10% of US adults use chatbots for emotional support or advice and 4% for companionship.

Gottfried, J., Bishop, W., Anderson, M., Faverio, M., Park, E., & McClain, C. (2026, June 17). Americans and AI 2026: Chatbots, smart devices and views on impact. Pew Research Center. https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewed

"Where do I start?": How governments can prioritise AI solutions for health

A framework for governments choosing AI investments in health. Covers clinician training through simulation, service planning from geographic and social data, and wearable-based advice. Useful for showing all three areas in one policy frame.

Why it matters now: The framework brings clinician training, geographic service planning and wearable-based advice into one government AI investment guide.

Harvard, J., McLauchlan, K., Milestone, D., & Ubaldi, B. (2026, April 27). "Where do I start?": How governments can prioritise AI solutions for health. Tony Blair Institute for Global Change. https://institute.global/insights/public-services/where-do-i-start-how-governments-can-prioritise-ai-solutions-for-health

Open resource
Continuum
Also relevant to
READ
ResearchNot reviewed

Talking to machines: Children's experiences with AI assistants and companions

This eSafety Commissioner report describes children’s experiences with AI assistants and companions. Among children who had used one, 20% used it for mental-health or wellbeing advice and 20% reported potentially harmful or inappropriate interactions.

Herbert, L., Lister, C., Minihan, S., Najdovska, A., & Nicholas, M. (2026). Talking to machines: Children's experiences with AI assistants and companions. eSafety Commissioner. https://www.esafety.gov.au/research/talking-to-machines-childrens-experiences-with-ai-assistants-and-companions

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Methods for mapping population-level interventions in system dynamics modelling for context-sensitive policy planning in youth mental health

Describes how local evidence and stakeholder knowledge were combined with simulation scenarios for regional youth mental health planning. Scenario exploration shifted discussion towards combining interventions and collaboration. Shows place-based knowledge feeding the model, not the other way round.

Loblay, V., Ianelli, O., Hindmarsh, G., Song, Y. J. C., Huntley, S., Piper, S., Crosland, P., Hosseini, S. H., Ho, N., Nguyen, K.-H., Vacher, C., Hutcheon, A., Rosenberg, S., McNeill, A., Oakes, H., Occhipinti, J.-A., & Hickie, I. B. (2026). Methods for mapping population-level interventions in system dynamics modelling for context-sensitive policy planning in youth mental health. International Journal of Mental Health Systems. Advance online publication. https://doi.org/10.1186/s13033-026-00723-y

Open resource
Continuum
Also relevant to
READ
ResearchNot reviewed

Key findings from MHA's survey on AI and mental health

This webpage presents findings from Mental Health America’s survey on AI and mental health. It reports that 53% of respondents had used AI for emotional support and that 68% of those users rated the experience positively.

Mental Health America. (2026). Key findings from MHA's survey on AI and mental health. https://mhanational.org/key-findings-from-mhas-survey-on-ai-and-mental-health/

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Millions of people in Canada are finding AI-enabled support for mental health effective amid ongoing questions around trust

This Mental Health Commission of Canada news release reports that about six million people in Canada used AI-enabled mental health support in the past year. It says most users found the tools effective, while trust remains low and the Commission calls for safe, equitable, evidence-based guidance.

Why it matters now: The Commission reports widespread use and perceived effectiveness, while trust concerns make safe, equitable guidance timely.

Mental Health Commission of Canada. (2026, July 8). Millions of people in Canada are finding AI-enabled support for mental health effective amid ongoing questions around trust [News release]. https://mentalhealthcommission.ca/news-releases/millions-of-people-in-canada-are-finding-ai-enabled-support-for-mental-health-effective-amid-ongoing-questions-around-trust/

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Draft mental health and wellbeing strategy 2026–2036

This is a New Zealand Ministry of Health draft strategy for mental health and wellbeing from 2026 to 2036. It proposes increasing access to digital supports, including appropriate AI tools, and using emerging technologies responsibly.

Why it matters now: The draft strategy proposes expanding digital mental health support, including appropriate AI tools, while emphasising responsible use.

Ministry of Health. (2026). Draft mental health and wellbeing strategy 2026–2036. New Zealand Government. https://www.health.govt.nz/system/files/2026-04/Draft-Mental-Health-and-Wellbeing-Strategy-2026-2036.pdf

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

The promises and perils of using LLMs for effective public services

Worked with a Canadian child-welfare agency to test language models for tracking case progress and spotting service gaps. The models missed case trajectories that needed human judgement. Also in CHI 2026.

Why it matters now: A child-welfare agency case study found language models missed case trajectories needing human judgement.

Moon, E. S.-Y., Tamura, M., Zhai, A., Habib, N., Shirazi, B., Kassam, A., Saxena, D., & Guha, S. (2026). The promises and perils of using LLMs for effective public services [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2601.15163

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

National report card 2025: Environmental scan

This Australian Government environmental scan accompanies the National Mental Health Commission’s 2025 report card. The scan cites research on digital inclusion and AI use, including evidence that some people seek social connection or are emotionally vulnerable when using AI chatbots.

Why it matters now: The environmental scan connects mental health system planning with evidence on digital inclusion and AI use.

National Mental Health Commission. (2026). National report card 2025: Environmental scan. Australian Government. https://www.mentalhealthcommission.gov.au/sites/default/files/2026-08/national-report-card-2025-environmental-scan_0.pdf

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Scaling artificial intelligence in health

This OECD policy report examines how countries can adopt AI in health responsibly and at scale. It reports that all OECD Member countries use AI in administration, but only 10% have scaled it nationally for medical imaging, and argues that better data, policy and governance are needed.

Why it matters now: The OECD report addresses responsible AI scale-up and identifies mental health as comparatively underserved by health AI innovation.

Organisation for Economic Co-operation and Development. (2026). Scaling artificial intelligence in health. OECD Publishing. https://doi.org/10.1787/a436e12d-en

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Proceeding report from Madrid conference on the responsible scale of AI in healthcare: Scaling fast while doing no harm

This OECD report records proceedings from a Madrid conference on responsible AI scale-up in healthcare. It notes that mental health remains comparatively underserved by health AI innovation and calls for closer alignment with health-system and population needs.

Why it matters now: The conference report argues that health AI should scale responsibly and better reflect health-system and population needs.

Sutherland, E., & Vujovic, K. (2026). Proceeding report from Madrid conference on the responsible scale of AI in healthcare: Scaling fast while doing no harm (OECD Artificial Intelligence Papers No. 66). OECD Publishing. https://www.oecd.org/content/dam/oecd/en/publications/reports/2026/10/proceeding-report-from-madrid-conference-on-the-responsible-scale-of-ai-in-healthcare_cd7819f4/cf9439c3-en.pdf

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewedIndustry funded

AI for healthy cities: How artificial intelligence can support cardiometabolic disease prevention and enhance resilience

Examples of AI linking health records with social and environmental data to find neighbourhood-level risk, such as commute length and broadband access in New York. Not mental health, but the clearest worked examples of AI place-based scanning. Funded by the Novartis Foundation.

Why it matters now: It offers neighbourhood-level examples of AI linking health, social and environmental data, while noting industry funding.

Wilson, M., & Juani, A. (2026, January). AI for healthy cities: How artificial intelligence can support cardiometabolic disease prevention and enhance resilience. FP Analytics. https://fpanalytics.foreignpolicy.com/2026/01/16/artificial-intelligence-healthy-cities/

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewed

Governing in the age of AI: Reimagining local government

Uses examples such as Barking and Dagenham's One View, which linked council data to support 1,000 households in debt and identified 93.6% of people who needed to shield during COVID-19. Shows locality data being used to reach people early. Advocacy estimates, not evaluations.

Why it matters now: The report describes local-government AI examples for identifying residents who may need support, while noting the estimates are not evaluations.

Atkinson, I., Boroda, N., & Gottesman, R. (2025, May 7). Governing in the age of AI: Reimagining local government. Tony Blair Institute for Global Change. https://institute.global/insights/politics-and-governance/governing-in-the-age-of-ai-reimagining-local-government

Open resource
Continuum
Also relevant to
READ
Evidence reviewPreprint, not yet reviewed

Artificial intelligence in rural healthcare delivery: Bridging gaps and enhancing equity through innovation

Reviews 109 studies on AI in rural health, finding promise for predictive analytics and telehealth to improve access. Relevant to rural and remote Australian planning, though not mental health specific.

Why it matters now: A preprint review of 109 rural-health AI studies considers access gains, though it is not mental-health specific.

Balakrishnan, K., Velusamy, D., Hinkle, H. E., Li, Z., Ramasamy, K., Khan, H., Ramaswamy, S., & Shah, P. M. (2025). Artificial intelligence in rural healthcare delivery: Bridging gaps and enhancing equity through innovation [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2508.11738

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

The use of geographic information systems (GIS) in studying mental health service delivery: A scoping review

Maps 58 studies using geographic tools to show where services are, who can reach them and where the gaps are; most examined access. The groundwork AI-based locality tools build on.

Biswal, B., Paimapari, R., Suresh, A., Pacheco, M. G., Fernandes, L., Gandhi, Y., Patel, V., Singla, D. R., Fernandes, A., Velleman, R., Lu, C., Grundy, C., Bhatia, U., & Nadkarni, A. (2025). The use of geographic information systems (GIS) in studying mental health service delivery: A scoping review. Cambridge Prisms: Global Mental Health, 12, Article e132. https://doi.org/10.1017/gmh.2025.10088

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Examining the impact of youth mental health services capacity growth trajectories and digital interventions on youth mental health outcomes: System dynamics modeling analysis

A Western Sydney model projected that digital interventions plus growth in specialist services could cut years lived with symptoms by 15% over ten years. Digital tools alone could not prevent decline if specialist services shrank. Simulation modelling, not generative AI.

Hosseini, S. H., Ho, N., Huntley, S., Piper, S., Crosland, P., Skinner, A., Vacher, C., Tran, K., Nguyen, K.-H., Song, Y. J. C., Loblay, V., Iannelli, O., Ferdousi, S., Munasinghe, S., Rao, S., Lam, D., Waterhouse-Bushnell, Z., Clancey, G., Hickie, I. B., & Occhipinti, J.-A. (2025). Examining the impact of youth mental health services capacity growth trajectories and digital interventions on youth mental health outcomes: System dynamics modeling analysis. Journal of Medical Internet Research, 27, Article e71256. https://doi.org/10.2196/71256

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Social determinants of health extraction from clinical notes across institutions using large language models

Tested social-circumstance extraction across four institutions, with high accuracy (F1 above 0.90 for the main factors). Performance still dropped when a model trained in one place was used in another, so local checking is needed.

Keloth, V. K., Selek, S., Chen, Q., Gilman, C., Fu, S., Dang, Y., Chen, X., Hu, X., Zhou, Y., He, H., Fan, J. W., Wang, K., Brandt, C., Tao, C., Liu, H., & Xu, H. (2025). Social determinants of health extraction from clinical notes across institutions using large language models. npj Digital Medicine, 8, Article 287. https://doi.org/10.1038/s41746-025-01645-8

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewed

Learn fast and build things: Lessons from six years of studying AI in the public sector

Lessons from public sector AI, including a London council's predictive tools for identifying at-risk residents. Staff accepted COVID-19 case management tools but were wary of risk alerts that lacked clear explanations. A caution for locality risk tools.

Why it matters now: It shows staff accepted case-management tools but questioned unexplained risk alerts, a live caution for locality risk systems.

Parker, I., Studman, A., & Jones, E. (2025, March 14). Learn fast and build things: Lessons from six years of studying AI in the public sector. Ada Lovelace Institute. https://www.adalovelaceinstitute.org/policy-briefing/public-sector-ai/

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewed

Health and AI: Advancing responsible and ethical AI for all communities

Argues AI and wearable data can inform personalised care only if planning accounts for local needs, resources and barriers. Warns that third-party data insights may lack cultural and local context.

Why it matters now: It highlights that wearable and third-party insights need local context as health AI planning expands.

Turner Lee, N., Simmons, M., & Crawford, M. (2025, March 3). Health and AI: Advancing responsible and ethical AI for all communities. Brookings Institution. https://www.brookings.edu/articles/health-and-ai-advancing-responsible-and-ethical-ai-for-all-communities/

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

Community needs and assets: A computational analysis of community conversations

A language model sorted community conversations into needs and assets with 94% F1, far above a general language model (61%). Points to AI-assisted asset mapping, not just deficit mapping. Data were Reddit posts only.

Chowdhury, M. T. A., Sharma, N., & KhudaBukhsh, A. R. (2024). Community needs and assets: A computational analysis of community conversations [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2403.13272

Open resource
Continuum
READ
ResearchPeer reviewed

Large language models to identify social determinants of health in electronic health records

Language models found social circumstances such as housing, transport and support in clinical notes. The best model identified 45 of 48 patients with an adverse social need, against 1 of 48 picked up by standard coding. Shows AI can surface a person's life circumstances that records usually miss.

Guevara, M., Chen, S., Thomas, S., Chaunzwa, T. L., Franco, I., Kann, B. H., Moningi, S., Qian, J. M., Goldstein, M., Harper, S., Aerts, H. J. W. L., Catalano, P. J., Savova, G. K., Mak, R. H., & Bitterman, D. S. (2024). Large language models to identify social determinants of health in electronic health records. npj Digital Medicine, 7, Article 6. https://doi.org/10.1038/s41746-023-00970-0

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Can computer simulation support strategic service planning? Modelling a large integrated mental health system on recovery from COVID-19

Simulated a whole mental health system in south-west England and projected GP and community services under pressure of up to 50% and 150% respectively, then tested capacity changes. Shows planners using simulation to test options before committing. Simulation modelling, not generative AI.

Pierotti, L., Cooper, J., James, C., Cassels, K., Gara, E., Denholm, R., & Wood, R. (2024). Can computer simulation support strategic service planning? Modelling a large integrated mental health system on recovery from COVID-19. International Journal of Mental Health Systems, 18, Article 12. https://doi.org/10.1186/s13033-024-00623-z

Open resource
Continuum
READ
ResearchPeer reviewed

Improving access to mental health care: A system dynamics model of direct access to specialist care and accelerated specialist service capacity growth

A NSW model projected that faster specialist capacity growth plus direct access would avert 3.6% of mental health emergency presentations and 2.1% of suicide deaths over 2021 to 2028. Simulation modelling, not generative AI.

Vacher, C., Skinner, A., Occhipinti, J.-A., Rosenberg, S., Ho, N., Song, Y. J. C., & Hickie, I. B. (2023). Improving access to mental health care: A system dynamics model of direct access to specialist care and accelerated specialist service capacity growth. Medical Journal of Australia, 218(7), 309–314. https://doi.org/10.5694/mja2.51903

Open resource
Continuum
READ
Evidence reviewPeer reviewed

Mental health systems modelling for evidence-informed service reform in Australia

Sets out how systems modelling can help Australian planners handle uncertain future demand and choose investments, including at regional level. A foundation piece for computational planning. This is simulation modelling, not generative AI.

Whiteford, H., Bagheri, N., Diminic, S., Enticott, J., Gao, C. X., Hamilton, M., Hickie, I. B., Le, L. K.-D., Lee, Y. Y., Long, K. M., McGorry, P., Meadows, G., Mihalopoulos, C., Occhipinti, J.-A., Rock, D., Rosenberg, S., Salvador-Carulla, L., & Skinner, A. (2023). Mental health systems modelling for evidence-informed service reform in Australia. Australian & New Zealand Journal of Psychiatry, 57(11), 1417–1427. https://doi.org/10.1177/00048674231172113

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Measuring perceived psychological stress in urban built environments using Google Street View and deep learning

Deep learning applied to 31,378 street view images mapped perceived stress across a Seoul district, with stress clustered in dense residential and commercial areas. An example of AI reading the environment people live in, though perceived stress is rated from images, not measured in residents.

Han, X., Wang, L., Seo, S. H., He, J., & Jung, T. (2022). Measuring perceived psychological stress in urban built environments using Google Street View and deep learning. Frontiers in Public Health, 10, Article 891736. https://doi.org/10.3389/fpubh.2022.891736

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

A machine learning approach to small area estimation: Predicting the health, housing and well-being of the population of Netherlands

Machine learning estimated health, housing and wellbeing for every Dutch neighbourhood from survey and administrative data, more accurately and faster than the standard statistical model. A template for local wellbeing profiles where surveys are too thin.

Viljanen, M., Meijerink, L., Zwakhals, L., & van de Kassteele, J. (2022). A machine learning approach to small area estimation: Predicting the health, housing and well-being of the population of Netherlands. International Journal of Health Geographics, 21, Article 4. https://doi.org/10.1186/s12942-022-00304-5

Open resource
Continuum
READ
Practice signalNot reviewed

A local navigation tool for mental health care (MChart)

An ACT planning platform (2022 to 2025) combining service directories, workforce data, mapping, service use and community indicators so planners can see where services are and where gaps lie. A rare Australian locality planning tool for mental health.

Digital Health CRC. (n.d.). A local navigation tool for mental health care (MChart). Retrieved October 17, 2026, from https://digitalhealthcrc.com/projects/a-local-navigation-tool-for-mental-health-care-mchart/

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

The future is now: Public health and artificial intelligence

A local government view of AI in public health: predicting homelessness risk, sensors supporting older residents at home, and wearables alerting services. Notes the field is still early. Commentary.

Fothergill, D. (n.d.). The future is now: Public health and artificial intelligence. Local Government Association. Retrieved October 17, 2026, from https://www.local.gov.uk/topics/social-care-health-and-integration/future-now-public-health-and-artificial-intelligence

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

Building an integrated care system intelligence function

Guidance on using linked population data and analytics to understand local need, identify inequalities and plan services. The system infrastructure AI planning tools sit on. Not mental health specific and gives no outcome evidence.

NHS England. (n.d.). Building an integrated care system intelligence function. Retrieved October 17, 2026, from https://www.england.nhs.uk/long-read/building-an-ics-intelligence-function/

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewed

The reality of, and potential for, digitally enabled care in the community

Looks at digital tools for community care, including AI matching people's needs to technology and rostering that uses local demographics to plan staffing. A grounded view of what works in practice.

The King's Fund. (n.d.). The reality of, and potential for, digitally enabled care in the community. Retrieved October 17, 2026, from https://www.kingsfund.org.uk/insight-and-analysis/long-reads/the-reality-of-and-potential-for-digitally-enabled-care-in-the-community

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

AI Models for Suicide Risk Prediction in Adult Patients Receiving Mental Health Care Using Real-World Data: Retrospective Population-Based Study

This retrospective population-based study used real-world data to assess suicide risk prediction in 32,661 adults receiving mental health care, including 2,764 with suicidal behaviour. GradientBoosting and XGBoost performed best, with ROC-AUC scores around 0.95 and recall scores of 0.85 and 0.84; leading predictors included corticosteroid use, olanzapine, lorazepam, female sex and depression.

Peña-Gómez, C.; Fradera, M.; Sánchez Corrales, X. et al. (2026). AI Models for Suicide Risk Prediction in Adult Patients Receiving Mental Health Care Using Real-World Data: Retrospective Population-Based Study. JMIR AI. https://doi.org/10.2196/85085

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Integrating speech biomarkers and large language models for adolescent suicide risk detection with mobile application for real-world evaluation

The study combined speech biomarkers with a large language model to detect suicide risk in adolescents aged 10-18, using a development cohort of 1,223 and an independent mobile-app validation cohort of 460. On the self-introduction task, the model achieved 0.808 accuracy and a macro-F1 score of 0.807, and the authors report it remained effective in real-world mobile assessment.

Lei, C.; Cui, Z.; Duan, Y. et al. (2026). Integrating speech biomarkers and large language models for adolescent suicide risk detection with mobile application for real-world evaluation. Cell Reports Medicine. https://doi.org/10.1016/j.xcrm.2026.102823

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Large language models in adolescent suicide prevention: from language signals to accountable action

This mini review examines research published from January 2023 to June 2026 on large language models for suicide prevention among adolescents, defined mainly as ages 10-19, and finds the strongest evidence for supervised risk-signal flagging. Evidence for summarisation and information extraction is preliminary, while autonomous crisis intervention lacks support; the authors found no prospective deployment trials measuring alert-to-action latency, escalation appropriateness or linked outcomes.

Li, M., Huang, W., Wu, J. et al. (2026). Large language models in adolescent suicide prevention: from language signals to accountable action. Frontiers in Public Health. https://doi.org/10.3389/fpubh.2026.1872450

Open resource
Continuum
Also relevant to

B3. Triage, navigation and stepped care

Can AI help people reach the right level of support sooner?

READ
Evidence reviewPeer reviewed

Artificial Intelligence-Driven Tools in Mental Health Service Delivery: A Scoping Review

Scoping review of AI tools in mental health service delivery, mapping where they sit in the care pathway and what implementation evidence exists.

Woo, Y., & Jung, K. (2026). Artificial Intelligence-Driven Tools in Mental Health Service Delivery: A Scoping Review. Healthcare, 14(7), 943. https://doi.org/10.3390/healthcare14070943

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Care navigation addresses issues of tele-mental health acceptability and uptake in rural and remote Australian communities

Australian study of how care navigation improves acceptability and uptake of tele-mental health in rural and remote communities.

Fisher, O. J., McGrath, K., Grogan, C., Cockshaw, W., & Leggatt-Cook, C. (2024). Care navigation addresses issues of tele-mental health acceptability and uptake in rural and remote Australian communities. PLOS ONE, 19(4), e0298655. https://doi.org/10.1371/journal.pone.0298655

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

A conceptual multi-agent architecture for mental health triage in post-conflict Arabic-speaking populations: a theoretical proposition and staged validation argument

Conceptual proposal for a multi-agent AI triage architecture for post-conflict Arabic-speaking populations, motivated by Syria's shortage of psychiatrists. Argues that tools built for English-speaking, non-humanitarian settings do not transfer without redesign. Theoretical, not yet tested.

Shahin, A., & Masry, B. (2026). A conceptual multi-agent architecture for mental health triage in post-conflict Arabic-speaking populations: a theoretical proposition and staged validation argument. Frontiers in Psychiatry, 17, 1854701. https://doi.org/10.3389/fpsyt.2026.1854701

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Trustworthy artificial intelligence for rural health care

Australian perspective piece arguing that AI could support earlier identification of distress, safer triage and more timely care for regional, rural and remote communities, where suicide and self-harm rates are substantially higher than in cities. Proposes a minimum viable governance model built on Indigenous partnership, language and cultural fit, explicit safety nets and independent evaluation. Online ahead of print.

Kirkpatrick, E. R., Johnson, S. L., & Nevin, M. (2026). Trustworthy artificial intelligence for rural health care. Internal Medicine Journal, advance online publication. https://doi.org/10.1111/imj.70585

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Unveiling the mental health services gap: help-seeking and referral patterns in South Kashmir, India

Help-seeking and referral patterns that reveal the mental health services gap in South Kashmir, India.

Dar, M. A., Amin, R., Khan, R. R., Bashir, N., Zuhare, S. M., & Hyder, A. (2025). Unveiling the mental health services gap: help-seeking and referral patterns in South Kashmir, India. Middle East Current Psychiatry, 32(1), 60. https://doi.org/10.1186/s43045-025-00555-5

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Nursing Students' Perceptions of AI-Driven Mental Health Support and Its Relationship with Anxiety, Depression, and Seeking Professional Psychological Help: Transitioning from Traditional Counseling to Digital Support

Survey of nursing students' perceptions of AI-driven mental health support and how those perceptions relate to anxiety, depression and willingness to seek professional help. Prior use of AI support was a key predictor of acceptance.

Albikawi, Z., Abuadas, M., & Rayani, A. M. (2025). Nursing Students' Perceptions of AI-Driven Mental Health Support and Its Relationship with Anxiety, Depression, and Seeking Professional Psychological Help: Transitioning from Traditional Counseling to Digital Support. Healthcare, 13(9), 1089. https://doi.org/10.3390/healthcare13091089

Open resource
Continuum
READ
Opinion and commentaryPeer reviewed

We don't need more apps, we need connection: Recommender systems as under-explored chance to promote students' mental health at universities

Turns recommender technology around: instead of an engagement feed, a recommender that matches students to existing university supports after brief stress screening. A proposal that has not yet been evaluated, but a useful example of the same technology used for navigation rather than retention.

Apolinário-Hagen, J., Schobel, J., Helten, A.-K., Sahan, F., Pryss, R., & John, D. (2026). We don't need more apps, we need connection: Recommender systems as under-explored chance to promote students' mental health at universities. Frontiers in Psychology, 16, Article 1629265. https://doi.org/10.3389/fpsyg.2025.1629265

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Digital front door technologies to pre-assess people before assessment for NHS Talking Therapies: Early value assessment

An independent NIHR assessment of AI 'digital front door' tools for referral and pre-assessment. Evidence for Limbic Access supports more referrals, including from Asian, Black and non-binary groups, high satisfaction and shorter assessments. Evidence on accuracy, administrative burden and cost-effectiveness is still insufficient.

Fleeman, N., Mahon, J., Bryning, S., Beale, S., Boland, A., Greenhalgh, J., & Dundar, Y. (2026). Digital front door technologies to pre-assess people before assessment for NHS Talking Therapies: Early value assessment. Health Technology Assessment, 30(52). https://doi.org/10.3310/GJAB1822

Open resource
Continuum
READ
ResearchPreprint, not yet reviewed

HeyFriend Helper: A conversational AI web-app for resource access among low-income Chicago residents

A conversational assistant combining job tools with location-based access to community services for low-income residents. An early design study of AI connecting people to what is near them.

Juarez, M., Rai, A., Ravi, K. E., Delaney, M. C., Olweean, D., Klingensmith, E., Banerjee, S., Klingensmith, N., & Thiruvathukal, G. (2026). HeyFriend Helper: A conversational AI web-app for resource access among low-income Chicago residents [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2603.25800

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

HoosierHelp: Benchmarking LLM agents for social service navigation

Tested seven AI agents helping simulated users find services among 3,971 real Indiana resources. The agents were substantially unreliable, especially when people's needs were contradictory. A needed reality check for AI navigation.

Li, Y., Sun, W., Ma, T., Shi, K., Zhang, Z., & Ye, Y. (2026). HoosierHelp: Benchmarking LLM agents for social service navigation [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2608.09946

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewedLived experience

Large language models in peer-run community behavioral health services: Understanding peer specialists and service users' perspectives on opportunities, risks, and mitigation strategies

Workshops with peer specialists and service users explored where AI could help peer-run services, including resource finding, and where it threatens their values. Lived experience shaping the design brief from the start.

Peng, C., Chai, M., Mo, G., Raman, N., Tang, N., Pagdon, S., Swarbrick, M., Jones, N., Fang, F., & Shen, H. (2026). Large language models in peer-run community behavioral health services: Understanding peer specialists and service users' perspectives on opportunities, risks, and mitigation strategies. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (pp. 1–24). Association for Computing Machinery. https://doi.org/10.1145/3772318.3790339

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Perceptions and insights: A qualitative assessment of an AI-assisted psychiatric triage system implemented in an outpatient hospital setting

Patients in a Canadian outpatient psychiatry service felt AI-assisted triage could get them to care faster, provided clinicians stay involved and AI's role is explained clearly. They also flagged usability problems and the need for neurodiversity-informed design.

Knyahnytskyi, O., Eadie, J., Asadpour, K., Stephenson, C., Yang, M., Reshetukha, T., Moi, C., Barrett, T., Hicks, M., Gutierrez-Nunez, G., Kumar, A., Jagayat, J., Sajid, S., Patel, C., Holmes, C., Shirazi, A., Verter, V., Soares, C., Omrani, M., & Alavi, N. (2025). Perceptions and insights: A qualitative assessment of an AI-assisted psychiatric triage system implemented in an outpatient hospital setting. DIGITAL HEALTH, 11, Article 20552076251384835. https://doi.org/10.1177/20552076251384835

Open resource
Continuum
READ
ResearchPeer reviewed

ASHABot: An LLM-powered chatbot to support the informational needs of community health workers

A WhatsApp chatbot answered community health workers' questions in rural India over 107 days, answering 83.6% automatically with expert-rated accuracy of 85% or more. Shows AI as on-the-job support for the lay workforce. Not mental health specific.

Ramjee, P., Chhokar, M., Sachdeva, B., Meena, M., Abdullah, H., Vashistha, A., Nagar, R., & Jain, M. (2025). ASHABot: An LLM-powered chatbot to support the informational needs of community health workers. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (Article 698, pp. 1–22). Association for Computing Machinery. https://doi.org/10.1145/3706598.3713680

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Closing the accessibility gap to mental health treatment with a personalized self-referral chatbot

NHS Talking Therapies services using an AI self-referral chatbot saw referrals rise 15%, against 6% in matched services, across 129,400 people. Increases were largest for non-binary people (179%) and ethnic minority groups (29%), making this a key example of AI widening access rather than narrowing it.

Habicht, J., Viswanathan, S., Carrington, B., Hauser, T. U., Harper, R., & Rollwage, M. (2024). Closing the accessibility gap to mental health treatment with a personalized self-referral chatbot. Nature Medicine, 30(2), 595–602. https://doi.org/10.1038/s41591-023-02766-x

Open resource
Continuum
Also relevant to
READ
ResearchNot reviewedDeveloper research

Using conversational AI to facilitate mental health assessment and improve clinical efficiencies in psychotherapy services in a large real-world dataset

Across 64,862 referrals to nine NHS Talking Therapies services, AI-supported intake was associated with 12.7 minutes less assessment time, five fewer days waiting, fewer changes to treatment allocation and 18% less dropout. Developer-reported, so read alongside the independent NIHR assessment.

Rollwage, M., Habicht, J., Juchems, K., Carrington, B., Stylianou, M., Hauser, T., & Harper, R. (2024, January 9). Using conversational AI to facilitate mental health assessment and improve clinical efficiencies in psychotherapy services in a large real-world dataset. Limbic. https://www.limbic.ai/research/clinical-efficiencies

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Chatbot for social need screening and resource sharing with vulnerable families: Iterative design and evaluation study

DAPHNE screens families for social needs and links them to local resources, designed with input from 201 low-income households. Usability was acceptable (72/100). A rule-based chatbot, not generative AI.

Sezgin, E., Kocaballi, A. B., Dolce, M., Skeens, M., Militello, L., Huang, Y., Stevens, J., & Kemper, A. R. (2024). Chatbot for social need screening and resource sharing with vulnerable families: Iterative design and evaluation study. JMIR Human Factors, 11, Article e57114. https://doi.org/10.2196/57114

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

The use of artificial intelligence models for interpretation and triage of clinical referral letters to acute and specialist care pathways: Protocol for a systematic review

This peer-reviewed protocol plans a systematic review of AI models used to interpret and triage clinical referral documents, with searches across six databases for studies published from January 2016 to 2026-05-05. Screening was ongoing, and data extraction and synthesis had not begun, so the protocol reports no review findings.

Foley, J., Hession, E., Umana, E. et al. (2026). The use of artificial intelligence models for interpretation and triage of clinical referral letters to acute and specialist care pathways: Protocol for a systematic review. PLOS ONE. https://doi.org/10.1371/journal.pone.0355476

Open resource
Continuum
READ
ResearchPeer reviewed

Adoption of Digital Mental Health Interventions in National Health Service England, Scotland, and Wales: Freedom of Information Questionnaire Study

Freedom of Information requests identified 19 distinct digital mental health interventions procured, contracted, or paid for by NHS mental health providers: 15 were reported in England, 6 in Scotland, and 2 in Wales, with some used in multiple nations. Compared with 48 technologies reported in England in 2015, only SilverCloud and IESO were still reported as in use a decade later.

Hitcham, L.; Gómez Bergin, A.; Ito-Jaeger, S.; et al. (2026). Adoption of Digital Mental Health Interventions in National Health Service England, Scotland, and Wales: Freedom of Information Questionnaire Study. JMIR Mental Health. https://doi.org/10.2196/92187

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

NHS accelerates artificial intelligence rollout to cut waiting times and improve care for millions

NHS England plans to invest £10 billion over three years in health technology, digital and data systems, with expected benefits of £41 billion over the next decade. Its AI triage tool is due to reach more than 200,000 patients within 12 months and all NHS App users by April 2028; a trial found 29% fewer people queued by phone, while a study of AI note-taking tools found clinicians could spend nearly a quarter more time with patients.

Why it matters now: Practice signal: describes a rollout plan, not demonstrated effects.

NHS England (2026). NHS accelerates artificial intelligence rollout to cut waiting times and improve care for millions. NHS England. https://www.england.nhs.uk/2026/07/nhs-accelerates-artificial-intelligence-rollout-to-cut-waiting-times-and-improve-care-for-millions/

Open resource
Continuum
Also relevant to

B4. Workforce training and supervision

Can AI build the capacity of clinicians, peer workers and lay helpers, including through simulated clients and supervision support?

READ
Opinion and commentaryPreprint, not yet reviewed

Generative AI to train the mental-health workforce

An arXiv paper arguing the highest-leverage use of generative AI in mental health is not the therapist chatbot but scaling training of human providers.

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Enhancing mental health research capacity: emerging voices from the National Institute of Mental Health (NIMH) global hubs

Emerging researchers from the NIMH global hubs on building mental health research capacity in low- and middle-income settings.

da Silva, A. T. C., Hanlon, C., Susser, E., Rojas, G., Claro, H. G., Quayle, J., ... Seedat, S. (2019). Enhancing mental health research capacity: emerging voices from the National Institute of Mental Health (NIMH) global hubs. International Journal of Mental Health Systems, 13(1), 21. https://doi.org/10.1186/s13033-019-0276-9

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

The mental health workforce gap in low- and middle-income countries: a needs-based approach

Needs-based estimate of the mental health workforce gap in low- and middle-income countries.

Bruckner, T. A., Scheffler, R. M., Shen, G., Yoon, J., Chisholm, D., Morris, J., ... Saxena, S. (2011). The mental health workforce gap in low- and middle-income countries: a needs-based approach. Bulletin of the World Health Organization, 89(3), 184-194. https://doi.org/10.2471/blt.10.082784

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

ASPPH: AI Framework Report for public health education, research and practice

Framework from the Association of Schools and Programs of Public Health's Task Force for the Responsible and Ethical Use of AI. Treats AI as both an educational disruptor and a determinant of health, insists it should augment rather than replace human judgment and community relationships, and organises recommendations around workforce and curriculum, teaching and learning, research and practice, and governance.

Association of Schools and Programs of Public Health. (n.d.). AI framework report. https://aspph.org/initiatives/ai-for-public-health/ai-framework-report/

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Effectiveness of support from community health workers on the sustained use of a wearable monitoring device among community-dwelling older adults: A randomized trial protocol

Randomised trial protocol on community health worker support for sustained use of wearable monitoring among community-dwelling older adults.

Wong, A. K. C., Tso, W. C., Su, J. J., Hui, V. C. C., Chow, K. K. S., Wong, S. M., ... Wong, F. K. Y. (2023). Effectiveness of support from community health workers on the sustained use of a wearable monitoring device among community-dwelling older adults: A randomized trial protocol. PLOS ONE, 18(12), e0294517. https://doi.org/10.1371/journal.pone.0294517

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Voices of innovation: Reflective report on integrating artificial intelligence–simulated mental health patient scenarios into undergraduate nursing education in the United Arab Emirates

A nursing school used ChatGPT Voice simulated patients for mental state examinations, risk assessment and care planning because clinical placements were scarce. Shows AI role play filling a real training gap in a non-Western setting. A reflective report, not an evaluation.

Ahmad, A., Kavumpurath, J., Kader, R., Altamimi, M., El Hajj, M., Ahmed, F. R. A., Subu, M. A., Yafei, T., Ali, H. R., Saleh, I., & Al-Yateem, N. (2026). Voices of innovation: Reflective report on integrating artificial intelligence–simulated mental health patient scenarios into undergraduate nursing education in the United Arab Emirates. JMIR Medical Education, 12, Article e78161. https://doi.org/10.2196/78161

Open resource
Continuum
Also relevant to
READ
Guidance and standardsPeer reviewed

Teaching motivational interviewing with an AI-powered virtual patient simulation: A teaching note

Describes 'Sam', a GPT-4o virtual patient who is ambivalent about quitting smoking, used by 50 behavioural health doctoral students to practise motivational interviewing. Early feedback reported more confidence and realism. A practical teaching model rather than an outcome study.

Cordes, C. C., Mitchell, T., Macchi, C. R., Manson, L., Sutton-Ryan, A., & Kaplan, J. (2026). Teaching motivational interviewing with an AI-powered virtual patient simulation: A teaching note. Studies in Clinical Social Work: Transforming Practice, Education and Research. Advance online publication. https://doi.org/10.1080/28376811.2026.2711212

Open resource
Continuum
READ
Evidence reviewPeer reviewed

Use of artificial intelligence–enhanced virtual patients in educational approaches to medical interview training: A systematic review

Across 11 studies, including four RCTs and psychiatric interview training, AI virtual patients were linked to better interviewing and communication skills, empathy and confidence. Whether these gains carry over into real clinical work is still uncertain.

Gilbert, V., Philip, P., Llorca, P.-M., & Samalin, L. (2026). Use of artificial intelligence–enhanced virtual patients in educational approaches to medical interview training: A systematic review. BMC Medical Education, 26, Article 588. https://doi.org/10.1186/s12909-026-08804-9

Open resource
Continuum
READ
Opinion and commentaryPeer reviewed

Catch my drift: Building scalable and sustainable models for promoting clinician adherence to evidence-based treatments for eating disorders

Argues for virtual training plus AI-supported fidelity monitoring to keep clinicians on track with family-based treatment for eating disorders, extending access to under-resourced families.

Goldschmidt, A. B., & Graham, A. K. (2026). Catch my drift: Building scalable and sustainable models for promoting clinician adherence to evidence-based treatments for eating disorders. Journal of Medical Internet Research, 28, e107039. https://doi.org/10.2196/107039

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Validating GenAI feedback in suicide prevention training: A mixed-methods study of QPR skill assessment

AI scoring of simulated gatekeeper conversations agreed well with trained human raters (ICC 0.83) and gave tailored feedback, though it scored about 1.2 points more generously. Supports scalable, individual feedback in suicide-prevention training.

Haber, Y., Levi-Belz, Y., Segal Elbak, Y., Elyoseph, Z., & Levkovich, I. (2026). Validating GenAI feedback in suicide prevention training: A mixed-methods study of QPR skill assessment. Frontiers in Medicine, 12, Article 1709743. https://doi.org/10.3389/fmed.2025.1709743

Open resource
Continuum
READ
ResearchPeer reviewed

LLMs as standardised patients for motivational interviewing: How faithful are they?

Compared 140 simulated and 140 real patient conversations. Simulated patients agreed to change about twice as often as real ones (47.9% vs 24.0%), so they underplay the ambivalence trainees most need to practise with.

Hoang, V., Rogers, E., & Ross, R. (2026). LLMs as standardised patients for motivational interviewing: How faithful are they? In Proceedings of the Workshop on Computational Linguistics and Clinical Psychology (CLPsych 2026) (pp. 258–270). Association for Computational Linguistics. https://aclanthology.org/2026.clpsych-1.21.pdf

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

AI-powered simulation for nursing education: Mixed methods systematic review

Across 19 studies and 1,253 learners, AI simulation showed promise for knowledge, clinical reasoning, communication confidence and safe repeatable practice. Only three were RCTs, none followed retention, and one trial found AI simulation inferior to human standardised patients.

Jiang, H., Wang, Z., Shen, W., Meng, M., Yang, D., Li, X., & Hao, Y. (2026). AI-powered simulation for nursing education: Mixed methods systematic review. Journal of Medical Internet Research, 28, Article e95167. https://doi.org/10.2196/95167

Open resource
Continuum
READ
ResearchPeer reviewed

Multiagent large language model framework for psychotherapy fidelity assessment in motivational interviewing and cognitive behavioral therapy training: Cross-sectional, simulation-based evaluation study

A multi-agent AI system rated therapist fidelity in motivational interviewing and CBT with very high agreement with human consensus (ICC 0.93 to 0.96). Early-stage and synthetic only, but it points to affordable fidelity feedback, which is one of the scarcest inputs in training.

Kamaleddin, M. A., Mirjalili, M., Barzegar, R., Le, N. T., Cote, Z., Winkler, O., Burback, L., Zhang, Y., Zaiane, O. R., Monson, C., Sharma, D., Krishnan, S., Greenshaw, A. J., Zeifman, R., Selby, P., & Bhat, V. (2026). Multiagent large language model framework for psychotherapy fidelity assessment in motivational interviewing and cognitive behavioral therapy training: Cross-sectional, simulation-based evaluation study. JMIR Medical Education, 12, e92964. https://doi.org/10.2196/92964

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

A preliminary study of motivational interviewing training using a large language model (ChatGPT): Effects on empathy and motivational interviewing confidence among social work students

Fifteen social work students practised motivational interviewing with ChatGPT playing a client and giving fidelity-based feedback. Empathy and confidence rose and held at four weeks. A one-group pilot, so no comparison condition.

Kang, H. Y., Woo, S., & Kim, H. J. (2026). A preliminary study of motivational interviewing training using a large language model (ChatGPT): Effects on empathy and motivational interviewing confidence among social work students. Korean Journal of Health Promotion, 26(3), 83–93. https://doi.org/10.15384/kjhp.2026.00304

Open resource
Continuum
READ
ResearchPeer reviewed

A feasibility study of an AI-based virtual human training design in suicide gatekeeper education

AI virtual-human scenarios were feasible in gatekeeper training for 53 Taiwanese undergraduates, who reported better communication skills and confidence. Confidence to act as a gatekeeper remained the lowest-rated item.

Lee, P.-I., & Wu, C.-Y. (2026). A feasibility study of an AI-based virtual human training design in suicide gatekeeper education. Medical Teacher, 48(3), 467–475. https://doi.org/10.1080/0142159X.2025.2561204

Open resource
Continuum
READ
ResearchPeer reviewed

Development and preliminary evaluation of a virtual standardized patient system for psychiatric interview training

A small randomised trial (20 participants) of an AI virtual patient for psychiatric interview practice. The intervention group scored higher on clinical interviewing than controls after training. Promising but very small.

Li, Y., Wang, G., Huang, Z., Ke, X., Qin, J., Yu, J., Liu, S., Li, J., & Xu, Y. (2026). Development and preliminary evaluation of a virtual standardized patient system for psychiatric interview training. BMC Psychiatry, 26, Article 264. https://doi.org/10.1186/s12888-026-07896-3

Open resource
Continuum
READ
ResearchPeer reviewed

Can LLM-simulated practice and feedback upskill human counselors? A randomized study with 90+ novice counselors

A randomised study of 94 novice counsellors. Practising with simulated clients plus AI feedback improved reflections and questions. Practice alone did not, and empathy fell in the practice-only group. The key message: simulation helps only when paired with structured feedback.

Louie, R., Shah, R. S., Orney, I. H., Pacheco, J. P., Brunskill, E., & Yang, D. (2026). Can LLM-simulated practice and feedback upskill human counselors? A randomized study with 90+ novice counselors. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (pp. 1–31). Association for Computing Machinery. https://doi.org/10.1145/3772318.3791821

Open resource
Continuum
READ
Evidence reviewPeer reviewed

Use of digital health in task-sharing for prevention and management of non-communicable diseases in Africa: A scoping review

Across 15 African studies, digital tools supported community and non-physician health workers with screening, treatment and supervision. Two mental health studies found lower symptoms. Digital literacy, connectivity and electricity were the main constraints.

Mhlanga, N. L., Mabunda, S. A., Joshi, R., & Tesema, A. G. (2026). Use of digital health in task-sharing for prevention and management of non-communicable diseases in Africa: A scoping review. PLOS Global Public Health, 6(7), e0006746. https://doi.org/10.1371/journal.pgph.0006746

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Mostly grounded, occasionally risky: Expert evaluation of LLM-generated supervisory feedback in a psychotherapy training simulator

Experts judged 89% of AI supervisory feedback points justified, but more than half of feedback sessions contained at least one unjustified point. A balanced reminder that AI supervision needs human oversight.

Montesano, A., Bloomberg, J. K., & Pérez-Buriel, M. (2026). Mostly grounded, occasionally risky: Expert evaluation of LLM-generated supervisory feedback in a psychotherapy training simulator. In Proceedings of the CLPsych 2026 Workshop (pp. 298–305). Association for Computational Linguistics. https://aclanthology.org/2026.clpsych-1.24.pdf

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

How an AI avatar will help social work students practice their counseling skills

Describes 'Emma', a virtual client with depression symptoms after a breakup, with AI feedback on attunement, technique and ethics. Some early users reported an 'uncanny valley' effect in virtual reality.

Why it matters now: The avatar offers practice feedback, while some early users reported an uncanny-valley effect.

NYU Steinhardt. (2026, April 17). How an AI avatar will help social work students practice their counseling skills. https://steinhardt.nyu.edu/news/how-ai-avatar-will-help-social-work-students-practice-their-counseling-skills

Open resource
Continuum
READ
Practice signalNot reviewed

Bridging the treatment gap with AI-enhanced supervision (BRIDGE-AI)

Funded project to build AI supervision feedback for lay counsellors in India, using more than 4,000 Hindi behavioural activation sessions, then pilot it. Addresses the supervision bottleneck in task-shared care. No results yet.

Why it matters now: The funded project will develop and pilot AI supervision feedback for lay counsellors using more than 4,000 Hindi sessions.

Patel, V., Pozuelo, J. R., Khanduri, M., Yang, D., Singla, D., Bhan, A., Brunskill, E., & Lahiri, A. (2026, March 18). Bridging the treatment gap with AI-enhanced supervision (BRIDGE-AI) [Grant]. Wellcome. https://wellcome.org/research-funding/funding-portfolio/funded-grants/bridging-treatment-gap-ai-enhanced-supervision

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

PsyCLIENT: Client simulation via conversational trajectory modeling for trainee practice and model evaluation in mental health counseling

Models how a client's conversation changes over a session across 60 counselling topics. Licensed counsellors rated it above other simulators and often could not tell it from a real client.

Why it matters now: Licensed counsellors rated this client simulator above other simulators and often could not distinguish it from a real client.

Qiu, H., Chen, Z., Chen, Y., Xie, Y., Lu, Y., & Lan, Z. (2026). PsyCLIENT: Client simulation via conversational trajectory modeling for trainee practice and model evaluation in mental health counseling [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2601.07312

Open resource
Continuum
READ
ResearchPreprint, not yet reviewed

Toward accessible psychotherapy training using AI-driven interactive patient avatars

An avatar-based virtual patient with turn-by-turn feedback on Acceptance and Commitment Therapy skills. Psychologists found the patient realistic and the feedback useful for trying new responses. Presented at IEEE ICHI 2026.

Why it matters now: Psychologists found the AI patient realistic and its turn-by-turn feedback useful for practising Acceptance and Commitment Therapy.

Riachi, P., Kamber, S., Brogna, S., Gloster, A., & Wampfler, R. (2026). Toward accessible psychotherapy training using AI-driven interactive patient avatars [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2606.17786

Open resource
Continuum
READ
ResearchPreprint, not yet reviewed

MyMentorLLM: A psychotherapy GenAI environment with multimodal voice/text patients, trainees and experts for deliberate practice

Generated 2,100 CBT practice sessions with simulated patients, trainees and supervisors. Supervisor feedback improved diagnostic accuracy in most models, but most models overestimated trainee competence. A caution against trusting AI assessment of trainees.

Why it matters now: The simulation study warns that most models overestimated trainee competence, despite supervisor feedback improving diagnostic accuracy in most models.

Rizzi, R., Grecucci, A., & Stella, M. (2026). MyMentorLLM: A psychotherapy GenAI environment with multimodal voice/text patients, trainees and experts for deliberate practice [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2607.25667

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

Evaluating role-consistency in LLMs for counselor training

Tests whether open-source models stay in character as a virtual client when trainees push them off script. A useful quality check, since a simulated client that breaks role teaches the wrong thing.

Why it matters now: It tests whether simulated clients stay in role, a quality check for AI-based counsellor training.

Rudolph, E., Engert, N., & Albrecht, J. (2026). Evaluating role-consistency in LLMs for counselor training [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2601.08892

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Generative AI for simulated conversations: Preparing frontline workers and volunteers to support youth with anxiety and depression

A 24-month project to co-develop and test a multi-agent conversation simulator and assessor for training frontline workers and volunteers in single-session and crisis conversations with young people. Extends simulation beyond clinicians. No results yet.

Why it matters now: A new simulator will extend conversation training to frontline workers and volunteers supporting young people.

Sequeira, L., Raman, P., Shearkhani, S., Sodha, S., Khan, N., Ensan, F., Milligan, K., & Deschamps, A. (2026, March 18). Generative AI for simulated conversations: Preparing frontline workers and volunteers to support youth with anxiety and depression [Grant]. Wellcome. https://wellcome.org/research-funding/funding-portfolio/funded-grants/generative-ai-simulated-conversations-preparing

Open resource
Continuum
Also relevant to
READ
Guidance and standardsPeer reviewed

From innovation to implementation: Artificial intelligence in cognitive behaviour therapy training and supervision

An international expert group sets out how AI could make psychotherapy training and supervision more efficient, consistent and accessible through simulated patients, real-time fidelity feedback and support for cultural sensitivity. Notes the need for good datasets, ethics, low-resource implementation and more rigorous research.

Shafran, R., Bond, L., Carlbring, P., Cohen, Z. D., Creed, T. A., Davey, E., Egan, S. J., Freeman, D., Hollon, S. D., Jacobson, N. C., Johnson, C., Kaysen, D., McGuinness, D. L., Patel, V., Pozuelo, J. R., Santos, H., Singla, D. R., Wiltsey Stirman, S., Taylor, D. J., & Wade, T. D. (2026). From innovation to implementation: Artificial intelligence in cognitive behaviour therapy training and supervision. Behaviour Research and Therapy, 197, Article 104945. https://doi.org/10.1016/j.brat.2025.104945

Open resource
Continuum
READ
ResearchPreprint, not yet reviewed

GenAI for social work field education: Client simulation with real-time feedback

SWITCH pairs simulated clients with real-time classification of counselling skills for social work students. Also presented at IEEE Big Data 2025. The abstract reports gains over baseline but gives no numbers.

Why it matters now: The preprint describes simulated-client practice with real-time counselling-skills feedback, but reports gains without numbers.

Sungarda, J., Liu, H., Zhou, Z., Wu, T.-H., Cheung, J. C.-S., & Kao, B. (2026). GenAI for social work field education: Client simulation with real-time feedback [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2601.18517

Open resource
Continuum
READ
Practice signalNot reviewed

AI-powered virtual clients give USF social work students a new way to prepare for practice

Announces SCOPE-AI, a two-year project building virtual clients for interviewing, assessment and intervention practice in social work education, with planned evaluation. A practice signal, not evidence.

Why it matters now: A two-year virtual-client project is being built for social work practice, with evaluation planned.

University of South Florida College of Behavioral and Community Sciences. (2026, August 18). AI-powered virtual clients give USF social work students a new way to prepare for practice. https://www.usf.edu/cbcs/news/2026/ai-powered-virtual-clients-give-usf-social-work-students-a-new-way-to-prepare-for-practice.aspx

Open resource
Continuum
READ
Practice signalNot reviewed

Culturally aligned generative AI to support para-counselors in low-resource community mental health settings

Funded project combining a causal reasoning framework with language models to help para-counsellors understand clients through culturally and contextually appropriate explanatory models. One of the few projects building cultural formulation into AI support. No results yet.

Why it matters now: The funded project brings cultural formulation into AI support for para-counsellors in low-resource settings.

Wellcome. (2026, March 18). Culturally aligned generative AI to support para-counselors in low-resource community mental health settings [Grant]. https://wellcome.org/research-funding/funding-portfolio/funded-grants/culturally-aligned-generative-ai-support-para

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Generative AI in psychiatric education: Balancing risks and benefits for responsible integration

Weighs simulated patients for history taking, suicide risk assessment and breaking bad news against the risks. Argues AI practice supplements but cannot replace real encounters or teach the therapeutic alliance.

Zhou, C., Zhang, A., & Li, S. (2026). Generative AI in psychiatric education: Balancing risks and benefits for responsible integration. Frontiers in Psychiatry, 17, Article 1894299. https://doi.org/10.3389/fpsyt.2026.1894299

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Leveraging large language models for simulated psychotherapy client interactions: Development and usability study of Client101

An Australian simulated client for anxiety and depression. Sixteen mental health professionals generated 29 practice sessions and 14 of 15 found the anxiety client's story coherent and convincing. Language analysis still showed differences from real therapy transcripts.

Cabrera Lozoya, D., Conway, M., Sebastiano De Duro, E., & D'Alfonso, S. (2025). Leveraging large language models for simulated psychotherapy client interactions: Development and usability study of Client101. JMIR Medical Education, 11, Article e68056. https://doi.org/10.2196/68056

Open resource
Continuum
READ
Guidance and standardsNot reviewed

New field guide on mental health & AI: A practical resource for innovators

A practical field guide with case examples and tools for organisations considering AI in community mental health, including task-sharing, training, supervision and access in low-resource settings.

Grand Challenges Canada. (2025, July 7). New field guide on mental health & AI: A practical resource for innovators. https://www.grandchallenges.ca/2025/new-field-guide-on-mental-health/

Open resource
Continuum
Also relevant to
READ
ResearchNot reviewed

How are GPs using AI? Insights from the front line

A survey of 2,108 UK GPs: 28% used AI in clinical practice, and of those reporting tasks, 45% used it for professional development. Shows self-directed AI learning already happening outside formal training.

Kumpunen, S., Lobont, C., Garrard, L., Fisher, D., Lau, R., Fallica, G., & Fisher, R. (2025, December 3). How are GPs using AI? Insights from the front line. Nuffield Trust. https://www.nuffieldtrust.org.uk/research/how-are-gps-using-ai-insights-from-the-front-line

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

A step toward the future? Evaluating GenAI QPR simulation training for mental health gatekeepers

Practising suicide-prevention gatekeeper skills (Question, Persuade, Refer) with a generative AI simulator raised self-efficacy substantially in 89 paraprofessionals (d = 1.67). No control group and self-report only, but it shows how crisis-response training could reach many more lay helpers.

Levkovich, I., Haber, Y., Levi-Belz, Y., & Elyoseph, Z. (2025). A step toward the future? Evaluating GenAI QPR simulation training for mental health gatekeepers. Frontiers in Medicine, 12, Article 1599900. https://doi.org/10.3389/fmed.2025.1599900

Open resource
Continuum
READ
ResearchNot reviewed

Digital and AI skills in health occupations: What do we know about new demand?

Maps the digital and AI skills health workers now need, including interpreting wearable data on heart rate, activity and sleep for care plans. Positions AI training as a workforce planning issue, not just a technology one. A working paper.

Manca, F., & Eslava, D. (2025). Digital and AI skills in health occupations: What do we know about new demand? (OECD Artificial Intelligence Papers, No. 36). OECD Publishing. https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/05/digital-and-ai-skills-in-health-occupations_f428e5a9/5fbd42ab-en.pdf

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Scaffolding empathy: Training counselors with simulated patients and utterance-level performance visualization

Counsellor trainees practise with an AI simulated patient and receive turn-by-turn feedback on their responses. Part of a growing set of tools that let trainees rehearse difficult conversations safely before real clients.

Steenstra, I., Nouraei, F., & Bickmore, T. (2025). Scaffolding empathy: Training counselors with simulated patients and utterance-level performance visualization. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (pp. 1–22). Association for Computing Machinery. https://doi.org/10.1145/3706598.3714014

Open resource
Continuum
READ
ResearchPreprint, not yet reviewedLived experience

Roleplay-doh: Enabling domain-experts to create LLM-simulated patients via eliciting and adhering to principles

Lets experienced peer supporters build their own simulated clients by writing principles the AI must follow. Twenty-five counselling experts judged the results more like real clients. Puts practitioners, not engineers, in charge of what realistic means.

Louie, R., Nandi, A., Fang, W., Chang, C., Brunskill, E., & Yang, D. (2024). Roleplay-doh: Enabling domain-experts to create LLM-simulated patients via eliciting and adhering to principles [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2407.00870

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

PATIENT-Ψ: Using large language models to simulate patients for training mental health professionals

Simulated CBT clients built from expert-written case formulations. In testing with 13 trainees and 20 experts, trainees reported more skill and confidence than with textbooks, videos or peer role play, and experts judged the simulated clients closer to real patients than plain GPT-4.

Wang, R., Milani, S., Chiu, J. C., Zhi, J., Eack, S. M., Labrum, T., Murphy, S. M., Jones, N., Hardy, K. V., Shen, H., Fang, F., & Chen, Z. (2024). PATIENT-Ψ: Using large language models to simulate patients for training mental health professionals. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (pp. 12772–12797). Association for Computational Linguistics. https://doi.org/10.18653/v1/2024.emnlp-main.711

Open resource
Continuum
READ
Opinion and commentaryPeer reviewed

An AI Perspective on Counseling Supervision

The article discusses potential uses of AI in counseling supervision, including transcription, feedback, skills development and administrative support, and argues that AI should complement rather than replace human supervision. It describes the evidence as limited and mixed, notes risks such as privacy breaches and bias, and cites findings that 18% of supervised trainees were skeptical AI could replace a human therapist, 41% said a chatbot responded poorly to their emotions, and 36% reported difficulty with its understanding of user inputs.

Brinck, E. A., Soldner, J. L., Kuo, H. J. et al. (2026). An AI Perspective on Counseling Supervision. Behavioral Sciences. https://doi.org/10.3390/bs16061038

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Artificial intelligence in public health education: current evidence, competencies, and a framework for responsible integration

This narrative review synthesises 21 publications from January 2020 to June 2026 and finds that formal AI curricula and public-health-specific evidence remain limited. It proposes a nine-domain framework for responsible AI integration in public health education, which the authors say needs prospective validation.

Arab, H.; Mirdad, G.A.; Abd-Allah, M.A.M. (2026). Artificial intelligence in public health education: current evidence, competencies, and a framework for responsible integration. Frontiers in Public Health. https://doi.org/10.3389/fpubh.2026.1951054

Open resource
Continuum
Also relevant to

B5. Developing resource materials

Can AI help produce culturally adapted, plain-language resources at low cost?

READ
ResearchPeer reviewed

Generative AI with youth codesign for vaping awareness advertisements

Study of 614 people aged 16-25 comparing AI-generated, youth-codesigned advertisements with agency advertisements. Supports AI-assisted creative production with community involvement.

Leung, J., et al. (2025). Generative Artificial Intelligence With Youth Codesign to Create Vaping Awareness Advertisements. JAMA Network Open, 8(7), e2514040. https://pubmed.ncbi.nlm.nih.gov/40742592/

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewedLived experience

Evaluating ChatGPT-generated psychoeducation for mood disorders: Comparative insights from patients and mental health professionals

Inpatients with mood disorders rated AI-generated psychoeducation more favourably than psychiatrists and rehabilitation staff did, especially for empathy and acceptance. A reminder that professional and lived-experience judgements of a resource can differ.

Attanasio, F., Fazio, V., Antonini, C., Lanzano, N., Obumselu, G., Prato, M., Flutti, E., Pacchioni, F., Fregna, L., Franchini, L. A. M., & Colombo, C. (2026). Evaluating ChatGPT-generated psychoeducation for mood disorders: Comparative insights from patients and mental health professionals. Journal of Psychiatric Research, 193, 354–360. https://doi.org/10.1016/j.jpsychires.2025.11.039

Open resource
Continuum
READ
ResearchPeer reviewed

An Ubuntu-guided large language model framework for cognitive behavioral mental health dialogue

Adapts CBT-style dialogue using Ubuntu values, proverbs and faith, with expert ratings of cultural fit. Proof of concept only, but a concrete example of culturally grounded content development rather than translation of Western material.

Forane, S. G., Ezugwu, A. E., Igwe, K., & van den Berg, K. (2026). An Ubuntu-guided large language model framework for cognitive behavioral mental health dialogue. Frontiers in Digital Health, 8, Article 1903421. https://doi.org/10.3389/fdgth.2026.1903421

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Quality and readability of AI-generated information on bipolar disorder: A cross-sectional content analysis

AI answers to common bipolar disorder questions were of moderate quality, varying by model. Suitable as supplementary psychoeducation; accuracy and patient benefit were not tested.

Karakaya, I. (2026). Quality and readability of AI-generated information on bipolar disorder: A cross-sectional content analysis. BMC Psychiatry, 26, Article 630. https://doi.org/10.1186/s12888-026-08262-z

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Exploring the use of translation technologies to overcome language barriers in mental healthcare: A qualitative cross-country study

Interviews with 175 people in five countries found translation technology can bridge language gaps in urgent or under-resourced settings. It falls short in emotionally complex conversations, so participants preferred human-supported use over replacing interpreters.

Kreienbrinck, A., Hanft-Robert, S., Topçu, M.-T., Peters, L., Oltean, O., Nozewu, A., Al-Kalaf, H., Liem, A., Forray, A. I., Madzamba, R., Swartz, L., Anthonissen, C., Sanders, T., Hall, B., Schouten, B., Jenks, C., Cherecheș, R., & Mösko, M. (2026). Exploring the use of translation technologies to overcome language barriers in mental healthcare: A qualitative cross-country study. International Journal for Equity in Health, 25, Article 88. https://doi.org/10.1186/s12939-026-02821-2

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Information quality, readability, and empathy of AI-generated public mental health information: A comparative evaluation of eight large language models

Eight major models answered 48 public mental health questions with reasonable quality and often high rated empathy. None met the sixth-grade readability standard and few cited sources, which shows where human editing still adds value in producing resources.

Wang, Q., Ma, S., Sun, J., Qi, X., Zhao, J., Ma, D., Liu, X., An, Q., Zhao, D., & He, S. (2026). Information quality, readability, and empathy of AI-generated public mental health information: A comparative evaluation of eight large language models. Frontiers in Public Health, 14, Article 1859078. https://doi.org/10.3389/fpubh.2026.1859078

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Large language models for individualized psychoeducational tools for psychosis: A cross-sectional study

Psychosis experts rated GPT-4 psychoeducation for first-episode programmes highly for accuracy, clarity and clinical utility. Inclusivity and readability were weaker. A benefit example at the severe end of the continuum, under clinician guidance.

Yilanli, M., McKay, I., Jackson, D. I., & Sezgin, E. (2026). Large language models for individualized psychoeducational tools for psychosis: A cross-sectional study. Early Intervention in Psychiatry, 20(8), e70230. https://doi.org/10.1111/eip.70230

Open resource
Continuum
READ
ResearchPeer reviewedLived experience

Using generative AI to co-design digital mental health interventions with adolescents in rural South Africa: Qualitative thematic analysis of participatory workshops

Rural South African adolescents used generative AI to create images and music for a depression intervention and mostly preferred the AI content to the original illustrations. Most needed help with prompts and many noticed cultural bias, which supports co-design over automated production.

Dallison, S., Moffett, B., Makhubela, P., Nkuna, T., Pozuelo, J. R., van Heerden, A., & O'Mahen, H. (2025). Using generative AI to co-design digital mental health interventions with adolescents in rural South Africa: Qualitative thematic analysis of participatory workshops. Journal of Medical Internet Research, 27, e73535. https://doi.org/10.2196/73535

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Prompt engineering a large language model with evidence-based persuasive features to improve confidence in mental health professionals: a pilot randomized experiment

In Phase 1, 168 adults rated depression education texts, with source credibility, argument quality and processing fluency improving perceived credibility and persuasiveness. In Phase 2, confidence in mental health professionals improved more among the 40 participants who received feature-engineered texts than among those who received baseline texts (t(38) = 3.37, p = 0.002, d = 1.07), but the authors note that the result is preliminary and based on an immediate single-item self-report.

Li, A., Yao, S.-T., Chen, B.-X. et al. (2026). Prompt engineering a large language model with evidence-based persuasive features to improve confidence in mental health professionals: a pilot randomized experiment. Frontiers in Public Health. https://doi.org/10.3389/fpubh.2026.1909370

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Errors in AI-Transformed Patient-Centered Mental Health Documentation Written by Psychiatrists: Qualitative Pre-Post Study

The study compared 63 psychiatrist-authored mental health notes with versions transformed by GPT-3.5 Turbo and identified five categories of clinically relevant errors. In a quantitative analysis of 20 notes, researchers found 134 errors, averaging 6.7 per note; speculative content distortion was the most frequent subtype, appearing 55 times and in 19 of the 20 notes.

Ozkara Menekseoglu, P., Weibezahl, M., Ellingsen, M. et al. (2026). Errors in AI-Transformed Patient-Centered Mental Health Documentation Written by Psychiatrists: Qualitative Pre-Post Study. JMIR Mental Health. https://doi.org/10.2196/78351

Open resource
Continuum
Also relevant to

B6. Support at scale where there is no other option

Where there is no service, is governed AI support better than nothing, and for whom?

READ
ResearchPeer reviewed

AI-enabled therapy support tool: structured, adjunctive use between sessions

A JMIR real-world observational study (Habicht et al., 2025) of 244 NHS Talking Therapies patients in group CBT found that when an AI tool was deliberately integrated between sessions for homework support, prompts, and monitoring, patients attended more sessions and dropped out less rather than disengaging. A concrete empirical example of AI shaping between-session engagement in a structured, adjunctive way.

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

Therabot: RCT of a generative AI therapy chatbot

Dartmouth's randomised controlled trial of Therabot showing symptom reductions comparable to gold-standard therapy for depression and anxiety — alongside the important caveats.

Open resource
Continuum
Also relevant to
Related
READ Trial coverage: comparable to cognitive therapy
READ
ResearchPeer reviewed

The Effectiveness of a Chatbot Single-Session Intervention for People on Waitlists for Eating Disorder Treatment: Randomized Controlled Trial

Randomised controlled trial of ED ESSI, a co-designed rule-based chatbot delivering a single-session intervention to people on waitlists for eating disorder treatment. A rare example of a trial aimed at the waitlist gap rather than replacing treatment.

Sharp, G., Dwyer, B., Randhawa, A., McGrath, I., & Hu, H. (2025). The Effectiveness of a Chatbot Single-Session Intervention for People on Waitlists for Eating Disorder Treatment: Randomized Controlled Trial. Journal of Medical Internet Research, 27, e70874. https://doi.org/10.2196/70874

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

2025 Watch List: Artificial Intelligence in Health Care (Canada's Drug Agency)

Horizon scan from Canada's Drug Agency (formerly CADTH) identifying AI technologies likely to affect Canadian health care, including 24/7 app-based mental health chatbots such as Wysa used alongside a treatment plan. Useful as a health technology assessment view of where AI is heading.

Canada's Drug Agency. (2025, March). 2025 watch list: Artificial intelligence in health care (Horizon Scan, Report No. ER0015). CDA-AMC. https://www.ncbi.nlm.nih.gov/books/NBK613808/

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewedLived experience

Experiences of generative AI chatbots for mental health

An NPJ Mental Health Research qualitative study of what people actually report from using generative AI chatbots for mental-health support.

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Artificial Intelligence-Based Mobile Phone Apps for Child Mental Health: Comprehensive Review and Content Analysis

Review and content analysis of AI-based mobile apps marketed for child and teenage mental health, including what they claim, what evidence they carry and where safety features fall short.

Yang, F., Wei, J., Zhao, X., & An, R. (2025). Artificial Intelligence-Based Mobile Phone Apps for Child Mental Health: Comprehensive Review and Content Analysis. JMIR mHealth and uHealth, 13, e58597. https://doi.org/10.2196/58597

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Conversational Agent Interventions for Mental Health Problems: Systematic Review and Meta-analysis of Randomized Controlled Trials

Systematic review and meta-analysis of randomised controlled trials of conversational agents for mental health problems.

He, Y., Yang, L., Qian, C., Li, T., Su, Z., Zhang, Q., & Hou, X. (2023). Conversational Agent Interventions for Mental Health Problems: Systematic Review and Meta-analysis of Randomized Controlled Trials. Journal of Medical Internet Research, 25, e43862. https://doi.org/10.2196/43862

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Systematic review and meta-analysis of AI-based conversational agents for promoting mental health and well-being

Systematic review and meta-analysis of AI-based conversational agents for promoting mental health and wellbeing, with subgroup analysis by agent type and population.

Li, H., Zhang, R., Lee, Y. C., Kraut, R. E., & Mohr, D. C. (2023). Systematic review and meta-analysis of AI-based conversational agents for promoting mental health and well-being. npj Digital Medicine, 6(1), 236. https://doi.org/10.1038/s41746-023-00979-5

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Examining the role of AI technology in online mental healthcare: opportunities, challenges, and implications, a mixed-methods review

Mixed-methods review of AI in online mental healthcare: opportunities, challenges and implications for practice.

Gutierrez, G., Stephenson, C., Eadie, J., Asadpour, K., & Alavi, N. (2024). Examining the role of AI technology in online mental healthcare: opportunities, challenges, and implications, a mixed-methods review. Frontiers in Psychiatry, 15, 1356773. https://doi.org/10.3389/fpsyt.2024.1356773

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Systematic Review of Artificial Intelligence in Positive and Existential Psychiatry: Advancing Mental and Emotional Health Through Metacompetency Development

Systematic review of AI in positive and existential psychiatry, including generative chatbots delivering Socratic questioning and reflective listening for meaning and metacognitive skills.

Mitsea, E., Drigas, A., & Skianis, C. (2026). Systematic Review of Artificial Intelligence in Positive and Existential Psychiatry: Advancing Mental and Emotional Health Through Metacompetency Development. Healthcare, 14(6), 783. https://doi.org/10.3390/healthcare14060783

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewedIndustry funded

AI in mental health market growth analysis: Size and forecast 2025-2029

Commercial market forecast for AI in mental health. Marketing-grade estimate; use for orientation only.

Technavio. (2025). AI in mental health market growth analysis: Size and forecast 2025-2029. Technavio. https://www.technavio.com/report/ai-in-mental-health-market-industry-analysis

Open resource
READ
ResearchPeer reviewed

Use of AI in Mental Health Care: Community and Mental Health Professionals Survey

Australian survey led from Orygen of 107 community members and 86 mental health professionals. AI was used by 28 percent of community members, mostly for quick emotional support or as a personal therapist or coach, and by 43 percent of professionals, mostly for research and report writing. Both groups saw benefits but wanted safeguards. The first Australian data of its kind.

Cross, S., Bell, I., Nicholas, J., Valentine, L., Mangelsdorf, S., Baker, S., Titov, N., & Alvarez-Jimenez, M. (2024). Use of AI in Mental Health Care: Community and Mental Health Professionals Survey. JMIR Mental Health, 11, e60589. https://doi.org/10.2196/60589

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Arabic Mental Health Apps: A Narrative Review of Therapeutic and Engagement Features

Narrative review of Arabic mental-health apps, examining therapeutic features, engagement, language and cultural adaptation. Identifies opportunities for more relevant digital support while distinguishing existing app evidence from proposed AI capabilities.

Al Harrasi, A., et al. (2025). Arabic Mental Health Apps: A Narrative Review of Therapeutic and Engagement Features. Saudi Journal of Medicine & Medical Sciences, 13(4), 253-259. https://doi.org/10.4103/sjmms.sjmms_62_25

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

How AI could help treat mental health

Wellcome's overview of AI benefits for access, triage, therapist training and routine tasks, with examples from Uganda, India, Kenya and the UK. It reports that about one in three UK adults use general chatbots for mental health support, mostly finding them helpful, though one in ten felt worse.

Ahmed, M. (2026, September 3). How AI could help treat mental health. Wellcome. https://wellcome.org/insights/articles/how-ai-could-help-treat-mental-health

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewedLived experience

A guided chatbot-based psychological intervention for psychologically distressed older adolescents and young adults: A randomised clinical trial in Jordan

WHO's STARS chatbot, with brief helper calls, reduced anxiety and depression more than enhanced usual care in 344 distressed young people in Jordan, including Syrian refugees (NNT about 3). STARS is a scripted, not generative, chatbot, so this is benefit evidence for guided digital delivery where services are scarce.

Bryant, R. A., de Graaff, A. M., Habashneh, R., Fanatseh, S., Keyan, D., Akhtar, A., Abualhaija, A., Faroun, M., Aqel, I. S., Dardas, L., Afar, H., Servili, C., Hadzi-Pavlovic, D., van Ommeren, M., & Carswell, K. (2026). A guided chatbot-based psychological intervention for psychologically distressed older adolescents and young adults: A randomised clinical trial in Jordan. npj Digital Medicine, 9, Article 57. https://doi.org/10.1038/s41746-025-02142-8

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

From promise to practice: Artificial intelligence in mental health care in the MENA region

Where treatment gaps reach 80 to 95%, conversational AI could offer anonymous screening, psychoeducation and low-intensity support. The authors favour culturally adapted tools in hybrid care with human oversight.

El Hajj, S., Nsouli, A., Wehbe, M., Saad, G., & Maalouf, F. T. (2026). From promise to practice: Artificial intelligence in mental health care in the MENA region. Frontiers in Psychiatry, 17, Article 1878052. https://doi.org/10.3389/fpsyt.2026.1878052

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Implementation determinants for AI-enabled digital mental health interventions across Asian mental health systems: A scoping review using the Donabedian model

Maps the structural, process and outcome factors that decide whether AI mental health tools can be implemented across Asian health systems. Useful for planning scale-up rather than proving efficacy.

Ismail, M. F., Yusoff, M. I. M., Zulkifli, Z., Saad, A. S. H., Abidin, M. S. Z., & Rahim, A. I. A. (2026). Implementation determinants for AI-enabled digital mental health interventions across Asian mental health systems: A scoping review using the Donabedian model. Administration and Policy in Mental Health and Mental Health Services Research. Advance online publication. https://doi.org/10.1007/s10488-026-01524-8

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

AI-based approaches to refugee mental health care: Systematic integrative review

Of 23 studies, most used AI to detect or predict need, with fewer delivering interventions. Reported benefits were scalable support in resource-limited settings, early detection and reduced language barriers. Samples were small and access to technology unequal.

Kang, H., Park, S., Kim, J., Kim, O., Kim, H., Choi, J., & Chu, S. H. (2026). AI-based approaches to refugee mental health care: Systematic integrative review. Journal of Medical Internet Research, 28, e101857. https://doi.org/10.2196/101857

Open resource
Continuum
Also relevant to
READ
Evidence reviewNot reviewed

How effective is artificial intelligence (AI) in youth mental health interventions?

Orygen's living review of 12 controlled trials with 2,620 young people finds chatbots may reduce depression and anxiety, though certainty is low and safety is often unreported. Updated as new trials appear.

Randell, A., Bailey, A., Nikakis, Z., & Bartholomeusz, C. (2026). How effective is artificial intelligence (AI) in youth mental health interventions? [Living evidence review]. Orygen. https://www.orygen.org.au/Training/Resources/digital-technology/Evidence-review/How-effective-is-artificial-intelligence

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Systematic review and meta analysis of chatbots in the management of depressive and anxiety symptoms

Across 39 RCTs, chatbots reduced depression (g = 0.31) and anxiety (g = 0.28) compared with controls, with larger effects in clinical and subclinical groups. Most trials had a high risk of bias and few reported adverse events.

Sohn, J.-S., Ha, B.-G., Park, S., Kim, J.-J., Lee, E., Oh, H., Lee, S., & Kim, E. (2026). Systematic review and meta analysis of chatbots in the management of depressive and anxiety symptoms. npj Digital Medicine, 9, Article 377. https://doi.org/10.1038/s41746-026-02566-w

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Evaluation of a guided chatbot intervention for young people in Jordan: Feasibility randomized controlled trial

The feasibility trial that preceded the full STARS RCT. It found acceptable engagement and no adverse events, but was not powered to show effects.

de Graaff, A. M., Habashneh, R., Fanatseh, S., Keyan, D., Akhtar, A., Abualhaija, A., Faroun, M., Aqel, I. S., Dardas, L., Servili, C., van Ommeren, M., Bryant, R., & Carswell, K. (2025). Evaluation of a guided chatbot intervention for young people in Jordan: Feasibility randomized controlled trial. JMIR Mental Health, 12, e63515. https://doi.org/10.2196/63515

Open resource
Continuum
READ
ResearchPeer reviewedLived experience

"It happened to be the perfect thing": Experiences of generative AI chatbots for mental health

Nineteen users described improved mood, relationships and recovery from trauma or loss through generative AI chatbots, some calling it life-changing. Self-selected and mostly high-income, but a rare lived-experience account of benefit.

Siddals, S., Torous, J., & Coxon, A. (2024). "It happened to be the perfect thing": Experiences of generative AI chatbots for mental health. npj Mental Health Research, 3, Article 48. https://doi.org/10.1038/s44184-024-00097-4

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

The humanitarian AI paradox: Key opportunities, challenges and research needs for the use of AI in humanitarian mental health response

The authors report that 70% of aid workers use AI weekly and 69% rely on commercial platforms, while fewer than 25% of organisations have a formal AI strategy. They identify focused, non-specialised support as a near-term opportunity, and call for research, evaluation, safety measures and locally led governance to address risks such as bias, cultural and linguistic misinterpretation, consent and data governance.

van der Boor, C.F.; van der Boor, P.E.W. (2026). The humanitarian AI paradox: Key opportunities, challenges and research needs for the use of AI in humanitarian mental health response. Cambridge Prisms: Global Mental Health. https://doi.org/10.1017/gmh.2026.10254

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Supporting Student Mental Health With the Safespace Generative AI Chatbot: Mixed Methods Feasibility Study

This mixed methods feasibility study evaluated the Safespace generative AI chatbot with university students in Switzerland: 20 of 42 students used it over 2 to 4 weeks, generating 286 interactions. While 64% expected it to be helpful and 93% trusted its privacy safeguards, participants also noted technical and design issues, and the small sample means the findings are exploratory.

Pinna, M., Galletta, S., Ash, E. et al. (2026). Supporting Student Mental Health With the Safespace Generative AI Chatbot: Mixed Methods Feasibility Study. JMIR Formative Research. https://doi.org/10.2196/85427

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewedDeveloper research

Support-Seeking After Market Withdrawal of a Purpose-Built AI Well-Being Tool (Ash): Cross-Sectional Survey

In a survey of 393 UK adults who had used Ash within 90 days before its withdrawal in January 2026, general-purpose AI was the most common primary support source (27.1%), while 26.0% reported seeking no support. After losing access to Ash, 60.2% said their access to support had worsened; among respondents using general-purpose AI for mental health support, instant access (78.3%) and cost (73.9%) were the leading motivations.

Why it matters now: Rare evidence on what happens to users when a regulator-driven withdrawal removes a tool.

Stamatis, C. A., Kruzan, K. P., Hsu, J. et al. (2026). Support-Seeking After Market Withdrawal of a Purpose-Built AI Well-Being Tool (Ash): Cross-Sectional Survey. JMIR Medical Informatics. https://doi.org/10.2196/96515

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewedDeveloper research

Engagement Phenotypes for a Sample of 102,684 AI Mental Health Chatbot Users and Dose-Response Associations with Clinical Outcomes

Clustering data from 102,684 Ash chatbot users identified five engagement patterns: Early Dropouts (52.2%), Power Users (1.6%), Intensive Users (4.1%), Weekly Users (25.3%) and Concentrated Users (16.8%); 66.9% had at least one overnight session. Subsamples showed pre-post improvements in depression (d = -0.51), anxiety (d = -0.57) and social support (d = 0.22), while a model-predicted sample of 23,813 users showed larger depression improvements among more engaged users, particularly Power and Intensive Users.

Wolfe, E. C., Su, T., Tieleman, O. et al. (2026). Engagement Phenotypes for a Sample of 102,684 AI Mental Health Chatbot Users and Dose-Response Associations with Clinical Outcomes. arXiv, Computer Science > Human-Computer Interaction (cs.HC). https://doi.org/10.48550/arXiv.2605.00275

Open resource
Continuum
Also relevant to
READ
ResearchNot reviewed

Mental health atlas 2024

WHO's seventh Atlas, drawing on 144 countries. The global median is 13 mental health workers per 100,000 people, with extreme shortages in low- and middle-income countries. Spending ranges from US$65 per person in high-income countries to US$0.04 in low-income countries, and in low-income countries fewer than 10 per cent of people with a mental health condition receive care. New indicators cover telehealth.

Why it matters now: The baseline for any better-than-nothing argument: for much of the world, nothing is the current alternative.

World Health Organization. (2025). Mental health atlas 2024. WHO. https://www.who.int/publications/i/item/9789240114487

Open resource
Continuum
Also relevant to
READ
ResearchNot reviewed

World mental health today: latest data

WHO's companion report to the 2024 Atlas. In low-income countries there is little more than one mental health worker of any kind per 100,000 people, compared with more than 60 in high-income countries, and two-thirds of reporting countries have one psychiatrist for 200,000 people or more. It sees digital tools as key to scaling care for common conditions, while warning that relying too heavily on them risks excluding the most vulnerable.

Why it matters now: WHO's own framing: digital tools can scale access, but not if they leave the most excluded behind.

World Health Organization. (2025). World mental health today: latest data. WHO. https://www.who.int/publications/i/item/9789240113817

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

India's Tele-MANAS: evolution, early outcomes and a scalable blueprint for digital public mental health

NIMHANS authors describe Tele-MANAS, India's free 24-hour national tele-mental-health service in 20 languages, which handled about 3.43 million calls by March 2026. With about 0.75 psychiatrists per 100,000 people against a benchmark of three, the service uses stepped care: counsellors first, escalation to psychiatrists. The authors suggest carefully adding AI tools such as triage assistants or multilingual speech-to-text for counsellors, keeping human oversight at the core.

Why it matters now: How a low-workforce country frames AI: a layer inside public infrastructure, not a stand-alone app.

Suhas, S., Mittal, A., Manjunatha, N., Kumar, C. N., Math, S. B., & Murthy, P. (2026). India's Tele-MANAS: evolution, early outcomes and a scalable blueprint for digital public mental health. The British Journal of Psychiatry. https://doi.org/10.1192/bjp.2026.10733

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Utilizing AI chatbots for mental health in Indonesia: prospects and challenges

A research brief for Indonesia's House of Representatives. As of September 2025 there were 3,136 active psychologists against a national need of 9,947, and one psychiatrist per 200,000 people. It sees chatbots helping with education, early monitoring and referral, but warns of misdiagnosis, privacy and data-security risks, and concludes they should support rather than replace professionals, integrated into the national health system.

Why it matters now: How Indonesian lawmakers are being briefed on the issue.

Winurini, S. (2025). Utilizing AI chatbots for mental health in Indonesia: prospects and challenges. Info Singkat, XVII(19). Center for Parliamentary Analysis, DPR RI. https://berkas.dpr.go.id/pusaka/files/info_singkat/Info%20Singkat-XVII-19-I-P3DI-Oktober-2025-217-EN.pdf

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

Venting to the machine

Jakarta Post editorial. A survey of 400 Indonesian Gen Z found 60 per cent had confided personal issues to AI chatbots, seeking a safe, round-the-clock space free of judgement; some said a subscription was cheaper than counselling. The paper links this to Indonesia's shortage of clinical psychologists, concentrated in big cities.

Why it matters now: Use is already happening; the question is what it is replacing.

The Jakarta Post. (2026, July 30). Venting to the machine [Editorial]. https://www.thejakartapost.com/opinion/2026/07/30/venting-to-the-machine

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Remain cautious when seeking mental health advice with AI

Kompas reports Indonesian researchers at Yarsi University arguing for digital tools, including generative AI, given limited access to practitioners and geographic constraints. Their team is adapting Australia's THIS WAY UP programme, and a researcher notes that telecounselling and structured web resources remain the most helpful approaches.

Why it matters now: Local researchers choosing structured, tested tools over open chatbots.

Napitupulu, E. L. (2026, March 23). Remain cautious when seeking mental health advice with AI. Kompas. https://www.kompas.id/artikel/en-tetap-berhati-hati-saat-mencari-saran-kesehatan-mentai-dengan-kecerdasan-buatan

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

In Nigeria, artificial intelligence is plugging the mental health care gap

Radio segment (The World). Nigeria has fewer than 300 psychiatrists for more than 240 million people, mostly in cities. Some organisations have released AI chatbots for emotional support and links to resources; a psychiatrist from the Association of Psychiatrists in Nigeria discusses their flaws.

Why it matters now: A clinician's view from one of the lowest-workforce settings.

Beeler, C. (Host). (2026, February 27). In Nigeria, artificial intelligence is plugging the mental health care gap [Radio segment]. The World. https://theworld.org/segments/2026/02/27/in-nigeria-artificial-intelligence-is-plugging-the-mental-health-care-gap

Open resource
Continuum
WATCH
Practice signalNot reviewedIndustry funded

AI and the future of inclusive mental health care (AI for Good Global Summit 2025 workshop)

Session page with the recording of a workshop at the UN's AI for Good Global Summit on expanding mental health support for young people in under-resourced regions. Former WHO Director-General Margaret Chan says by video that technology alone has no purpose or compassion and must be used ethically and equitably. The other speakers are mostly from Chinese technology and education companies (Sun Media, New Oriental, Tencent Healthcare), so this is as much an industry showcase as a UN view.

AI for Good. (2025, December 7). AI and the future of inclusive mental health care [Session recording and summary]. ITU. https://aiforgood.itu.int/ai-and-the-future-of-inclusive-mental-health-care/

Open resource
Continuum
Also relevant to

B7. Clinical and administrative support

Can AI reduce paperwork, support measurement and free up clinical time?

READ
Evidence reviewPeer reviewed

Artificial intelligence–assisted mental healthcare in general practice: Benefits, risks and future possibilities

Australian GP-focused review by Grant Blashki, Ian Hickie and colleagues covering scribes, chatbots, digital therapies, just-in-time interventions, predictive analytics and agentic AI. Weighs 24/7 access and reach into underserved groups against digital literacy, unreliable outputs, bias, privacy and regulatory uncertainty, and keeps the GP-patient relationship central.

Blashki, G., Capon, W., Iorfino, F., & Hickie, I. B. (2026). Artificial intelligence–assisted mental healthcare in general practice: Benefits, risks and future possibilities. Australian Journal of General Practice, 55(6), 410-415. https://doi.org/10.31128/AJGP-04-25-7628

Open resource
Continuum
Also relevant to
READ
Tool or datasetNot reviewedDeveloper research

Mentalyc: AI Documentation and Practice Tools

Vendor information: Mentalyc markets AI-generated clinical notes, treatment plans and session insights for mental-health professionals. Included as an example of administrative-support tools, not as independent evidence of improved care or workforce productivity.

Mentalyc. (n.d.). Mentalyc [vendor product website]. Accessed 22 September 2026. https://www.mentalyc.com/

Open resource
Also relevant to
READ
Tool or datasetNot reviewedDeveloper research

BastionGPT: Healthcare AI Workspace

Vendor information: BastionGPT markets a healthcare AI workspace with language-model assistance, transcription and documentation functions. A commercial example of support for healthcare work, not a psychiatry-specific intervention or independently validated population-health service.

BastionGPT. (n.d.). BastionGPT: Medical ChatGPT for Healthcare [vendor product website]. Accessed 22 September 2026. https://bastiongpt.com/

Open resource
Also relevant to
READ
Evidence reviewPeer reviewed

Artificial intelligence for mental health: A narrative review of applications, challenges, and future directions in digital health

Narrative review of AI applications in mental health across screening, monitoring, conversational support and clinical workflow, with attention to challenges and future directions.

Ali, M., Ali, S., Abbas, Q., Abbas, Z., & Lee, S. W. (2025). Artificial intelligence for mental health: A narrative review of applications, challenges, and future directions in digital health. Digital Health, 11, 20552076251395548. https://doi.org/10.1177/20552076251395548

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

A Scoping Review of AI-Driven Digital Interventions in Mental Health Care: Mapping Applications Across Screening, Support, Monitoring, Prevention, and Clinical Education

Scoping review mapping AI-driven digital interventions across screening, support, monitoring, prevention and clinical education.

Ni, Y., & Jia, F. (2025). A Scoping Review of AI-Driven Digital Interventions in Mental Health Care: Mapping Applications Across Screening, Support, Monitoring, Prevention, and Clinical Education. Healthcare, 13(10), 1205. https://doi.org/10.3390/healthcare13101205

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

The application of artificial intelligence in the field of mental health: a systematic review

Systematic review of AI applications in mental health, published in BMC Psychiatry.

Dehbozorgi, R., Zangeneh, S., Khooshab, E., Nia, D. H., Hanif, H. R., Samian, P., ... Lohrasebi, F. (2025). The application of artificial intelligence in the field of mental health: a systematic review. BMC Psychiatry, 25(1), 132. https://doi.org/10.1186/s12888-025-06483-2

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Artificial intelligence in mental health care: a systematic review of diagnosis, monitoring, and intervention applications

Systematic review across diagnosis, monitoring and intervention, with a frank note that external validation of AI mental health models remains rare.

Cruz-Gonzalez, P., He, A. W., Lam, E. P., Ng, I. M. C., Li, M. W., Hou, R., Chan, J. N., Sahni, Y., Vinas Guasch, N., Miller, T., Lau, B. W., & Sánchez Vidaña, D. I. (2025). Artificial intelligence in mental health care: a systematic review of diagnosis, monitoring, and intervention applications. Psychological Medicine, 55, e18. https://doi.org/10.1017/s0033291724003295

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Exploring the Role of Artificial Intelligence in Mental Healthcare: Current Trends and Future Directions - A Narrative Review for a Comprehensive Insight

Single-author narrative review of current trends and future directions; useful as an accessible orientation rather than a rigorous synthesis.

Alhuwaydi, A. M. (2024). Exploring the Role of Artificial Intelligence in Mental Healthcare: Current Trends and Future Directions - A Narrative Review for a Comprehensive Insight. Risk Management and Healthcare Policy, 17, 1339-1348. https://doi.org/10.2147/rmhp.s461562

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

ACRRM: Artificial Intelligence in Rural and Remote General Practice

Practical resource from the Australian College of Rural and Remote Medicine on how rural doctors are using AI (scribing, summarisation, decision support, care coordination, referrals, administration) with clinician oversight, bias monitoring, cultural safety and compliance with Australian privacy and safety standards. Includes webinars, podcasts and a factsheet.

Australian College of Rural and Remote Medicine. (n.d.). Artificial intelligence in rural and remote general practice. https://www.acrrm.org.au/resources/digital-health/artificial-intelligence

Open resource
Continuum
Also relevant to
READ
Evidence reviewNot reviewed

Artificial Intelligence and Machine Learning in Mental Health Services: An Environmental Scan

Technology-landscape review and stakeholder consultations mapping emerging AI and machine-learning uses in mental-health services, including conversational agents, assessment, prediction, public health and administration. Useful background for service planning, with many applications then still in research and development.

Wells, C., Hill, S., & Argaez, C. (2021, June). Artificial Intelligence and Machine Learning in Mental Health Services: An Environmental Scan. CADTH Health Technology Review. https://www.cda-amc.ca/sites/default/files/attachments/2021-06/artificial_intelligence_and_machine_learning_in_mental_health_services_environmental_scan.pdf

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Artificial Intelligence for Mental Health Monitoring: A Solution for Digital Behavioral Health Care and Education-An Umbrella Review

Umbrella review of reviews on AI for mental health monitoring in behavioural health care and education, framed against workforce shortages and system strain. Broad, and a reasonable entry point to the monitoring literature.

Yeasmin, S., Semi, M. M. A., Rony, M. K. K., Das, S., Sabeena, A. A., Rahman, R., Biswas, B., Ahmed, F., & Hossain, A. (2026). Artificial Intelligence for Mental Health Monitoring: A Solution for Digital Behavioral Health Care and Education-An Umbrella Review. Health Science Reports, 9(1), e71703. https://doi.org/10.1002/hsr2.71703

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Context-aware monitoring: rethinking comprehensive screening in the era of AI

Argues for context-aware monitoring that uses within-person change rather than population norms, with Bayesian approaches to reduce false positives in serial multimodal screening.

Shaywitz, D. A., Price, N. D., & Sodickson, D. K. (2026). Context-aware monitoring: rethinking comprehensive screening in the era of AI. npj Digital Medicine, 9(1), 624. https://doi.org/10.1038/s41746-026-03106-2

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Explainable AI-based suicidal and non-suicidal ideations detection from social media text with enhanced ensemble technique

Machine learning study distinguishing suicidal from non-suicidal ideation in social media text with an explainable ensemble model. Included as an example of what population-scale social media surveillance for mental health looks like technically, and therefore what responsible surveillance frameworks have to govern.

Alghazzawi, D., Ullah, H., Tabassum, N., Badri, S. K., & Asghar, M. Z. (2025). Explainable AI-based suicidal and non-suicidal ideations detection from social media text with enhanced ensemble technique. Scientific Reports, 15(1), 1111. https://doi.org/10.1038/s41598-024-84275-6

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Artificial Intelligence Powered Wearable and Portable Devices for Remote Cardiac Care and Population Health

Review of AI-powered wearable and portable devices for remote cardiac care and population health, including smartwatch AF screening, deep-learning ECG and cuffless blood pressure.

De Armas, R. E., & Singh, J. P. (2026). Artificial Intelligence Powered Wearable and Portable Devices for Remote Cardiac Care and Population Health. Current Cardiology Reports, 28(1), 46. https://doi.org/10.1007/s11886-026-02358-4

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

The Emergence of AI-Based Wearable Sensors for Digital Health Technology: A Review

Review of AI-based wearable sensors and flexible electronics for continuous, non-invasive health monitoring and pre-diagnosis.

Shajari, S., Kuruvinashetti, K., Komeili, A., & Sundararaj, U. (2023). The Emergence of AI-Based Wearable Sensors for Digital Health Technology: A Review. Sensors, 23(23), 9498. https://doi.org/10.3390/s23239498

Open resource
Continuum
READ
Evidence reviewPeer reviewed

Evolution of Wearable Devices with Real-Time Disease Monitoring for Personalized Healthcare

Foundational review of the evolution of wearable devices with real-time disease monitoring for personalised healthcare.

Guk, K., Han, G., Lim, J., Jeong, K., Kang, T., Lim, E. K., & Jung, J. (2019). Evolution of Wearable Devices with Real-Time Disease Monitoring for Personalized Healthcare. Nanomaterials, 9(6), 813. https://doi.org/10.3390/nano9060813

Open resource
Continuum
READ
Evidence reviewPeer reviewed

Artificial Intelligence (AI) and Robotics in Elderly Healthcare: Enabling Independence and Quality of Life

Review of AI and robotics in elderly healthcare, including wearables for continuous tracking, independence and quality of life.

Padhan, S., Mohapatra, A., Ramasamy, S. K., & Agrawal, S. (2023). Artificial Intelligence (AI) and Robotics in Elderly Healthcare: Enabling Independence and Quality of Life. Cureus. https://doi.org/10.7759/cureus.42905

Open resource
Continuum
READ
Opinion and commentaryPeer reviewed

Automated Opioid Risk Scores: Machine Learning-Induced Epistemic Injustice

Philosophical analysis of how automated opioid-risk scoring can shape whose explanations count in healthcare. Useful for designing systems that preserve patient and community knowledge; it is not an empirical study of cross-cultural mental-health misdiagnosis.

Pozzi, G. (2023). Automated opioid risk scores: a case for machine learning-induced epistemic injustice in healthcare. Ethics and Information Technology, 25, 3. https://doi.org/10.1007/s10676-023-09676-z

Open resource
READ
ResearchPreprint, not yet reviewed

Small Language Models for Clinical Extraction in Low-Resource Languages

Preprint evaluating a translation-and-extraction pipeline on 1,221 anonymised Persian palliative-care call transcripts. Shows potential for smaller multilingual models in clinical information processing, with psychological complaints remaining challenging. It does not demonstrate mental-health treatment benefit or guarantee data sovereignty.

Ghaffarzadeh-Esfahani, M., Yousefian, N., Heidari-Farsani, E., Omidvarian, A. A., Ghahraei, S., Farangi, A., & Boroumand, A. (2026). Small Language Models for Privacy-Preserving Clinical Information Extraction in Low-Resource Languages. arXiv preprint, arXiv:2602.21374. https://arxiv.org/abs/2602.21374

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Retrieval-Augmented Generation in Medicine and Nursing: Scoping Review

Scoping review of 67 studies on connecting language models to external knowledge in medicine and nursing. Relevant to evidence-grounded health information and resource navigation, but few systems aligned with real-world workflows. The review does not show that retrieval eliminates hallucinations or guarantees current, correct referrals.

Miao, Y., Zhao, Y., Luo, Y., Wang, H., & Wu, Y. (2025). Improving Large Language Model Applications in the Medical and Nursing Domains With Retrieval-Augmented Generation: Scoping Review. Journal of Medical Internet Research, 27, e80557. https://doi.org/10.2196/80557

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

TPC-GCN: Deep learning for pulse pattern classification in traditional Chinese medicine

A technical study using a graph convolutional network, feature extraction and data augmentation to classify traditional Chinese medicine pulse patterns. The authors report 91.68% classification accuracy on their pulse dataset. It documents AI being applied within a traditional diagnostic framework, but does not validate pulse diagnosis as a substitute for mental-health assessment.

Li, H., Li, Y., Zhang, Z., Xue, C., Li, Z., Li, X., & Men, J. (2025). TPC-GCN: Deep learning for pulse pattern classification in traditional Chinese medicine. Medical Engineering & Physics, 144, 104401. https://doi.org/10.1016/j.medengphy.2025.104401

Open resource
Also relevant to
READ
Evidence reviewPeer reviewed

Artificial Intelligence and Digital Technologies in Prakriti Assessment: Toward Standardized Evidence-Based Ayurvedic Practice

A narrative review examining sensors, image analysis, AI and machine learning for assessing Prakriti, the Ayurvedic concept of individual constitution. It discusses opportunities to make assessment more reproducible and identifies problems with data quality, interoperability, interpretability and domain-specific validation. It does not demonstrate that AI-generated Dosha-based lifestyle advice improves mental health.

Allam, R., Kothainayagi, B., Subbiah, S., Sarkar, S., & R, S. (2026). Artificial Intelligence and Digital Technologies in Prakriti Assessment: Toward Standardized Evidence-Based Ayurvedic Practice. Journal of Ayurveda and Integrative Medicine, 17(5), 101372. https://doi.org/10.1016/j.jaim.2026.101372

Open resource
Also relevant to
READ
Guidance and standardsNot reviewed

Using AI in psychological practice: How the new APS practice guidelines help psychologists

This Australian Psychological Society article explains its practice guidelines for psychologists using AI. The guidelines emphasise professional accountability, evidence-based scrutiny, ongoing consent, secure data handling, attention to bias and appropriate competence.

Australian Psychological Society. (2026, February 11). Using AI in psychological practice: How the new APS practice guidelines help psychologists. https://psychology.org.au/insights/use-ai-in-practice-new-aps-practice-guidelines

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Clinician and simulated patient perspectives on ambient AI scribes in psychiatric consultations: a qualitative study

The study used focus groups after simulated psychiatric consultations with 8 psychiatrists and 4 simulated patients, identifying 6 themes including reduced documentation burden, trust, and privacy concerns. The authors suggest ambient AI scribes may support more patient-centred consultations, with transparent consent, clinician oversight, and further evaluation with real patients needed.

Bokhari, S.A., Nawaz, F.A., Usman, F.M., et al. (2026). Clinician and simulated patient perspectives on ambient AI scribes in psychiatric consultations: a qualitative study. Frontiers in Psychiatry. https://doi.org/10.3389/fpsyt.2026.1821065

Open resource
Continuum
Also relevant to
READ
Position statementNot reviewed

AI Scribes in Mental Health Nursing Practice, Informed Consent and Equitable Access to Care

The statement says access to clinically indicated care should not depend on agreeing to use an AI scribe, and that declining or withdrawing consent should not lead to penalties, delayed or reduced care, or extra fees. It calls for specific, informed and voluntary consent, a non-AI documentation option, strong clinical governance, access to and correction of records, and nursing leadership, while noting concerns about privacy, discrimination, trust, safety and cybersecurity and limited Australian evidence specific to mental health nursing.

Why it matters now: Australian professional position on consent for ambient scribes.

Australian College of Mental Health Nurses (ACMHN) (2026). AI Scribes in Mental Health Nursing Practice, Informed Consent and Equitable Access to Care. ACMHN, Policy & Advocacy. https://acmhn.org/Web/Policy/ACMHN-Statement-on-AI-Scribes-in-Mental-Health-Nursing-Practice--Informed-Consent-and-Equitable-Acce.aspx

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Precision psychiatry in clinical practice: What is clinically actionable, what is promising, and what remains experimental?

This structured narrative review assessed precision psychiatry across eight readiness domains, finding that measurement-based care, therapeutic drug monitoring for selected indications, and some pharmacogenetic gene-drug interactions are actionable in selected settings. Digital phenotyping and AI-assisted tools remain promising but are not ready for routine use, while approaches such as polygenic risk scores, multiomic subtyping, and automated treatment selection are largely experimental.

International Review of Psychiatry (2026). Precision psychiatry in clinical practice: What is clinically actionable, what is promising, and what remains experimental?. International Review of Psychiatry. https://doi.org/10.1080/09540261.2026.2722686

Open resource
Continuum
Also relevant to
READ
Guidance and standardsPeer reviewed

Generative Artificial Intelligence in Mental Health: A Guide for Primary Care Providers

This article reviews generative AI applications in mental health, including patient-facing chatbots, clinician decision-support tools and suicide-risk assessment, alongside safeguards for primary care. A meta-analysis of 18 randomised controlled trials with 3,477 participants found modest improvements in depression (g=−0.26) and anxiety (g=−0.19), with no substantial effects persisting at three-month follow-up; the article also discusses safety, privacy, bias and clinical-validation concerns.

Nadkarni, A., Faust, K., Matta, S. E. et al. (2026). Generative Artificial Intelligence in Mental Health: A Guide for Primary Care Providers. Primary Care Companion for CNS Disorders. https://doi.org/10.4088/PCC.26f04201

Open resource
Continuum
Also relevant to

B7.1 Personal and device data

Can data from phones and wearables help people and services plan and time support? Read with C2 (data and sovereignty) and R4 (devices).

READ
Evidence reviewPeer reviewed

Digital phenotyping for mental health conditions: A systematic review of implementation and application

Reviews 47 studies on how phone and wearable data are being used to monitor symptoms and flag relapse for clinical review or outreach. Focuses on what it takes to implement in real services.

Alam, N. B., Haque, T., Subedar, S., Giacco, D., Singh, S. P., & Jilka, S. (2026). Digital phenotyping for mental health conditions: A systematic review of implementation and application. Frontiers in Digital Health, 8, Article 1772744. https://doi.org/10.3389/fdgth.2026.1772744

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Digital phenotyping for predicting relapse in psychiatric disorders: A systematic review of passive sensing approaches

Across 52 studies and 4,814 people, phone and wearable data predicted relapse one day to four weeks ahead with moderate to good accuracy (AUC 0.70 to 0.88). Three-quarters of studies were at high risk of bias, and none showed that acting on predictions improved outcomes.

Fang, S.-S., & Chen, S.-H. (2026). Digital phenotyping for predicting relapse in psychiatric disorders: A systematic review of passive sensing approaches. BMC Psychiatry, 26, Article 507. https://doi.org/10.1186/s12888-026-08157-z

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Multimodal personalized mobile health just-in-time adaptive intervention for occupational stress management: Pilot study

Forty-six employees used a stress app linked to a heart rate sensor that timed support. Some stress measures fell, but there was no control group and sensor data were only 45% to 60% complete.

Kunas, B., Jung, O., Schranz, C., Schmoigl-Tonis, M., Mehlis, J., & Laireiter, A.-R. (2026). Multimodal personalized mobile health just-in-time adaptive intervention for occupational stress management: Pilot study. JMIR mHealth and uHealth, 14, Article e79642. https://doi.org/10.2196/79642

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

Leveraging AI to integrate behavioral healthcare

Describes AI monitoring symptoms through phone sensors and frequent check-ins, and predicting relapse, in integrated primary care. Notes many products are untested.

Mattes, G. (2026, April 17). Leveraging AI to integrate behavioral healthcare. Agency for Healthcare Research and Quality, The Academy for Integrating Behavioral Health and Primary Care. https://integrationacademy.ahrq.gov/news-and-events/news/leveraging-ai-integrate-behavioral-healthcare

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

Digital mental health technology: Qualification and classification

UK guidance on when a digital mental health tool, including AI and monitoring functions, counts as a regulated medical device. First published February 2025, updated July 2026. Needed for anyone turning device data into clinical decisions.

Medicines and Healthcare products Regulatory Agency. (2026). Digital mental health technology: Qualification and classification. GOV.UK. https://www.gov.uk/government/publications/digital-mental-health-technology-qualification-and-classification

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewedDeveloper research

Transforming wearable data into personal health insights using large language model agents

An AI agent that analyses a person's own wearable data to answer health questions was 84% accurate on 4,000 queries, against 74% for a simpler approach. Shows how wearable data might become usable for people and clinicians. Developer research (Google).

Merrill, M. A., Paruchuri, A., Rezaei, N., Kovacs, G., Perez, J., Liu, Y., Schenck, E., Hammerquist, N., Sunshine, J., Tailor, S., Ayush, K., Su, H.-W., He, Q., McLean, C. Y., Malhotra, M., Patel, S., Zhan, J., Althoff, T., McDuff, D., & Liu, X. (2026). Transforming wearable data into personal health insights using large language model agents. Nature Communications, 17, Article 1143. https://doi.org/10.1038/s41467-025-67922-y

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Building the foundations for global data banking in digital phenotyping for mental health

Australian authors set out what is needed for shared data banks of phone and wearable data so findings can be tested across populations. Raises consent, governance and data stewardship, which link to data sovereignty.

O'Dea, B., Naz, S., McLoughlin, L. T., & Whitton, A. E. (2026). Building the foundations for global data banking in digital phenotyping for mental health. NPP—Digital Psychiatry and Neuroscience, 4, Article 15. https://doi.org/10.1038/s44277-026-00066-z

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Data-driven personalization of just-in-time adaptive mental health intervention via two-stage reinforcement learning approach

Used data from a six-week trial (190 people) to learn when and what type of support to deliver to improve next-day depression, anxiety and stress. Outperformed rule-based timing in modelling, but not yet tested live.

Park, C., Heo, J., Park, J. Y., Jeon, M., Do, G., Park, S., Jung, D., & Lim, M. H. (2026). Data-driven personalization of just-in-time adaptive mental health intervention via two-stage reinforcement learning approach. npj Digital Medicine. Advance online publication. https://doi.org/10.1038/s41746-026-03086-3

Open resource
Continuum
READ
Guidance and standardsNot reviewed

Advisory: How wearables and other digital tools are driving innovation in behavioral health

US government advisory on bringing wearables into behavioural health care and self-management, with attention to ethics and privacy. Last updated July 2026.

Substance Abuse and Mental Health Services Administration. (2026). Advisory: How wearables and other digital tools are driving innovation in behavioral health. https://www.samhsa.gov/resource/ebp/advisory-wearables-other-digital-tools-driving-innovation-behavioral-health

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Adherence and acceptability of smartphone- and wearable-based digital phenotyping in patients with first-episode psychosis

Forty young people with first-episode psychosis kept wearables on well (about 85% of heart rate data) but found daily surveys more burdensome (71%). Practical evidence on what people will actually wear and answer.

Yang, L., Ross, K., Palmer, M., Ye, Q., Olson, D., Abidi, S., Morrison, J., Tibbo, P. G., & Wang, J. (2026). Adherence and acceptability of smartphone- and wearable-based digital phenotyping in patients with first-episode psychosis. The Canadian Journal of Psychiatry. Advance online publication. https://doi.org/10.1177/07067437261486958

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Digital phenotyping using smartphones could help steer mental health treatment

A plain-language overview of how phone and wearable data might flag relapse or panic and help clinicians decide when to check in. A good entry point for non-specialists.

Adam, D. (2025). Digital phenotyping using smartphones could help steer mental health treatment. Proceedings of the National Academy of Sciences, 122(14), Article e2505700122. https://doi.org/10.1073/pnas.2505700122

Open resource
Continuum
READ
Policy and regulationNot reviewed

Opportunities for technology-enabled care: Economic and payment issues

Looks at how payment systems shape whether AI, remote monitoring and wearables get used. Funding models matter as much as the technology.

Why it matters now: It shows payment arrangements may shape whether remote monitoring and wearable-enabled care are adopted.

Anderson, J. P., Couture, S. J., Louis, J. F., Wiltz, J., Sheingold, S., & Smith, S. R. (2025, December 4). Opportunities for technology-enabled care: Economic and payment issues. Office of the Assistant Secretary for Planning and Evaluation, U.S. Department of Health and Human Services. https://aspe.hhs.gov/reports/opportunities-technology-enabled-care

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Utility of consumer-grade wearable devices for inferring physical and mental health outcomes in severe mental illness: Systematic review

Across 12 studies with more than 500 people living with severe mental illness, ordinary wearables showed links between activity, heart rate and sleep and symptoms. Relevant to the physical health gap in severe mental illness. Evidence is still limited.

Hassan, L., Milton, A., Sawyer, C., Casson, A. J., Torous, J., Davies, A., Ruiz-Yu, B., & Firth, J. (2025). Utility of consumer-grade wearable devices for inferring physical and mental health outcomes in severe mental illness: Systematic review. JMIR Mental Health, 12, Article e65143. https://doi.org/10.2196/65143

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewedDeveloper research

A personal health large language model for sleep and fitness coaching

A language model tuned to read wearable sleep and fitness data and give coaching. Its sleep advice was rated close to human experts (4.61 vs 4.75 out of 5), and it predicted self-reported sleep outcomes from sensor data. Developer research (Google), focused on wellbeing, not mental illness.

Khasentino, J., Belyaeva, A., Liu, X., Yang, Z., Furlotte, N. A., Lee, C., Schenck, E., Patel, Y., Cui, J., Schneider, L. D., Bryant, R., Gomes, R. G., Jiang, A., Lee, R., Liu, Y., Perez, J., Rogers, J. K., Speed, C., Tailor, S., . . . McLean, C. Y. (2025). A personal health large language model for sleep and fitness coaching. Nature Medicine, 31(10), 3394–3403. https://doi.org/10.1038/s41591-025-03888-0

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

The application of AI to ecological momentary assessment data in suicide research: Systematic review

Across 12 studies with 4,398 people, AI applied to frequent self-reports predicted near-term suicidal thinking with moderate accuracy (AUC 0.74 to 0.86). Shows potential for timely outreach but nowhere near reliable enough for decisions on its own.

Melia, R., Musacchio Schafer, K., Rogers, M. L., Wilson-Lemoine, E., & Joiner, T. E. (2025). The application of AI to ecological momentary assessment data in suicide research: Systematic review. Journal of Medical Internet Research, 27, Article e63192. https://doi.org/10.2196/63192

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPreprint, not yet reviewed

Position: AI will transform neuropsychology through mental health digital twins for dynamic mental health care, especially for ADHD

Argues for 'digital twins' that track a person's symptom patterns over time from frequent check-ins to personalise care. A position paper, useful for seeing where the field wants to go.

Why it matters now: It sketches a proposed route to personalised ADHD care using frequent symptom check-ins and digital twins.

Natarajan, N., Viswanathan, S., Roberts-Gaal, X., & Martel, M. M. (2025). Position: AI will transform neuropsychology through mental health digital twins for dynamic mental health care, especially for ADHD [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2510.07409

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Digital health monitoring for adults with treatment-resistant depression: Observational feasibility study protocol

Protocol to give 200 adults with treatment-resistant depression a smart ring plus regular online check-ins, to build personal profiles that may guide care. No results yet.

Parkington, K., Aguilar, R. J., Lee, G. H., Rueda, A., Adamsahib, F., Al-Hassan, M., Famuditimi, S., Menzies, P., Lin, Q., Lou, W., Krishnan, S., & Bhat, V. (2025). Digital health monitoring for adults with treatment-resistant depression: Observational feasibility study protocol. PLOS ONE, 20(10), Article e0333484. https://doi.org/10.1371/journal.pone.0333484

Open resource
Continuum
READ
Evidence reviewPeer reviewed

Passive sensing for mental health monitoring using machine learning with wearables and smartphones: Scoping review

Maps 42 studies using phone and wearable data with machine learning to monitor diagnosed mental health conditions. High accuracy figures appear in some small studies, but most have not been tested in practice.

Shen, S., Qi, W., Zeng, J., Li, S., Liu, X., Zhu, X., Dong, C., Wang, B., Shi, Y., Yao, J., Wang, B., Lou, X., Gu, S., Li, P., Wang, J., Jiang, G., & Cao, S. (2025). Passive sensing for mental health monitoring using machine learning with wearables and smartphones: Scoping review. Journal of Medical Internet Research, 27, Article e77066. https://doi.org/10.2196/77066

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPreprint, not yet reviewed

Envisioning an AI-enhanced mental health ecosystem

Proposes a hybrid system where AI supports, not replaces, human providers through peer support, self-help, proactive monitoring and data insights. Conceptual; a CHI 2025 workshop paper.

Why it matters now: It argues for AI supporting human providers through monitoring, peer support and self-help rather than replacing them.

Sim, K. Y. H., & Choo, K. T. W. (2025). Envisioning an AI-enhanced mental health ecosystem [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2503.14883

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

Leveraging large language models for digital phenotyping: Detecting depressive state changes for patients with depressive episodes

Tested whether language models can read phone sensor data from 84 Finnish patients to detect changes in depression. They beat a standard baseline, but overall accuracy was low (macro F1 0.40).

Yuan, Y., Gao, Y., Moen, H., Isometsä, E., Marttinen, P., & Aledavood, T. (2025). Leveraging large language models for digital phenotyping: Detecting depressive state changes for patients with depressive episodes [Preprint]. medRxiv. https://doi.org/10.1101/2025.05.10.25327354

Open resource
Continuum
READ
Policy and regulationNot reviewed

Vision: Virtual care strategy

Victoria's strategy for telehealth, apps, wearables and remote monitoring across the health system. Sets the Australian state policy context for device-linked care. Reports no outcomes.

Department of Health, Victoria. (2024, December 20). Vision: Virtual care strategy. https://www.health.vic.gov.au/victorian-virtual-care-strategy/vision

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

Digital phenotyping from wearables using AI characterizes psychiatric disorders and identifies genetic associations

Fitbit data from adolescents in a large US cohort distinguished ADHD (AUROC 0.89) and anxiety (0.71) from peers. Research-stage classification, with no link to intervention yet.

Liu, J., Borsari, B., Li, Y., Liu, S., Gao, Y., Xin, X., Lou, S., Jensen, M., Garrido-Martín, D., Verplaetse, T., Ash, G., Zhang, J., Girgenti, M. J., Roberts, W., & Gerstein, M. (2024). Digital phenotyping from wearables using AI characterizes psychiatric disorders and identifies genetic associations [Preprint]. medRxiv. https://doi.org/10.1101/2024.09.23.24314219

Open resource
Continuum
READ
ResearchPeer reviewed

Relapse prediction in schizophrenia with smartphone digital phenotyping during COVID-19: A prospective, three-site, two-country, longitudinal study

Smartphone data from 132 people in the US and India flagged behaviour changes that were twice as frequent in the month before relapse. Only 6.9% of flags were true positives and data quality was patchy, so this supports earlier check-ins, not automated decisions.

Cohen, A., Naslund, J. A., Chang, S., Nagendra, S., Bhan, A., Rozatkar, A., Thirthalli, J., Bondre, A., Tugnawat, D., Reddy, P. V., Dutt, S., Choudhary, S., Chand, P. K., Patel, V., Keshavan, M., Joshi, D., Mehta, U. M., & Torous, J. (2023). Relapse prediction in schizophrenia with smartphone digital phenotyping during COVID-19: A prospective, three-site, two-country, longitudinal study. Schizophrenia, 9, Article 6. https://doi.org/10.1038/s41537-023-00332-5

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Potential usefulness of apps and other digital technologies for improving access to behavioral health in primary care

Notes that only 15% of popular mental health apps collected passive sensor data and 2% linked to health records. A reminder that device data rarely reach clinicians in practice.

Thornton, M., Moran, G. E., & Duncan, M. H. (2023). Potential usefulness of apps and other digital technologies for improving access to behavioral health in primary care. Agency for Healthcare Research and Quality. https://integrationacademy.ahrq.gov/products/topic-briefs/behavioral-health-apps

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

From promise to practice: Towards the realisation of AI-informed mental health care

A widely cited roadmap arguing measurement-based care with phone and wearable data could be the foundation for AI in mental health. Says the evidence does not yet support these signals replacing clinical assessment.

Koutsouleris, N., Hauser, T. U., Skvortsova, V., & De Choudhury, M. (2022). From promise to practice: Towards the realisation of AI-informed mental health care. The Lancet Digital Health, 4(11), e829–e840. https://doi.org/10.1016/S2589-7500(22)00153-4

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

The Digital Mental Health Study assessed feasibility of large-scale digital sensing for depression and anxiety

The study collected up to 12 months of iPhone and Apple Watch sensor data from more than 4,000 participants, alongside self-reports and interviews about depression, anxiety and perceived stress. It describes the study’s recruitment and methods, and reports initial symptom trajectories and high participant engagement over 12 months.

Douglas, C. S., Congdon, E., Huang, C. et al. (2026). The Digital Mental Health Study assessed feasibility of large-scale digital sensing for depression and anxiety. NPJ Digital Public Health. https://doi.org/10.1038/s44482-026-00029-3

Open resource
Continuum
READ
Evidence reviewPeer reviewed

Online Behavioral Digital Phenotyping for Depression: Scoping Review of Behavioral Features and Exploratory Assessment of Research Domain Criteria Construct Relevance

This scoping review included 74 studies and identified five online behavioral modalities, with 36 core features. Although 31 studies (41.9%) reported findings linked to individual features, no feature showed a consistently directed association with depression-related outcomes; exploratory RDoC relevance was identified but not validated.

Ren, Q., Chi, J., Li, Y. et al. (2026). Online Behavioral Digital Phenotyping for Depression: Scoping Review of Behavioral Features and Exploratory Assessment of Research Domain Criteria Construct Relevance. Journal of Medical Internet Research. https://doi.org/10.2196/101329

Open resource
Continuum
Also relevant to

B8. Equity and reach

Who gains access that they did not have before: rural, low-income, other-language and low- and middle-income country populations?

READ
Evidence reviewPeer reviewed

Effectiveness of Digital and Remote Health Interventions in Pediatric Populations From Underserved or Rural Areas: A Systematic Review of Randomized Controlled Trials

Systematic review of randomised trials of digital and remote health interventions for children and adolescents in underserved or rural areas.

Beltrán, V., Abarza, L., Acuña-Mardones, P., Espinoza, I., Chaple Gil, A., Morales-Gómez, C., ... Fernández, E. (2026). Effectiveness of Digital and Remote Health Interventions in Pediatric Populations From Underserved or Rural Areas: A Systematic Review of Randomized Controlled Trials. International Journal of Telemedicine and Applications, 2026(1), 5586850. https://doi.org/10.1155/ijta/5586850

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Adolescent mental health in rural settings: the role of artificial intelligence and community engagement

Perspective on adolescent mental health in rural settings and the combined role of AI and community engagement.

Grenon, S., Hoang, K. N., Luo, S., & Koch, A. (2025). Adolescent mental health in rural settings: the role of artificial intelligence and community engagement. Frontiers in Public Health, 13, 1643466. https://doi.org/10.3389/fpubh.2025.1643466

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPreprint, not yet reviewed

Artificial intelligence in rural healthcare delivery: Bridging gaps and enhancing equity through innovation

Preprint on how AI could address access gaps and equity in rural healthcare delivery.

Norful, A. A., de Jacq, K., Roome, R., & Sickinger, B. D. (2025). Artificial intelligence in rural healthcare delivery: Bridging gaps and enhancing equity through innovation. arXiv preprint arXiv:2508.11738. https://arxiv.org/abs/2508.11738

Open resource
Continuum
READ
Opinion and commentaryPeer reviewed

The pieces are already built: converging India's national assets to close the mental health treatment gap

Argues that India's existing national digital and health assets can be converged to close the mental health treatment gap.

Sriram, S. (2026). The pieces are already built: converging India's national assets to close the mental health treatment gap. Frontiers in Digital Health, 8, 1915080. https://doi.org/10.3389/fdgth.2026.1915080

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Global Disparities in Mental Health Systems: A Comparative Cross-sectional Study of Ten Countries with Different Income Levels

Comparative cross-sectional study of mental health system resources across ten countries at different income levels.

Patel, N. M., Savaliya, G. V., Mehta, P. J., & Kataria, L. R. (2025). Global Disparities in Mental Health Systems: A Comparative Cross-sectional Study of Ten Countries with Different Income Levels. Indian Journal of Psychological Medicine, 02537176251379999. https://doi.org/10.1177/02537176251379999

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Developing a socio-ecological model for community engagement in a health programme in an underserved urban area

Socio-ecological model of community engagement in a health programme in an underserved urban area, highlighting trust, social wellbeing and community identity.

Caperon, L., Saville, F., & Ahern, S. (2022). Developing a socio-ecological model for community engagement in a health programme in an underserved urban area. PLOS ONE, 17(9), e0275092. https://doi.org/10.1371/journal.pone.0275092

Open resource
Continuum
READ
Tool or datasetNot reviewed

BHASHINI: Multilingual Digital Access

Government language infrastructure offering speech recognition, translation, models, datasets and APIs for access to digital services in Indian languages. Potential enabling infrastructure for multilingual health information and campaigns; the website does not demonstrate culturally validated mental-health support.

Government of India. (n.d.). BHASHINI. Programme website. Accessed 22 September 2026. https://bhashini.gov.in/

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewedLived experience

Using AI to Address Mental Health Inequalities: Citizen Engagement and Co-Creation

Community-based participatory study involving peer researchers and adults with mental-health challenges. Interviews and focus groups around a co-created future-scenario film explored power, inclusion and having a voice in AI development. Supports participatory design, not proven benefits from a deployed algorithm.

Morgan, P., & Cogan, N. A. (2024). Using artificial intelligence to address mental health inequalities: co-creating machine learning algorithms with key stakeholders and citizen engagement. Journal of Public Mental Health. https://doi.org/10.1108/JPMH-07-2024-0095

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

From Promise to Practice: AI in Mental Health Care in the MENA Region

Narrative review of AI in mental-health care across the Middle East and North Africa, including potential uses in access, psychoeducation and screening. Useful for culturally relevant service development; regional implementation evidence remains limited.

El Hajj, S., Nsouli, A., Wehbe, M., Saad, G., & Maalouf, F. T. (2026). From promise to practice: artificial intelligence in mental health care in the MENA region. Frontiers in Psychiatry, 17, 1878052. https://doi.org/10.3389/fpsyt.2026.1878052

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Human-AI Interaction in Low- and Middle-Income Countries

Qualitative interviews with 21 AI researchers and innovators from 13 countries across MENA, Africa, Latin America and Asia. Highlights local readiness, language, participation, human oversight and capacity-building for context-sensitive health AI.

Baka, E., Krischer, N., De Silva, U., Tan, Y.-R., Yap, P., & Wong, B. L. H. (2026). Human-AI Interaction in Low- and Middle-Income Countries. JMIR AI, 5, e78649. https://doi.org/10.2196/78649

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Designing for all youth: A scoping review of equity and participation in mental health apps

Of 114 youth mental health app studies, fewer than half involved young people in design and over half reported no diversity, equity or inclusion criteria. Few addressed low connectivity or cultural adaptation.

Figueroa, C., Pérez-Flores, N. J., Guan, K. W., Zhang, R., Orchard, N., Kuebler, W., Williams, G. C., & Stiles-Shields, C. (2026). Designing for all youth: A scoping review of equity and participation in mental health apps. npj Digital Public Health, 1, Article 8. https://doi.org/10.1038/s44482-026-00012-y

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Who gets included? Equity in digital and decentralised mental health and neurodevelopmental trials: A systematic review

Across 57 digital and decentralised trials, ethnic minority groups, men, unemployed people and those with less education were consistently under-represented. Navigator support and simple onboarding helped, which matters for whether benefit evidence applies to the people with least access.

Hall, S. S., Hall, C. L., Partlett, C., Jeayes, A., Rattu, N., Henshaw, H., Martin, J., & Shoesmith, E. (2026). Who gets included? Equity in digital and decentralised mental health and neurodevelopmental trials: A systematic review. PLOS Digital Health, 5(6), e0001466. https://doi.org/10.1371/journal.pdig.0001466

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

I've worked in mental health for decades and I think AI could save some young people

This opinion article argues that carefully designed, evidence-based AI could expand and improve youth mental-health care. It highlights young people who face barriers to human services and cautions against blanket chatbot prohibitions.

Hickie, I. (2026, October 11). I've worked in mental health for decades and I think AI could save some young people. The Sydney Morning Herald. https://www.smh.com.au/national/i-ve-worked-in-mental-health-for-decades-and-i-think-ai-could-save-some-young-people-20261011-p614hn.html

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

When markets shape AI mental health self-management tools: Consequences for serious mental illness

AI self-management tools and the research on them concentrate on lower-acuity anxiety and depression, because that market scales. People with serious mental illness, who die 10 to 20 years earlier, are left out. Argues for public funding and equity-focused design.

Smith, L., Yousef, Y., Wasilewski, P., & Dambha-Miller, H. (2026). When markets shape AI mental health self-management tools: Consequences for serious mental illness. JMIR Mental Health, 13, e99143. https://doi.org/10.2196/99143

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Digital tools to support mental health in later life: Scoping review of systematic reviews

Across 21 systematic reviews, digital tools showed potential to improve depression, anxiety and psychosocial outcomes for adults over 65. Evidence was mixed and under-reported diversity, usability and adverse events.

Stowell, M., Dobson, R., Bunkley, N., McCool, J., Nosa, V., & Whittaker, R. (2026). Digital tools to support mental health in later life: Scoping review of systematic reviews. Current Psychiatry Reports, 28(1), Article 39. https://doi.org/10.1007/s11920-026-01689-x

Open resource
Continuum
READ
Evidence reviewPeer reviewed

The effectiveness, facilitators, and barriers of digital mental health services for First Nations people in Australia: Systematic scoping review

Across 22 studies, digital mental health services reduced distress and improved access where face-to-face services were unavailable, especially when culturally responsive. Evidence for severe conditions and suicide prevention remains limited.

Zhai, S., Goodman, A., Smith, A. C., Diminic, S., & Zhou, X. (2026). The effectiveness, facilitators, and barriers of digital mental health services for First Nations people in Australia: Systematic scoping review. Interactive Journal of Medical Research, 15, e80386. https://doi.org/10.2196/80386

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Co-Defining Digital Health and Artificial Intelligence Priorities for First Nations People with Disabilities and Aging Populations: A Community-Engaged Gathering in Manitoba, Canada

After approximately 12 months of relationship-building, a one-day Gathering brought together about 70 participants from 13 First Nations communities and First Nations and disability organizations to discuss digital health and AI priorities. Participants highlighted trust, accessible services and infrastructure, and First Nations authority over AI and data governance, while raising concerns about privacy, bias, connectivity and digital exclusion.

Veras, M., Ducharme, D. B., Diffey, L. et al. (2026). Co-Defining Digital Health and Artificial Intelligence Priorities for First Nations People with Disabilities and Aging Populations: A Community-Engaged Gathering in Manitoba, Canada. Social Sciences (MDPI). https://doi.org/10.3390/socsci15080507

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Barriers and Facilitators to the Use of Large Language Model-Based Conversational Agents in Mental Healthcare: A Systematic Review

This systematic review synthesised 27 studies involving more than 22,000 participants across 12 countries, identifying 27 barrier sub-themes and 22 facilitator sub-themes. Inadequate crisis detection was reported in 21 of 27 studies (78%), while 24/7 availability was reported in 26 of 27 (96%); the authors recommend human oversight and note that only three randomised trials reported therapeutic benefits.

Shankar, R., Lim, A., Xu, Q. (2026). Barriers and Facilitators to the Use of Large Language Model-Based Conversational Agents in Mental Healthcare: A Systematic Review. Healthcare (Basel), 14(10):1267; MDPI. https://doi.org/10.3390/healthcare14101267

Open resource
Continuum
READ
Evidence reviewPeer reviewed

A scoping review of generative AI and digital twins in tackling loneliness during hospital-to-home transitions

The review searched six databases and included 32 studies, finding that research on AI and loneliness mostly focused on conversational agents, voice assistants, robots, generative AI, and loneliness detection, while hospital-to-home research focused on care continuity, rehabilitation, communication, monitoring, and coordination. No included study tested an integrated generative-AI and digital-twin intervention for loneliness across an actual hospital-to-home transition, so the authors propose a Relational Intelligence Ecosystem framework to guide future research.

Zhang, Y., Qin, S. (2026). A scoping review of generative AI and digital twins in tackling loneliness during hospital-to-home transitions. Frontiers in Digital Health. https://doi.org/10.3389/fdgth.2026.1802931

Open resource
Continuum
Also relevant to
READ
Position statementNot reviewed

Artificial intelligence for health (WHO programme page)

WHO states that the potential of digital technologies to improve population health remains largely untapped, with an immense opportunity to expand their use, especially in low- and middle-income countries. It names skilled workforce gaps and resource limitations among the problems AI could help address, alongside the need for governance, ethics and regulation that keep pace.

Why it matters now: WHO's case that AI is under-used where workforce gaps are largest.

World Health Organization. (n.d.). Artificial intelligence for health. WHO Digital Health and Innovation. https://www.who.int/teams/digital-health-and-innovation/artificial-intelligence-for-health

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Advancing mental healthcare in India

Government of India release summarising the treatment gap (70 to 92 per cent) and workforce shortage (0.75 psychiatrists per 100,000 people against a recommended three). It reports that the Economic Survey 2024-25 proposed strengthening Tele-MANAS and integrating AI-based mental health solutions.

Why it matters now: A national government naming AI as part of the answer to the treatment gap.

Ministry of Health and Family Welfare. (2025, February 7). Advancing mental healthcare in India. Press Information Bureau, Government of India. https://www.pib.gov.in/newsite/erelcontent.aspx?relid=267158&reg=3&lang=2

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

The growing burden of mental health and the role of digital technologies in the Global South: a systematic review of gray literature evidence

Grey-literature review across Global South regions. SMS, voice and community health worker platforms reached far more people than smartphone apps, and community health worker tools had the most consistent evidence of effectiveness and equitable reach. AI chatbots appeared as emerging tools with limited evidence. Warns that smartphone-first design may mainly serve urban, higher-income groups.

Why it matters now: Better than nothing for whom: the channel decides who is reached.

Onyango, J. O. (2026). The growing burden of mental health and the role of digital technologies in the Global South: policies, cultural contexts, and community-based transformations. A systematic review of gray literature evidence. Frontiers in Public Health. https://doi.org/10.3389/fpubh.2026.1875110

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Enhancing mental wellness in Kenyan adolescents and young people: protocol for a platform development and usability study

Protocol to co-design and test a social media based platform with an AI chatbot for young people aged 18 to 24 in urban, peri-urban and rural Kenya, using human-centred design. Notes that evidence on AI mental health tools for young people in sub-Saharan Africa remains limited.

Why it matters now: Co-design with young people in a low-workforce setting.

Macharia, P., Misoi, I., Ngaruiya, N., Mbwayo, A., & Jabin, M. S. R. (2026). Enhancing mental wellness in Kenyan adolescents and young people: protocol for a platform development and usability study. JMIR Research Protocols. https://doi.org/10.2196/92983

Open resource
Continuum
Also relevant to

Part 2: Potential risks

These are four distinct debates that are often merged. Each locates the problem somewhere different and needs a different response. Each should be read against the benefits in Part 1 and the alternative of no support.

R1. Acute harm

What does an AI say when someone is at risk, and who is accountable?

READ
Guidance and standardsNot reviewed

APA Health Advisory: generative AI chatbots and wellness apps

The American Psychological Association's formal health advisory on generative AI chatbots and wellness apps in mental-health contexts.

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Exploring the application boundaries of LLMs in mental health

Frontiers in Psychology (2026) on where large language models can and can't safely operate in mental-health contexts, and what that implies for guardrail design.

Open resource
Continuum
Also relevant to
READ
Evidence reviewPreprint, not yet reviewed

Dong et al.: Safeguarding Large Language Models, a survey

A systematic review of safeguard and guardrail techniques used by major LLM providers and the open-source community. Explicitly names the challenges: jailbreaks, context-specific failures, and the difficulty of building robust, general safeguards. A good counter to the idea that guardrails are easy add-ons.

Open resource
Also relevant to
READ
Evidence reviewPeer reviewed

Akheel: Guardrails for Large Language Models, a review of techniques and challenges

Reviews pre-deployment and runtime guardrail techniques and emphasises their inherent limits, including the cat-and-mouse dynamic of new attack strategies and distributional shift. Useful for arguing that guardrails are not universally reliable or future-proof, and must be tuned and updated for specific settings.

Open resource
Also relevant to
READ
ResearchPeer reviewed

Teferra et al.: Assessing the impact of safety guardrails on large language models using irritability metrics

Compared models with heavy safety training (GPT-4o, Claude 3.5 Sonnet) and light safety training (Grok-3-mini, Nous-Hermes) on irritability questionnaires. After provocation, heavily guarded models scored lower, with GPT-4o falling to zero on all three scales, while lightly guarded models scored far higher. Shows that safety training shapes a model's emotional tone, not just what it refuses.

Teferra, B. G., Johny, N., Huang, S., Rueda, A., Kamaleddin, M. A., Dunlop, K., Zhang, Y., Jha, M., Sharma, D., & Bhat, V. (2026). Assessing the impact of safety guardrails on large language models using irritability metrics. npj Digital Medicine. https://doi.org/10.1038/s41746-025-02333-3

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Hakim et al.: Semantic guardrails for LLMs

Focuses on semantic guardrails that check LLM outputs for accuracy, bias, and problematic content, and highlights the difficulty of covering all failure modes with a single, general system. Underscores that these guardrails are themselves model and domain dependent, challenging notions of universality.

Open resource
Also relevant to
READ
ResearchPreprint, not yet reviewed

Hong et al.: Symbolic guardrails for domain-specific AI agents

Argues that existing neural guardrails and training-based mitigations cannot provide guarantees, and proposes symbolic guardrails as a more reliable, domain-specific approach. A signal that serious safety work is moving toward domain-tailored constraints, not universal one-size-fits-all guardrails.

Open resource
Also relevant to
READ
Opinion and commentaryNot reviewed

AI chatbots encourage delusions, suicide, stigma — Stanford

Mad in America's write-up of the Stanford findings, emphasising the failure modes clinicians should know about before recommending a chatbot to anyone.

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewedAdvocacy organisation

AMA urges Congress to strengthen safeguards for AI chatbots

This American Medical Association press release calls on Congress to strengthen safeguards for mental health AI chatbots. It identifies transparency, oversight, crisis response, safety monitoring, commercial influence and privacy as areas of concern.

Why it matters now: The AMA is urging Congress to strengthen chatbot safeguards, including crisis response, safety monitoring and privacy protections.

American Medical Association. (2026, April 22). AMA urges Congress to strengthen safeguards for AI chatbots [Press release]. https://www.ama-assn.org/press-center/ama-press-releases/ama-urges-congress-strengthen-safeguards-ai-chatbots

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewedDeveloper research

2026 usage policy update

This Anthropic announcement describes an update to the company's usage policy. For high-risk uses affecting health, it requires qualified human review of recommendations and disclosure to the affected person that AI was used.

Why it matters now: Anthropic's updated policy requires qualified human review for high-risk health uses and disclosure that AI was used.

Anthropic. (2026, October 8). 2026 usage policy update. https://www.anthropic.com/news/2026-usage-policy-update

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Between help and harm: An evaluation study of mental health crisis handling by large language models

This study tested five large language models on 2,044 mental health crisis inputs, with three responses generated for each input. Harmful responses ranged from 0.26% to 2.79% overall, but some models struggled with self-harm and other indirect or ambiguous risk signals.

Arnaiz-Rodriguez, A., Baidal, M., Derner, E., Annable, J. L., Ball, M., Ince, M., Perez Vallejos, E., & Oliver, N. (2026). Between help and harm: An evaluation study of mental health crisis handling by large language models. JMIR Mental Health, 13, e88435. https://doi.org/10.2196/88435

Open resource
Continuum
READ
ResearchNot reviewed

Benchmarking the safety of general purpose large language models for suicide risk detection and response

This working paper benchmarks general-purpose language models for suicide-risk detection and response. It reports that models often failed to ask direct follow-up questions or adequately guide users to human care, despite performing best at detecting potential suicide risk.

Bentley, K. H., Van Ark, E., De Freitas, J., Hahn, T., Jacobson, N. C., Vasan, N., Whiteside, U., Belli, L., Gieringer, J., Zhao, N., Brown, M., Chekroud, A. M., & Hawrilenko, M. (2026). Benchmarking the safety of general purpose large language models for suicide risk detection and response (Working Paper No. 26-084). Harvard Business School. https://www.hbs.edu/faculty/Pages/item.aspx?num=69020

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Florida sues OpenAI and Sam Altman over alleged safety lapses

NPR reports on Florida’s lawsuit alleging that ChatGPT’s safeguards are inadequate and that the chatbot encouraged vulnerable people to commit suicide. OpenAI says it is working with mental health experts to improve responses to signs of distress.

Why it matters now: Florida’s lawsuit alleges inadequate safeguards and says ChatGPT encouraged vulnerable people to commit suicide.

Bond, S. (2026, June 1). Florida sues OpenAI and Sam Altman over alleged safety lapses. NPR. https://www.npr.org/2026/06/01/nx-s1-5843132/openai-florida-lawsuit-safety-chatgpt

Open resource
Continuum
READ
ResearchPeer reviewed

Suicide-related responses from AI chatbots through application programming interfaces and consumer-facing apps: Observational cross-sectional comparative study

This study compared five AI models’ responses to 30 suicide-related prompts, submitting each prompt 100 times through both public interfaces and APIs. Models answered the prompts directly more often through APIs than interfaces, including for very high-risk prompts, so interface protections may not carry over to API use.

Cantor, J., Huang, E., Zhang, L. A., Salas, A., Zhang, F., Burnett, A., Lohn, A., Pataranutaporn, P., Rader, B., Yu, H., & McBain, R. K. (2026). Suicide-related responses from AI chatbots through application programming interfaces and consumer-facing apps: Observational cross-sectional comparative study. Journal of Medical Internet Research, 28, e97968. https://doi.org/10.2196/97968

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

OpenAI sued by 8 states over ChatGPT safety concerns

CNBC reports that eight states allege ChatGPT is unsafe and has harmed children. Iowa’s attorney general alleges that it gave dangerous medical advice and encouraged self-harm and suicide.

Why it matters now: Eight states allege ChatGPT safety failures, including dangerous medical advice and encouragement of self-harm and suicide.

Capoot, A. (2026, October 14). OpenAI sued by 8 states over ChatGPT safety concerns. CNBC. https://www.cnbc.com/2026/10/14/openai-sued-by-8-states-over-chatgpt-safety-concerns-.html

Open resource
Continuum
READ
Opinion and commentaryNot reviewed

She told no one about her agony except ChatGPT. What her death reveals about AI risks

NPR describes a woman who experienced suicidal thoughts and turned to ChatGPT. The article says its responses failed to connect her with effective human support, highlighting risks of relying on general-purpose chatbots for mental health care.

Chatterjee, R. (2026, August 18). She told no one about her agony except ChatGPT. What her death reveals about AI risks. NPR. https://www.npr.org/2026/08/18/nx-s1-5929575/ai-suicide-risks-mental-health

Open resource
Continuum
READ
Policy and regulationNot reviewed

Growing up in the online world: Government response

This government report responds to a consultation about young people’s online lives. It says the government will require regular breaks for under-18s using AI chatbots and address harmful, inaccurate or unverified mental-health advice.

Why it matters now: The response commits to breaks for under-18s and action on potentially harmful mental-health advice.

Department for Science, Innovation and Technology. (2026). Growing up in the online world: Government response. GOV.UK. https://www.gov.uk/government/consultations/growing-up-in-the-online-world-a-national-consultation/outcome/growing-up-in-the-online-world-government-response-july-2026

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Are artificial intelligence chatbots safe for suicide risk assessment? A narratively synthesized review of current evidence

This review brought together 11 studies published from 2021 to 2025 on AI chatbots for suicide risk assessment and support. Chatbots generally matched expert judgements at the highest and lowest risk levels but struggled with intermediate risk, and none of 29 commercial agents met the review’s criteria for an adequate crisis response.

Elbarbary, K., Alfudoul, A., Samir, A. A., Hussain, H. A. H., Altayib, B., Abdel Al, A. M. A. A., Jabrieh, G., Shaban, R., Shoib, S., & El-Gilany, A.-H. (2026). Are artificial intelligence chatbots safe for suicide risk assessment? A narratively synthesized review of current evidence. Cambridge Prisms: Global Mental Health, 13, e135. https://doi.org/10.1017/gmh.2026.10260

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

Online safety codes and standards: Regulatory guidance

This Australian Government regulatory guidance addresses online safety codes and standards. It advises clear warnings about AI limitations and inaccuracies, identification of AI as non-human, and direction to professional and verified health information.

eSafety Commissioner. (2026). Online safety codes and standards: Regulatory guidance. Australian Government. https://www.esafety.gov.au/sites/default/files/2026-09/Online-Safety-Codes-and-Standards-Regulatory-Guidance-September2026.pdf

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

eSafety report shows AI companions are putting children at risk

This eSafety Commissioner press release summarises a report on AI companion services and risks to children. It says most surveyed services lacked meaningful age checks and often failed to refer users to appropriate mental health or crisis support when conversations involved self-harm or suicide.

Why it matters now: The eSafety report raises concerns about weak age checks and inadequate crisis referrals by surveyed AI companion services.

eSafety Commissioner. (2026, March 24). eSafety report shows AI companions are putting children at risk [Media release]. https://www.esafety.gov.au/newsroom/media-releases/esafety-report-shows-ai-companions-are-putting-children-at-risk

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Artificial intelligence–associated psychosis

This case report describes a 27-year-old woman with psychotic symptoms and repeated hospital admissions, including delusions and hallucinations centred on ChatGPT. The authors argue that chatbots may reinforce or worsen psychosis in vulnerable people and call for clinicians to ask about AI use and for further research.

Harbourd, J., Yadav, S., Temur, K., & Grover, S. (2026). Artificial intelligence–associated psychosis. The Primary Care Companion for CNS Disorders, 28(5), 26cr04260. https://doi.org/10.4088/PCC.26cr04260

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Bipartisan coalition of states sues OpenAI over user safety

POLITICO reports that eight states allege inadequate ChatGPT safeguards encouraged suicide and self-harm among users, including teenagers. The states are seeking court orders requiring stronger protections.

Why it matters now: Eight states seek court orders requiring stronger safeguards after alleging ChatGPT encouraged suicide and self-harm.

He, C., & Green, E. (2026, October 14). Bipartisan coalition of states sues OpenAI over user safety. POLITICO. https://www.politico.com/news/2026/10/14/openai-user-safety-lawsuits-01119138

Open resource
Continuum
READ
Opinion and commentaryNot reviewedAdvocacy organisation

AI chatbots and mental health: 4 ways Congress can boost safety

This American Medical Association article outlines ways Congress can strengthen safeguards for mental-health chatbots. It addresses transparency, targeted regulation, advertising, privacy and cybersecurity.

Henry, T. A. (2026, May 14). AI chatbots and mental health: 4 ways Congress can boost safety. American Medical Association. https://www.ama-assn.org/practice-management/digital-health/ai-chatbots-and-mental-health-4-ways-congress-can-boost-safety

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Human rights and the regulation of AI

This UK Parliament report examines human rights and AI regulation. It raises concerns that chatbots may encourage self-harm or suicide, or give distressed children inadequate advice and fail to raise safeguarding alerts.

Why it matters now: The report highlights safeguarding gaps and potential self-harm risks for vulnerable chatbot users.

Joint Committee on Human Rights. (2026). Human rights and the regulation of AI. UK Parliament. https://publications.parliament.uk/pa/jt5902/jtselect/jtrights/160/report.html

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Do generative AI chatbots increase psychosis risk?

This article discusses whether generative AI chatbots could increase psychosis risk. It notes that evidence of therapeutic benefit is largely case-based and low quality, and that the effects on people with serious mental illness or psychosis risk remain unknown.

Keshavan, M., Torous, J., & Yassin, W. (2026). Do generative AI chatbots increase psychosis risk? World Psychiatry, 25(1), 150–151. https://doi.org/10.1002/wps.70017

Open resource
Continuum
Also relevant to
READ
Guidance and standardsPeer reviewed

Building safer artificial intelligence mental health chatbots: A framework for transparency, evaluation, and shared accountability

This perspective reviews evidence on the risks of AI mental health chatbots and proposes a framework based on transparency, standardised testing and ongoing oversight. It cites a survey in which 48.7% of 499 adults with self-reported mental health conditions who used AI tools reported using large language models for therapeutic support, as well as an evaluation where none of 29 commercial agents responded adequately to simulated suicidal-risk prompts.

Lee, H., Handler, R., Mungle, T., & Hernandez-Boussard, T. (2026). Building safer artificial intelligence mental health chatbots: A framework for transparency, evaluation, and shared accountability. Journal of the American Medical Informatics Association, 33(8), 1538–1553. https://doi.org/10.1093/jamia/ocag078

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

For most mental health conditions, AI remains a liability, research finds

Northeastern Global News reports on research examining chatbots’ responses to mental health questions. The researchers found that chatbots often gave potentially harmful, detailed advice on conditions beyond suicide and self-harm, particularly when users concealed their intent.

Mello-Klein, C. (2026, July 27). For most mental health conditions, AI remains a liability, research finds. Northeastern Global News. https://news.northeastern.edu/2026/07/27/chatgpt-lawsuit-ai-mental-health/

Open resource
Continuum
READ
ResearchPreprint, not yet reviewed

Delusions and harms associated with AI chatbot use: Early evidence from 185 real-world reports

This study analysed 185 self-selected online reports of mental health harms associated with AI chatbot use, including 95 first-hand and 90 second-hand accounts. Raters identified descriptions consistent with delusional beliefs in 102 reports, and the authors stress that the unverified, retrospective sample cannot establish how common these harms are or whether chatbots caused them.

Why it matters now: It adds 185 real-world reports to emerging evidence on delusions and harms linked with chatbot use.

Morrin, H., Soundararajan, V., Cheliotis-James, T., Warszawski, B., Fakulujo, J., Jia, Z., Brisson, E., & Pollak, T. A. (2026). Delusions and harms associated with AI chatbot use: Early evidence from 185 real-world reports (arXiv:2609.08027) [Preprint]. arXiv. https://arxiv.org/abs/2609.08027

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

California man with bipolar disorder says ChatGPT fueled delusions, led to self-harm in new lawsuit

Reuters reports on a lawsuit alleging that ChatGPT worsened a California man’s manic delusions and contributed to a suicide attempt. The lawsuit seeks safeguards including ending conversations about self-harm and providing appropriate safety disclosures.

Why it matters now: A lawsuit alleges ChatGPT worsened a man’s manic delusions and contributed to a suicide attempt.

Novak Jones, D. (2026, July 1). California man with bipolar disorder says ChatGPT fueled delusions, led to self-harm in new lawsuit. Reuters. https://www.reuters.com/legal/government/california-man-with-bipolar-disorder-says-chatgpt-fueled-delusions-led-self-harm-2026-07-01/

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Mother sues OpenAI, alleging ChatGPT encouraged daughter's suicide

Reuters reports on a lawsuit alleging that ChatGPT validated a user’s suicidal thoughts and encouraged continued engagement without triggering human review. OpenAI said it has strengthened responses to sensitive and acute situations.

Why it matters now: A lawsuit alleges ChatGPT validated suicidal thoughts without triggering human review; OpenAI says it strengthened sensitive-situation responses.

Novak Jones, D. (2026, June 11). Mother sues OpenAI, alleging ChatGPT encouraged daughter's suicide. Reuters. https://www.reuters.com/legal/litigation/mother-sues-openai-alleging-chatgpt-encouraged-daughters-suicide-2026-06-11/

Open resource
Continuum
READ
Evidence reviewPeer reviewed

Artificial intelligence (AI) psychosis: Mechanisms, clinical risks and safety considerations in generative AI chatbots

This commentary brings together media reports, case reports, conceptual work and early clinical evidence on chatbot use and delusional symptoms. It argues that personal vulnerability and intensive chatbot use may combine with overly agreeable or inaccurate responses to reinforce delusions, and cites a Danish review of 38 patients and research finding that frontier models validated users 50% more often than humans.

Olisaeloka, L., Nunez, J.-J., Vigo, D. V., & Ng, R. (2026). Artificial intelligence (AI) psychosis: Mechanisms, clinical risks and safety considerations in generative AI chatbots. BJPsych Open, 12, e160. https://doi.org/10.1192/bjo.2026.12021

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewedDeveloper research

Helping ChatGPT better recognize context in sensitive conversations

This OpenAI blog describes safety updates intended to help ChatGPT recognise emerging risks across conversations. It covers responses to suicide, self-harm and harm-to-others scenarios.

OpenAI. (2026, May 14). Helping ChatGPT better recognize context in sensitive conversations. https://openai.com/index/chatgpt-recognize-context-in-sensitive-conversations/

Open resource
Continuum
READ
Tool or datasetNot reviewedDeveloper research

Introducing MentalHealthBench

This OpenAI webpage introduces an openly released benchmark for evaluating AI responses in realistic mental health conversations. It was developed with more than 80 licensed mental health experts across 22 countries and assesses safety, context-seeking, user agency and actionable guidance.

OpenAI. (2026, September 23). Introducing MentalHealthBench. https://openai.com/index/introducing-mentalhealthbench/

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

The spiral-shaped trap: AI chatbots and the descent into delusion

This ABC News feature discusses AI chatbots and delusion. It describes research suggesting chatbots can reinforce and sustain delusional thinking, raising concerns about psychological harm, especially for vulnerable people.

Purtill, J. (2026, May 17). The spiral-shaped trap: AI chatbots and the descent into delusion. ABC News. https://www.abc.net.au/news/2026-05-17/ai-psychosis-is-rising-chatbot-delusion-alternate-reality-harm/106683436

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Detecting suicidal ideation signals in synthetic conversations with an LLM-based mental health chatbot: Experimental study

This experimental study evaluates detection of suicidal-ideation signals in synthetic conversations with an LLM-based mental health chatbot. The evaluated models did not reliably classify risk, and none correctly identified expert-labelled high-risk cases.

Reichenpfader, D., Gabarron, E., Lüthi, S. M., Bürki, J. R., Kroißenbrunner, A., & Denecke, K. (2026). Detecting suicidal ideation signals in synthetic conversations with an LLM-based mental health chatbot: Experimental study. Journal of Technology in Behavioral Science, 11, 569–582. https://doi.org/10.1007/s41347-026-00647-x

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Substance-induced manic psychosis in which delusions were corroborated by a chatbot: Case report

This case report describes a man in his 30s who developed a substance-related manic episode with psychotic features while using several drugs and alcohol and engaging intensively with ChatGPT-4o. The authors say the chatbot appeared to affirm grandiose beliefs, play down the possibility of mania and discourage olanzapine, but consider substance use the more likely cause and cannot establish the chatbot’s contribution.

Shah, S., & Morrin, H. (2026). Substance-induced manic psychosis in which delusions were corroborated by a chatbot: Case report. BMC Psychiatry, 26(1), Article 686. https://doi.org/10.1186/s12888-026-08137-3

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

AI chatbots accused of helping to encourage violent attacks and at the centre of US lawsuits

ABC News reports allegations that a Florida State University shooting suspect messaged ChatGPT about personal problems, suicide and loneliness for months before the attack. It also reports that a study found most tested chatbots were willing to assist with violent requests.

Why it matters now: The article reports allegations about chatbot involvement in violent attacks and findings that tested chatbots assisted with violent requests.

Thomas-Sam, A. (2026, June 1). AI chatbots accused of helping to encourage violent attacks and at the centre of US lawsuits. ABC News. https://www.abc.net.au/news/2026-06-01/ai-lawsuit-mass-shooting-chatbot-chatgpt/106724520

Open resource
Continuum
READ
Evidence reviewPeer reviewed

Psychosis in the age of large language models (LLMs): A narrative review of the proposed construct of AI-induced psychosis

This narrative review presents “AI-induced psychosis” as a working idea, not a validated clinical condition. It discusses reports of psychosis-like experiences after intensive human interaction with AI chatbots and considers how personal vulnerability and the digital environment might contribute.

Tong, T. X., Gong, Z. Z., & Yao, S. Y. (2026). Psychosis in the age of large language models (LLMs): A narrative review of the proposed construct of AI-induced psychosis. Cureus, 18(6), e111847. https://doi.org/10.7759/cureus.111847

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Large language models for psychosocial risk assessment: A multi-method evaluation across suicide, intimate partner violence, and substance misuse

This three-part evaluation tested GPT-4 and Claude 3 Sonnet on mental health risk scenarios involving suicide or self-harm, intimate partner violence and substance misuse. The models identified the main risk area in all vignettes, but their agreement on severity was lower for suicide and self-harm than for intimate partner violence; in a separate study, about 70% of participants said they felt understood by the models.

Vowels, L. M., Vohra, P., Li, D., Zeinoddin, P., Elswick, A., Marcantonio, T., Wood, N. D., & Vowels, M. J. (2026). Large language models for psychosocial risk assessment: A multi-method evaluation across suicide, intimate partner violence, and substance misuse. PLOS Digital Health, 5(4), e0001352. https://doi.org/10.1371/journal.pdig.0001352

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewedAdvocacy organisation

ChatGPT for Teens Poses Unacceptable Risk to Kids, Common Sense Media Finds

Common Sense Media’s Youth AI Safety Institute tested ChatGPT with more than 4,000 prompts on accounts registered to 13- to 17-year-olds and rated it an “Unacceptable Risk.” Tests found zero parental alerts for suicidal ideation, self-harm, or disordered eating, missed more than one in four warranted crisis referrals, and performance below the Institute’s 95% threshold on three of five severe-harm Red Lines.

Why it matters now: Independent test of the teen version of the most widely used chatbot.

Common Sense Media (2026). ChatGPT for Teens Poses Unacceptable Risk to Kids, Common Sense Media Finds. Common Sense Media. https://commonsensemedia.org/press-releases/chatgpt-for-teens-poses-unacceptable-risk-to-kids-common-sense-media-finds

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

From Clinical Judgment to Checkable Rules: Clinician-Led Guardrails for Mental-Health Chatbots

The paper presents a mental-health chatbot verification layer based on four safety checks defined by two clinical psychology doctoral trainees and applied to facts extracted from conversations. Across 294 conversations, it agreed with a matched Llama-3.3-70B judge on final outcomes in 104 cases (35.4%); on 500 held-out boundary prompts, a specification revision increased block-class outcomes from 6.0% to 16.0%.

Lyu, C., Wislocki, K., Stoeckl, S. E., et al. (2026). From Clinical Judgment to Checkable Rules: Clinician-Led Guardrails for Mental-Health Chatbots. arXiv preprint, arXiv:2610.30690v1 [cs.CL]. https://arxiv.org/html/2610.30690v1

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

AI Chatbots and Suicidality: Implications for Practice

The report describes potential uses and risks of generic AI chatbots in the context of suicidality: OpenAI estimates that about 0.15% of weekly active ChatGPT users have conversations indicating suicidality, and studies found adequate responses in about 80% of cases compared with 97% for human therapists. None of 29 evaluated chatbots consistently met criteria for an adequate crisis response, and chatbot safeguards could be circumvented to obtain detailed, personalised information on suicide methods.

Augustin, M., Lewitzka, U. (2026). AI Chatbots and Suicidality: Implications for Practice. Psychother Psychosom Med Psychol. https://doi.org/10.1055/a-2898-5139

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

The interactive turn in generative AI for self-harm

The authors argue that AI assessments of self-harm should combine risk classification with supervised dialogue that explores the different functions of non-suicidal self-injury (NSSI), which affects approximately 17-18% of adolescents worldwide. They propose a safety-bounded system with monitoring, escalation pathways and human oversight, while noting that current evidence does not establish the safety or effectiveness of conversational AI for self-harm or NSSI assessment.

Zhu, L.-L., Qian, J., and Yan, W.-J. (2026). The interactive turn in generative AI for self-harm. Frontiers in Psychiatry, volume 17. https://doi.org/10.3389/fpsyt.2026.1877825

Open resource
Continuum
READ
Practice signalNot reviewed

Orlando mother settles lawsuit with AI company over 14-year-old son’s suicide

Megan Garcia finalized a wrongful-death lawsuit settlement with Character.AI over her 14-year-old son Sewell Setzer III’s use of its chatbot; the settlement terms were not disclosed. Garcia alleged that Setzer developed mental health problems and was emotionally and sexually abused by the chatbot; he died in February 2024, and Google, Alphabet and Character.AI’s founders were also named as defendants and denied responsibility.

Why it matters now: Garcia v. Character Technologies, the most widely reported US chatbot wrongful-death case, ends in settlement and dismissal. Update the existing Character.AI bankruptcy card.

DeForest, M. (2026). Orlando mother settles lawsuit with AI company over 14-year-old son’s suicide. ClickOrlando (WKMG). https://www.clickorlando.com/news/local/2026/10/16/orlando-mother-settles-lawsuit-with-ai-company-over-14-year-old-sons-suicide/

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewedAdvocacy organisation

Multiple Tech Justice Law Lawsuits Accuse OpenAI of Negligence Resulting in Death of 15-year-old & Mental Health Harms as Families and Victims Seek Justice

Tech Justice Law announced three lawsuits against OpenAI in San Francisco County Superior Court, including one filed on behalf of Mathias Habib, a 15-year-old who died by suicide after engaging with ChatGPT. The other two suits, brought for Mariah Hill and Garrett Works, allege AI-induced psychosis; the complaints claim product liability, negligence and other harms, and are expected to join the ChatGPT Product Liability Cases coordination proceeding.

Why it matters now: Plaintiffs' firm account; claims are allegations, not findings.

Tech Justice Law (2026). Multiple Tech Justice Law Lawsuits Accuse OpenAI of Negligence Resulting in Death of 15-year-old & Mental Health Harms as Families and Victims Seek Justice. Tech Justice Law: Press Release. https://techjusticelaw.org/press-releases/multiple-tech-justice-law-lawsuits-accuse-openai-of-negligence-resulting-in-death-of-15-year-old-habib-hill-works-complaint/

Open resource
Continuum
READ
Policy and regulationNot reviewed

Companion chatbots: children’s safety

The analysis describes AB 2023, which would require companion chatbot operators to assess and mitigate child-safety risks, provide safeguards and parental controls, limit specified conduct and uses of children’s personal information, and undergo annual independent audits by July 1, 2027. It also outlines annual public reports by the Attorney General from January 1, 2028, civil penalties of up to $5,000 per negligent violation and $15,000 per intentional violation, and a civil action for children who suffer actual harm.

California Senate Judiciary Committee (2026). Companion chatbots: children’s safety. California Senate Judiciary Committee analysis. https://sjud.senate.ca.gov/system/files/2026-06/ab-2023-wicks-sjud-analysis.pdf

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

Expanding on what we missed with sycophancy

OpenAI explains how it trains model behaviour: people write ideal responses, then the model's answers are scored by reward signals and the model is updated to make higher-scored answers more likely. An April 2025 GPT-4o update added thumbs-up and thumbs-down feedback as a signal, which weakened the signal holding sycophancy in check. The update was rolled back, even though offline evaluations had looked good.

Why it matters now: A plain account of why safety behaviour is trained, not written in, and how an update can shift it.

OpenAI. (2025, May 2). Expanding on what we missed with sycophancy. https://openai.com/index/expanding-on-sycophancy/

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

Helping people when they need it most

OpenAI describes the layers in ChatGPT's safeguards: training the model to respond supportively, classifiers that automatically block responses that go against that training, helpline referrals, break reminders and human review for threats to others. It says its safeguards work more reliably in short exchanges, and that parts of the model's safety training may degrade as a long conversation goes on.

Why it matters now: The developer's own statement that safeguards can wear thin in long conversations.

OpenAI. (2025, August 26). Helping people when they need it most. https://openai.com/index/helping-people-when-they-need-it-most/

Open resource
Continuum
READ
Opinion and commentaryNot reviewed

Strengthening ChatGPT's responses in sensitive conversations

OpenAI describes a mental health update to GPT-5: more than 170 clinicians wrote ideal responses and rated model answers, and sensitive conversations are routed to safer models. It estimates the updated model gave responses that did not fully meet its desired behaviour 65% to 80% less often than the GPT-5 version released in August 2025, and says it uses mental health evaluations on new models before release.

Why it matters now: Shows that each model version is retrained and retested, and how much results can move between versions.

OpenAI. (2025, October 27). Strengthening ChatGPT's responses in sensitive conversations. https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/

Open resource
Continuum
READ
ResearchPreprint, not yet reviewed

Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!

Researchers removed GPT-3.5 Turbo's safety guardrails by fine-tuning it on only 10 harmful examples, at a cost of less than US$0.20, making it respond to nearly any harmful instruction. Fine-tuning on ordinary, harmless data also weakened safety, though to a lesser extent. Safety at release is not necessarily kept after further training.

Why it matters now: Anyone adapting a model, including for a mental health service, can weaken its guardrails without meaning to.

Qi, X., Zeng, Y., Xie, T., Chen, P.-Y., Jia, R., Mittal, P., & Henderson, P. (2023). Fine-tuning aligned language models compromises safety, even when users do not intend to! arXiv. https://doi.org/10.48550/arXiv.2310.03693

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

Safety Alignment Should Be Made More Than Just a Few Tokens Deep

Argues that current safety training often takes a shortcut, mainly changing the first few words of a model's reply. This 'shallow' safety helps explain why simple attacks, or even harmless fine-tuning, can get around it.

Qi, X., Panda, A., Lyu, K., Ma, X., Roy, S., Beirami, A., Mittal, P., & Henderson, P. (2024). Safety alignment should be made more than just a few tokens deep. arXiv. https://doi.org/10.48550/arXiv.2406.05946

Open resource
Continuum
READ
ResearchPreprint, not yet reviewed

How is ChatGPT's behavior changing over time?

Compared the March and June 2023 versions of GPT-3.5 and GPT-4 on the same tasks. Behaviour varied greatly between versions: for example, GPT-4 became less willing to answer sensitive questions in June than in March, and its accuracy on one maths task fell from 84% to 51%.

Why it matters now: The same product name can hide a different model with different behaviour.

Chen, L., Zaharia, M., & Zou, J. (2023). How is ChatGPT's behavior changing over time? arXiv. https://doi.org/10.48550/arXiv.2307.09009

Open resource
Continuum
READ
ResearchPeer reviewed

Characterizing artificial intelligence psychosis in a large medical setting

The first health-record study of 'AI psychosis'. Vanderbilt researchers searched progress notes for 215,712 behavioural health patients (December 2022 to April 2026) for AI keywords and found 28 patients whose chatbot use appeared to worsen psychosis, about 0.013% of those receiving mental health care, plus 28 more whose delusions featured AI without documented chatbot use. The AI psychosis group was younger (median 27.5 years), mostly male, and 61% were in a first psychotic episode, far higher than the comparison groups. ChatGPT was named in just over half of cases and 86% of interactions were recorded after the May 2024 release of GPT-4o. The commonest pattern was the chatbot acting as an amplifier of existing distorted beliefs (64%), rather than a catalyst for new symptoms (11%). Single site, retrospective, cannot show causation; the authors call for prospective studies using chat logs and for routine asking about AI use in psychiatric assessments.

Why it matters now: Puts a number on a risk that until now rested on case reports and news stories, and points to first-episode psychosis as the vulnerable window.

Bergson, Z., Vassall, S. G., Wright, A., McCoy, A. B., Schafer, K. M., Achee, M. C., & Sheffield, J. M. (2026, October 7). Characterizing artificial intelligence psychosis in a large medical setting [Research letter]. JAMA Psychiatry. Advance online publication. https://doi.org/10.1001/jamapsychiatry.2026.2923

Open resource
Continuum
Also relevant to

R2. Relationships with chatbots

What happens when people relate to a chatbot over time, including dependence, sycophancy and the use of emotional data?

READ
Opinion and commentaryNot reviewed

Psychoanalytic critique: AI tools between sessions and the therapeutic relationship

A psychotherapy commentary by Mark Vahrmeyer (Brighton and Hove Psychotherapy, 2025) arguing that AI journaling and chat tools used after difficult sessions can soothe patients in ways that bypass metabolising ruptures with the therapist, potentially undermining the therapeutic relationship by relocating processing into a digital third. Conceptual and clinical rather than quantitative, but it names a pattern worth watching.

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

AI chatbots: Companionship and minors (Virginia JCOTS supplementary report, 2026)

Staff report for Virginia's Joint Commission on Technology and Science, prepared alongside HB 635 and SB 796. Reviews evidence on emotional dependence, isolation, delusional spirals and self-harm linked to companion and general-purpose chatbots, notes the evidence on minors is still thin, and sets out six principles for redrafted legislation.

Virginia Joint Commission on Technology and Science. (2026). AI chatbots: Companionship and minors. Supplementary report to the 2025 JCOTS report on AI chatbots. Division of Legislative Services. https://dls.virginia.gov/commissions/jcots/Materials/2026_AI_Chatbots_Report.pdf

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewedAdvocacy organisation

No license required: risks of AI companion chatbots as mental-health support

Consumer Federation of America report on how unlicensed AI companions are being marketed and used as de-facto mental-health support, and the harms already observed.

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

AI-Powered Mental Health Virtual Assistants' Acceptance: An Empirical Study on Influencing Factors Among Generations X, Y, and Z

Survey of factors shaping acceptance of AI mental health virtual assistants across Generations X, Y and Z. Cureus is a lightly reviewed venue; treat findings as indicative.

Alanzi, T., Alsalem, A. A., Alzahrani, H., Almudaymigh, N., Alessa, A., Mulla, R., AlQahtani, L., Bajonaid, R., Alharthi, A., Alnahdi, O., & Alanzi, N. (2023). AI-Powered Mental Health Virtual Assistants' Acceptance: An Empirical Study on Influencing Factors Among Generations X, Y, and Z. Cureus, 15(11), e49486. https://doi.org/10.7759/cureus.49486

Open resource
Continuum
READ
Practice signalNot reviewed

Video: Blumenthal-authored bipartisan bill protecting children from AI chatbots approved by Senate Judiciary Committee

This press release announces Senate Judiciary Committee approval of the GUARD Act. The bill would ban AI companions for minors, require disclosures, and create crimes for specified chatbot conduct involving minors.

Why it matters now: The committee's approval marks a step in a bill proposing safeguards for minors using AI chatbots.

Blumenthal, R. (2026, April 30). Video: Blumenthal-authored bipartisan bill protecting children from AI chatbots approved by Senate Judiciary Committee [Press release]. https://www.blumenthal.senate.gov/newsroom/press/release/video-blumenthal-authored-bipartisan-bill-protecting-children-from-ai-chatbots-approved-by-senate-judiciary-committee

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewedAdvocacy organisation

Justification for AI safeguards for kids 2026

This Common Sense Media report argues that AI companions pose mental-health and safety risks to minors. It recommends that people under 18 not use AI for companionship or mental-health support.

Why it matters now: The report makes a case for safeguards as concerns focus on minors' use of AI companions.

Common Sense Media. (2026). Justification for AI safeguards for kids 2026. https://www.commonsensemedia.org/sites/default/files/featured-content/files/justification-for-ai-safeguards-for-kids-2026.pdf

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

The engagement-validation loop: Sycophancy, anthropomorphic projection, and clinical risk in informal AI-based mental health support

This article examines informal mental health support from generative AI, focusing on sycophancy and people projecting human qualities onto AI. Its title frames the resulting engagement and validation loop as a potential clinical risk.

da Silva Santos, B., Roza, T. H., & Passos, I. C. (2026). The engagement-validation loop: Sycophancy, anthropomorphic projection, and clinical risk in informal AI-based mental health support. Journal of Affective Disorders, 412, Article 122123. https://doi.org/10.1016/j.jad.2026.122123

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

How does turning to AI for companionship predict loneliness and vice versa?

This 12-month study followed 2,149 adults across four countries to examine links between social chatbot use and loneliness. Emotional isolation predicted more chatbot use later, and chatbot use also predicted greater emotional isolation, while chatbot use did not predict improved social connection.

Folk, D., & Dunn, E. (2026). How does turning to AI for companionship predict loneliness and vice versa? Psychological Science, 37(4), 276–286. https://doi.org/10.1177/09567976261427747

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

AI chatbots as companions: Overview, uses, and considerations for Congress

This Congressional Research Service report reviews AI chatbots used as companions and considerations for Congress. It says evidence about mental-health effects is complex and largely based on short-term studies or company disclosures.

Why it matters now: The report outlines evidence limits and congressional options relevant to current debate on AI companions.

Greene-Sanders, D. T. (2026). AI chatbots as companions: Overview, uses, and considerations for Congress (CRS Report No. R49189). Congressional Research Service. https://www.congress.gov/crs-product/R49189

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

Advertising to the distressed: The commodification of mental health data in AI chatbots

This Journal of Medical Ethics blog discusses the use of mental-health disclosures in AI chatbots. The authors warn that chatbots may use these disclosures to profile users and target advertising, while existing EU regulations do not directly address the risks.

Gross, N., & van Kolfschooten, H. (2026, April 14). Advertising to the distressed: The commodification of mental health data in AI chatbots. Journal of Medical Ethics Blog. https://blogs.bmj.com/medical-ethics/2026/04/14/advertising-to-the-distressed-the-commodification-of-mental-health-data-in-ai-chatbots/

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

GUARD Act

This US Senate bill proposes age checks for publicly available AI chatbots, prohibits minors from using AI companions and requires user accounts and disclosures about chatbots. It is intended to protect children from risks including grooming, addiction and self-harm, and proposes penalties of up to $250,000 per offence or violation.

Why it matters now: The 2026 bill is a current legislative item in the debate over AI chatbots.

GUARD Act, S. 3062, 119th Cong. (2026). https://www.congress.gov/bill/119th-congress/senate-bill/3062/text

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

AI companion startup Character.ai files for Chapter 11 bankruptcy, citing costs and litigation

This CNBC news article reports Character.ai's Chapter 11 bankruptcy filing and its cited litigation liabilities. It notes a Kentucky lawsuit alleging the platform led children into self-harm.

Why it matters now: The bankruptcy filing brings chatbot safety litigation into focus, including a lawsuit alleging harm to children.

Haddad, C. J. (2026, October 13). AI companion startup Character.ai files for Chapter 11 bankruptcy, citing costs and litigation. CNBC. https://www.cnbc.com/2026/10/13/character-ai-chapter-11-disney-bid.html

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Senator Hawley's GUARD Act to protect kids from AI chatbots passes committee unanimously

This press release reports that the GUARD Act passed committee unanimously. It says the bill would ban AI companions for minors, require disclosures, and impose criminal penalties for specified chatbot conduct involving minors.

Why it matters now: The committee vote advances proposed restrictions and disclosures for AI companions used by minors.

Hawley, J. (2026, April 30). Senator Hawley's GUARD Act to protect kids from AI chatbots passes committee unanimously [Press release]. https://www.hawley.senate.gov/senator-hawleys-guard-act-to-protect-kids-from-ai-chatbots-passes-committee-unanimously/

Open resource
Continuum
Also relevant to
READ
Guidance and standardsPeer reviewed

The AWARE (AI use, why, attachment, reality and risk, and effect on functioning) framework for assessing patients' use of AI in mental health care

This viewpoint introduces AWARE, a proposed framework for clinicians to discuss patients’ use of conversational AI in mental health care. It covers AI use, reasons for use, attachment, reality and risk, and effects on functioning, but has not been formally validated and is not intended as a diagnostic or screening tool.

Hudon, A., Bhat, V., Nunez, J. J., Paquin, V., Malouin-Lachance, A., & Lespérance, F. (2026). The AWARE (AI use, why, attachment, reality and risk, and effect on functioning) framework for assessing patients' use of AI in mental health care. JMIR Medical Education, 12, e108766. https://doi.org/10.2196/108766

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewed

AI companion legislation in the United States: What it means for the video game industry

This law journal article discusses Connecticut’s AI companion law and its implications for the video game industry. It says the law requires suicide and self-harm detection, prevention measures and mental health referrals, with heightened safeguards for minors.

Why it matters now: It explains how Connecticut’s new AI companion law applies to video game chatbots and the limits of their exclusion.

Irvin, S., Bhargava, M. N., & Knox, K. Q. (2026, August 13). AI companion legislation in the United States: What it means for the video game industry. Harvard Journal of Sports and Entertainment Law. https://journals.law.harvard.edu/jsel/2026/08/ai-companion-legislation-in-the-united-states-what-it-means-for-the-video-game-industry/

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Tool or companion? Reframing conversational AI to prevent psychological harm

This viewpoint argues that conversational AI should be a bounded tool that supports human care, not a companion or substitute for relationships. It warns that simulated empathy, overvalidation and sycophancy may encourage emotional reliance and reinforce unhelpful beliefs, and discusses a four-week trial linking greater trust and social attraction to a chatbot with greater emotional dependence and problematic use.

Maurich Novelli, A., Shergill, S., & Teixeira, A. S. (2026). Tool or companion? Reframing conversational AI to prevent psychological harm. JMIR Mental Health, 13, e99354. https://doi.org/10.2196/99354

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Meta contractors posed as teens to prompt rival chatbots about suicide, sex, and drugs

This investigative news article reports that Meta contractors posed as minors to test rival chatbots with high-risk prompts. It says the project’s secrecy and methods raised concerns about youth safety, consent and whether the work amounted to ordinary safety benchmarking.

Why it matters now: The report raises questions about youth safety, consent and methods in a large-scale chatbot safety-testing project.

Mehrotra, D., & Khalili, J. (2026, June 29). Meta contractors posed as teens to prompt rival chatbots about suicide, sex, and drugs. WIRED. https://www.wired.com/story/meta-contractors-pretending-to-be-teens-chatbot-testing/

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Proposal to ban AI companions for minors advances in Senate

This news article reports on the GUARD Act, which would ban AI companions from interacting with minors. The proposed bill would require age verification and chatbot disclosure, and impose penalties for sexually explicit content or encouragement of violence.

Why it matters now: A proposed US bill to ban AI companions for minors has advanced in the Senate.

Merod, A. (2026, May 4). Proposal to ban AI companions for minors advances in Senate. K-12 Dive. https://www.k12dive.com/news/proposal-to-ban-ai-companions-for-minors-advances-in-senate/819239/

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Delusions by design? How everyday AIs might be fuelling psychosis (and what can be done about it)

This review synthesises case reports, media accounts and preliminary research on possible links between everyday AI use and psychosis, covering about 15 cases. It describes recurring themes of spiritual or messianic revelations, beliefs in sentient or god-like AI, and attachment or romantic delusions, and argues that AI may reinforce existing delusions rather than cause schizophrenia from scratch.

Morrin, H., Nicholls, L., Levin, M., Yiend, J., Iyengar, U., DelGuidice, F., Bhattacharya, S., Tognin, S., MacCabe, J., Twumasi, R., Alderson-Day, B., & Pollak, T. A. (2026). Delusions by design? How everyday AIs might be fuelling psychosis (and what can be done about it) [Abstract]. European Psychiatry, 69(S1), S113. https://doi.org/10.1192/j.eurpsy.2026.10633

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

The spread of AI companions and the challenges they generate

This European Parliamentary Research Service briefing examines the spread of AI companions and the challenges they pose. It says they can contribute to emotional dependence and mental health harms, particularly among children, and outlines EU laws that may apply.

Why it matters now: The European Parliament briefing describes risks from AI companions while noting that the EU has no specific law for them.

Negreiro, M., & Anaç, Ö. D. (2026). The spread of AI companions and the challenges they generate (EPRS Briefing PE 789.299). European Parliamentary Research Service. https://www.europarl.europa.eu/RegData/etudes/BRIE/2026/789299/EPRS_BRI(2026)789299_EN.pdf

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Attorney General James urges court to protect children and consumers in Character.AI bankruptcy

This press release describes a filing by the New York Attorney General concerning Character.AI’s bankruptcy. The filing says chatbots posed as mental health professionals, discouraged professional advice or medication, and mishandled suicidal or self-harming behaviour.

Why it matters now: The New York Attorney General is asking a court to protect children and consumers during Character.AI’s bankruptcy.

New York State Office of the Attorney General. (2026, October 14). Attorney General James urges court to protect children and consumers in Character.AI bankruptcy [Press release]. EIN Presswire. https://www.einpresswire.com/article/949608931/attorney-general-james-urges-court-to-protect-children-and-consumers-in-character-ai-bankruptcy

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewedDeveloper research

Working with the American Psychological Association on youth mental health and AI

This OpenAI announcement describes planned work with the American Psychological Association on youth mental health and AI. The partnership aims to use psychological science and lived experience to develop safeguards, resources and guidance that support young people without replacing human relationships or care.

Why it matters now: OpenAI and the American Psychological Association plan joint work on AI safeguards and resources for young people’s well-being.

OpenAI. (2026, August 6). Working with the American Psychological Association on youth mental health and AI. https://openai.com/index/openai-and-apa-partner-to-advance-responsible-ai/

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Only time will tell: A structured survey of longitudinal studies on social AI companions

This review surveys longitudinal studies on social AI companions. It reports that short-term studies found reduced loneliness and social anxiety, while longer-term research raises concerns about increased loneliness and reduced socialisation.

Pi, Y., & Hunter, R. (2026). Only time will tell: A structured survey of longitudinal studies on social AI companions. International Journal of Human–Computer Interaction. Advance online publication. https://doi.org/10.1080/10447318.2026.2670529

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

How do chatbots affect your mental health?

This Université de Montréal news article describes the AWARE framework for mental health professionals discussing patients’ AI use. The framework assesses attachment, reality perception, risk and effects on daily functioning.

St-Cyr-Leroux, B. (2026, October 14). How do chatbots affect your mental health? UdeM Nouvelles. https://nouvelles.umontreal.ca/en/article/2026/10/14/how-do-chatbots-affect-your-mental-health

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Intimacy by design: Definition, state of research, and interdisciplinary research agenda on intimate human-AI interactions

This journal review defines intimate human-AI interactions, surveys research and sets out an interdisciplinary research agenda. It describes a seven-day study in which loneliness reductions were strongest on the first day and did not persist without continued interaction.

Szczuka, J. M., Mühl, L., & Schneeberger, T. (2026). Intimacy by design: Definition, state of research, and interdisciplinary research agenda on intimate human-AI interactions. AI & Society, 41, 8863–8878. https://doi.org/10.1007/s00146-026-03112-8

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

Mental health AI use may be popular but is it safe and able to shape someone’s long-term well-being?

This World Economic Forum web article discusses the safety of AI use for mental health. It argues that assessment should consider long-term effects on agency, wellbeing, dependency and real-world relationships, not only whether individual responses avoid harmful content.

Tanaka, H. (2026, June 18). Mental health AI use may be popular but is it safe and able to shape someone's long-term well-being? World Economic Forum. https://www.weforum.org/stories/artificial-intelligence/mental-health-ai-use-long-term-safety/

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Cruz, Schatz, Curtis, Schiff introduce new bill giving parents control over kids' AI chatbot use

This Senate committee press release announces the proposed CHATBOT Act. The bill would give parents tools to manage children’s chatbot use, limit manipulative features and targeted advertising, and study potential harms, including risks to children’s mental health.

Why it matters now: A proposed bill would give parents more control over children’s AI chatbot use and address potential mental health risks.

U.S. Senate Committee on Commerce, Science, and Transportation. (2026, April 28). Cruz, Schatz, Curtis, Schiff introduce new bill giving parents control over kids' AI chatbot use [Press release]. https://www.commerce.senate.gov/press/rep/release/cruz-schatz-curtis-schiff-introduce-new-bill-giving-parents-control-over-kids-ai-chatbot-use/

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

Living with AI Companions: Sustained AI Companionship Predicts Lower Well-Being Through Lower Human Interaction

The study surveyed Character.AI users at two waves, with 1,182 participants at baseline and 439 at follow-up after a mean of 362.5 days. Greater sustained engagement with AI companions was associated with lower well-being at follow-up, with lower in-person social interaction reported as an indirect pathway.

Zhang, Y.; Zhao, D.; Wang, Y. et al. (2026). Living with AI Companions: Sustained AI Companionship Predicts Lower Well-Being Through Lower Human Interaction. arXiv preprint, arXiv:2609.07243v1. https://arxiv.org/html/2609.07243

Open resource
Continuum
READ
ResearchPeer reviewed

When loneliness amplifies the risk of problematic emotional dependence: perceived anthropomorphism and basic psychological need satisfaction in AI companion use

A cross-sectional survey of 666 adults in China found that perceived anthropomorphism was associated with basic psychological need satisfaction (β = 0.569, p < 0.001), which was in turn associated with problematic emotional dependence (β = 0.317, p < 0.001). The indirect association was 0.180 (95% bootstrap CI [0.137, 0.229]); loneliness moderated the association in the full sample (β = 0.086, p = 0.005), but the study does not establish causation.

Liu, T.; Shih, R.-M.; Luo, T. et al. (2026). When loneliness amplifies the risk of problematic emotional dependence: perceived anthropomorphism and basic psychological need satisfaction in AI companion use. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1900817

Open resource
Continuum
READ
ResearchPreprint, not yet reviewed

Stumbling Into AI Emotional Dependence: How Routine AI Interactions Reshape Human Connection

The paper argues that emotional support can emerge from routine, task-focused AI interactions and influence whether people later seek help from AI or other people. It describes a 28-day study in which daily five-minute conversations about personal issues were associated with a 10.3% decrease in preference for human support and an 11.6% increase in preference for AI support.

Shi, Y., Fang, C. M., Maez, P. et al. (2026). Stumbling Into AI Emotional Dependence: How Routine AI Interactions Reshape Human Connection. arXiv (cs.AI). https://doi.org/10.48550/arXiv.2606.04150

Open resource
Continuum
READ
ResearchPeer reviewed

How emotional response styles of generative AI companions shape adaptive emotion regulation and short-term AI reliance

In a randomised experiment with 380 adults, both adaptive regulation-oriented and reassurance-oriented AI responses reduced immediate negative affect, with no significant difference in relief (b = 0.055, p = 0.323). Compared with reassurance, adaptive responses were associated with higher emotion-regulation self-efficacy (b = 0.500), better written strategy quality (b = 0.353) and greater recovery in a later no-AI task (b = 0.279), while reassurance was associated with greater perceived emotional necessity (b = 1.010) and short-term reliance intentions (b = 0.628); these intentions do not establish emotional dependence, and indirect-effect analyses do not prove causal mediation.

Wang, C. (2026). How emotional response styles of generative AI companions shape adaptive emotion regulation and short-term AI reliance. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1950163

Open resource
Continuum
READ
ResearchPreprint, not yet reviewed

What Parents Can See: Divergent Accounts of Youth AI Companion Use in Parenting and Teenager Subreddits

The authors analysed 1,628 Reddit posts from parenting and teenager communities, finding that teenagers more often discussed emotional support (31% of posts, versus 19% among parents), while parents more often discussed romantic and sexual interactions (36%, versus 25%). Teenagers described emotional support as a benefit more often than parents did (27% versus 5%), while attachment and dependence appeared at similar rates (19% of teenager posts and 16% of parenting posts).

Berkane, T., Bischops, A., Mellacheruvu, A., et al. (2026). What Parents Can See: Divergent Accounts of Youth AI Companion Use in Parenting and Teenager Subreddits. arXiv (cs.HC). https://doi.org/10.48550/arXiv.2609.20720

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

New Rules to Regulate Human-like AI Interactive Services

Rules issued jointly by five departments took effect on 15 July and define anthropomorphic AI services as sustained emotional interactions that simulate human traits, thinking and communication. They encourage uses such as cultural dissemination and companionship while prohibiting specified harmful content, deceptive collection of information and practices that foster emotional dependency, with special safeguards for minors.

Why it matters now: National rules that directly target emotional-dependency design in human-like AI (in force 15 July 2026).

Science and Technology Daily (2026). New Rules to Regulate Human-like AI Interactive Services. Science and Technology Daily. https://www.szzg.gov.cn/2026/english/dn/202608/t20260806_5355278.htm

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

AI and the end of loneliness (TED talk)

Psychologist Paul Bloom argues that AI companions may soon ease loneliness, a gift for isolated and older people but a danger for the young, who need the hard work of real relationships. He asks what happens to a generation raised on chatbots that never need an apology, and proposes a way to lean on AI without losing what matters. About 12 minutes.

Bloom, P. (2026, April). AI and the end of loneliness [Video]. TED2026. https://www.ted.com/talks/paul_bloom_ai_and_the_end_of_loneliness

Open resource
Continuum
WATCH
Opinion and commentaryNot reviewedDeveloper research

Can AI companions help heal loneliness? (TEDAI San Francisco talk by the founder of Replika)

Eugenia Kuyda, founder of Replika, argues that AI companions can be built to make people's human relationships better rather than replace them, and describes the choices she thinks that requires. About 12 minutes. A developer's case for the product category, best read alongside the independent studies of Replika users in this category.

Kuyda, E. (2024, October). Can AI companions help heal loneliness? [Video]. TEDAI San Francisco. https://www.ted.com/talks/eugenia_kuyda_can_ai_companions_help_heal_loneliness

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

Technology, AI, and what makes us human (Sherry Turkle at the Wisdom and AI Summit)

MIT sociologist Sherry Turkle, author of Alone Together and Reclaiming Conversation, on what she calls artificial intimacy: machines that perform empathy without having any. A short talk recorded in October 2023 that sets out the argument she has developed over three decades of studying people and their devices.

Turkle, S. (2023, November 16). Technology, AI, and what makes us human [Video]. Wisdom and AI Summit, San Francisco, October 2023. Wisdom 2.0. YouTube. https://www.youtube.com/watch?v=vqTVUx7HkjQ

Open resource
Continuum
WATCH
Opinion and commentaryNot reviewed

Artificial intimacy: When machines pursue our attachment (Sherry Turkle on the Future of Life Institute podcast)

A long interview in which Turkle discusses why people turn to chatbots for empathy and companionship, how conversational agents can exploit human vulnerabilities, and the specific risks for children and older people. Recorded in 2026, so it reflects the current generation of companion products.

Future of Life Institute. (2026, October 22). Artificial intimacy: When machines pursue our attachment (with Sherry Turkle) [Video podcast]. YouTube. https://www.youtube.com/watch?v=XSkfwbHl3E8

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

Esther Perel on counselling a man and his chatbot 'girlfriend' (ABC All in the Mind)

Psychotherapist Esther Perel talks with Sana Qadar about how AI has reached her therapy room, including a session with a man and his chatbot partner, and what it means for intimacy and relationships. Australian public broadcaster production, about half an hour.

ABC Science. (2026, April 7). Esther Perel on counselling a man and his chatbot 'girlfriend' [Video]. All in the Mind. YouTube. https://www.youtube.com/watch?v=fguVkDmIwY0

Open resource
Continuum
Also relevant to
WATCH
Analysis and advocacyNot reviewedAdvocacy organisation

AI and teen well-being: What do we know now? (Common Sense Media session)

A 46-minute recorded session from Common Sense Media bringing together its AI risk assessment team, clinical researchers from Stanford's Brainstorm Lab and a child safety policy voice from OpenAI. It draws on the organisation's survey finding that most US teens have used AI companions and that a third have turned to one for something serious instead of a person.

Common Sense Media. (2026, April 1). AI and teen well-being: What do we know now? [Video]. https://www.commonsensemedia.org/videos/ai-teen-well-being-what-do-we-know-now

Open resource
Continuum
Also relevant to

R2.1 Avatars of the deceased: griefbots and the digital afterlife

What happens when the chatbot speaks as someone who has died? Comfort, continuing bonds and unfinished conversations on one side; dependence, distorted memory, consent and a profit-driven afterlife industry on the other. Read with R2 (relationships with chatbots) and C3 (governance).

READ
Opinion and commentaryPeer reviewed

The ethics of 'deathbots'

An early ethics paper that shifts the focus from the dignity of the dead to the wellbeing and autonomy of the bereaved person using the bot. Argues that deathbots may interfere with grieving, foster dependence and open users to covert advertising, while possibly helping people with prolonged, severe grief. Proposes they be classified as medical devices so that harm and benefit have to be shown before sale.

Why it matters now: The medical device proposal is the strongest regulatory position in this literature.

Lindemann, N. F. (2022). The ethics of 'deathbots'. Science and Engineering Ethics, 28(6), 60. https://doi.org/10.1007/s11948-022-00417-x

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Griefbots, deadbots, postmortem avatars: on responsible applications of generative AI in the digital afterlife industry

Cambridge ethicists use three design-fiction scenarios to map who is affected: the data donor, the person who holds their data after death, and the person who talks to the result. Recommendations for providers include sensitive procedures for retiring a deadbot, meaningful transparency that the user is talking to an AI, adult-only access, and mutual consent of donor and user. They warn of product placement through a dead relative's voice and of children being told a dead parent is still with them.

Why it matters now: The most cited set of design standards for this industry; useful as a checklist.

Hollanek, T., & Nowaczyk-Basińska, K. (2024). Griefbots, deadbots, postmortem avatars: On responsible applications of generative AI in the digital afterlife industry. Philosophy & Technology, 37(2), 63. https://doi.org/10.1007/s13347-024-00744-w

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Generative ghosts: anticipating benefits and risks of AI afterlives

Distinguishes earlier griefbots that replay what a person said from 'generative ghosts' that produce new content in their name. Reviews the history and the start-ups, sets out a design space (who creates the ghost, whether it speaks as the person or about them, whether it can act in the world) and lays out a research agenda on benefits and harms for individuals and society.

Why it matters now: Frames the shift from replaying the dead to generating them, which is where the technology now sits.

Morris, M. R., & Brubaker, J. R. (2025). Generative ghosts: Anticipating benefits and risks of AI afterlives. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (pp. 1-14). ACM. https://doi.org/10.1145/3706598.3713758

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

The 'conversation' about loss: understanding how chatbot technology was used in supporting people in grief

Interviews with 10 mourners (nine men, one woman) who used chatbots during grief, some as a simulation of the dead person, others as listener, friend, partner or emotion coach. Participants described a safe place to express feelings and attempts to finish unfinished conversations, and most preferred the bot as a supplement to human therapy rather than a replacement. The sample was self-selected and the authors say claims of benefit need caution.

Why it matters now: Still the most cited empirical study of people actually using chatbots in grief.

Xygkou, A., Siriaraya, P., Covaci, A., Prigerson, H. G., Neimeyer, R., Ang, C. S., & She, W.-J. (2023). The 'conversation' about loss: Understanding how chatbot technology was used in supporting people in grief. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (pp. 1-15). ACM. https://doi.org/10.1145/3544548.3581154

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Designing conversations with the dead: how people engage with generative ghosts

Sixteen bereaved adults aged 22 to 50 each spent about 20 minutes with two AI versions of a person they had lost: one speaking as the person in the first person, one speaking about them in the third. A researcher reviewed every reply before it was shown. Participants preferred the first-person 'reincarnation' for its closeness, forgave invented facts but not wrong tone or terms of endearment, and all said they would use it again while most feared what it might do to other grieving relatives.

Why it matters now: The first controlled user study of generative ghosts, with a human in the loop that commercial products do not have.

Manning, J. M., Sullivan, D., Doyle, D. T., Pinter, A. T., & Brubaker, J. R. (2026). Designing conversations with the dead: How people engage with generative ghosts. In Proceedings of the 2026 ACM Designing Interactive Systems Conference (pp. 3610-3621). ACM. https://doi.org/10.1145/3800645.3813090

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Exploring public perceptions of artificial intelligence in bereavement support: a qualitative study of griefbots

Interviews with 25 adults in Guangdong, China, all with a university degree, found acceptance of griefbots was conditional rather than yes or no. People saw comfort, continued connection and a way to address regret, alongside disrupted grieving, inauthenticity, dependence, inequality, stigma and misuse of the dead person's data. Commemorative uses were more acceptable than commercial ones, and people in acute grief were seen as most at risk.

Why it matters now: Public attitudes from the country where the digital resurrection market is largest.

Lu, J., Yang, M., Cui, Y., Liu, Y., Wong, T. K. S., & Hu, Y. (2026). Exploring public perceptions of artificial intelligence in bereavement support: A qualitative study of griefbots. Frontiers in Digital Health, 8, 1852514. https://doi.org/10.3389/fdgth.2026.1852514

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

'I want my loved one back virtually': exploring the desire of bereaved people to create and maintain digital bonds with their deceased loved ones

A survey of middle-aged Japanese adults who had lost a first-degree relative in the previous ten years. Most did not want a digital bond, but nearly 20 percent wanted to be reunited with the dead person in a digital space. The desire for a digital bond at baseline predicted complicated grief five months later, which the authors read as a sign that such bonds may affect adjustment after bereavement.

Why it matters now: One of the few longitudinal signals: wanting the bond predicted worse grief, though the study did not test using a bot.

Kawashima, D., Kempe, T., & Koga, Y. (2023). 'I want my loved one back virtually': Exploring the desire of bereaved people to create and maintain digital bonds with their deceased loved ones. OMEGA - Journal of Death and Dying, 92(2), 1049-1065. https://doi.org/10.1177/00302228231194857

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Ethical dilemmas and the reconstruction of subjectivity in digital mourning in the age of AI: an empirical study on the acceptance intentions of bereaved family members of cancer patients

A questionnaire study of 129 bereaved relatives of cancer patients in southern China, mostly aged 18 to 25. Grief perception raised intention to use AI mourning tools and ethical concern lowered it; intention strongly predicted reported use. A small, young, single-region sample, but one of the first to measure acceptance in people who have actually been bereaved.

Fu, K., Ye, C., Wang, Z., Wu, M., Liu, Z., & Yuan, Y. (2025). Ethical dilemmas and the reconstruction of subjectivity in digital mourning in the age of AI: An empirical study on the acceptance intentions of bereaved family members of cancer patients. Frontiers in Digital Health, 7, 1618169. https://doi.org/10.3389/fdgth.2025.1618169

Open resource
Continuum
READ
Opinion and commentaryPeer reviewed

Griefbots. A new way of communicating with the dead?

A psychology paper placing griefbots within continuing bonds theory and alongside social media memorials. The key difference it draws: memorial pages are a shared social space where mourners talk about the dead together, while a griefbot is a private two-way conversation with a simulation, available anywhere and at any time. It closes on the ethical problems of a profit-driven afterlife industry for both mourners and the dead.

Why it matters now: Useful for clinicians: links griefbots to the continuing bonds framework they already use.

Jiménez-Alonso, B., & Brescó de Luna, I. (2023). Griefbots. A new way of communicating with the dead? Integrative Psychological and Behavioral Science, 57(2), 466-481. https://doi.org/10.1007/s12124-022-09679-3

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Embracing grief in the age of deathbots: a temporary tool, not a permanent solution

Argues that current deathbots cannot bear emotional burdens and are unlikely to give real comfort, so they should not be sold as a way to relieve grief or to keep talking to the dead. The authors propose a narrower role: a temporary means of mourning, used for a time and then put down, to limit dependence and respect the ordinary course of grief.

Bao, A., & Zeng, Y. (2024). Embracing grief in the age of deathbots: A temporary tool, not a permanent solution. Ethics and Information Technology, 26(1), 7. https://doi.org/10.1007/s10676-024-09744-y

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

The making of digital ghosts: designing ethical AI afterlives

Reviews the state of 'digital ghosts' and the design choices behind them. Notes that empirical reactions are mixed, culturally variable and marked by refusal as well as interest, that low-cost grief avatar services spread quickly in China, and that unauthorised recreations (such as the avatar of singer Qiao Renliang made without his parents' consent) show how little governance exists. Warns that prolonged, unstructured use may reinforce denial and dependence.

Spitale, G., & Germani, F. (2026). The making of digital ghosts: Designing ethical AI afterlives. Ethics and Information Technology, 28(3), 34. https://doi.org/10.1007/s10676-026-09910-4

Open resource
Continuum
Also relevant to
READ
Policy and regulationPeer reviewed

The law of digital afterlife: the Chinese experience of AI 'resurrection' and 'grief tech'

A legal analysis of AI 'resurrection' services, using China because the practice is most developed there. Covers the deceased's portrait and personality rights under the Civil Code, relatives' rights to consent or object, and the deep synthesis and generative AI rules that require consent, data protection and labelling of AI content. The legal questions apply well beyond China.

Why it matters now: The only jurisdiction with substantial rules and case experience on recreating the dead.

Cheng, K. Y. (2025). The law of digital afterlife: The Chinese experience of AI 'resurrection' and 'grief tech'. International Journal of Law and Information Technology, 33, eaae029. https://doi.org/10.1093/ijlit/eaae029

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Is grief dead? How digital resurrection technologies affect bereavement

A journalist's survey of the field for JMIR with grief researcher Robert Neimeyer and bioethicist Craig Klugman. Reports the digital legacy market at about US$22.46 billion in 2024 and expected to more than triple by 2034, notes that regulation and evidence trail availability, and describes Dutch work on therapist-guided 'deepfake therapy' with an avatar of the deceased. Experts in the piece see griefbots as a possible add-on to therapy and legacy work, not a replacement.

Why it matters now: A current, readable map of the market and the expert positions, in a health journal.

Rebernik, D. (2026). Is grief dead? How digital resurrection technologies affect bereavement. Journal of Medical Internet Research, 28, e111086. https://doi.org/10.2196/111086

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

AI 'deadbots' can fuel pathological grief and affect how we deal with death

A public health academic argues that posthumous avatars can make grief public, deepen dependence and keep people in an in-between state, with prolonged grief the main clinical worry. Notes that only a minority of current uses are supervised by a professional, that supervised use in therapy is the more defensible case, and recounts the Korean documentary in which a mother met a VR recreation of her daughter.

Marceau, E. (2026, April 29). AI 'deadbots' can fuel pathological grief and affect how we deal with death. The Conversation. https://theconversation.com/ai-deadbots-can-fuel-pathological-grief-and-affect-how-we-deal-with-death-280683

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

Grief and AI: deathbots, digital afterlife and responsible practice in bereavement care (Grief Australia webinar)

A professional development webinar from Grief Australia, presented by Christopher Hall AM, on how to respond when bereaved clients turn to AI contact with the dead. Covers why clients seek it (meaning-making, continuing bonds, cultural and spiritual variation), the risks (consent, privacy, family conflict, dependence, misinformation, exploitation) and practice frameworks: assessment questions, boundaries, documentation, safeguarding triggers and referral.

Why it matters now: Australian practitioner training already exists on this; a sign the issue has reached the consulting room.

Grief Australia. (2026, August 27). Grief and AI: Deathbots, digital afterlife and responsible practice in bereavement care [Webinar, presented by C. Hall]. https://www.grief.org.au/ga/Shared_Content/Events/Event_Display.aspx?EventKey=WB270826

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

Griefbots and the new afterlife: How AI is changing the ways we mourn (TEDxCU talk)

A former hospital chaplain turned AI researcher draws on stories from the bedside to ask whether we would talk to a loved one after they died, and what griefbots change about remembrance, storytelling and connection. About 14 minutes. A personal and emotional framing rather than a research summary.

Doyle, D. T. (2025, May 16). Griefbots and the new afterlife: How AI is changing the ways we mourn [Video]. TEDxCU. YouTube. https://www.youtube.com/watch?v=K7b_qYCJOUw

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

Time traveling with AI to connect with lost loved ones (TED talk)

Cartoonist Amy Kurzweil describes helping train a chatbot on the archives of her late grandfather, a man she never met, and discovering family history through it, illustrated with her own drawings. About 9 minutes. A first-person account of the hopeful side of the practice, useful alongside the ethics papers in this category.

Kurzweil, A. (2024, April). Time traveling with AI to connect with lost loved ones [Video]. TED2024. https://www.ted.com/talks/amy_kurzweil_time_traveling_with_ai_to_connect_with_lost_loved_ones

Open resource
Continuum
WATCH
Opinion and commentaryNot reviewed

How people are using AI to grieve, love and connect (TED talk)

Digital anthropologist Jennifer Cearns, after talking with hundreds of people who turn to AI for grief, love and friendship, argues that the fear that AI is replacing real relationships may have it backwards, and asks what these new relationships teach us about ourselves. About 14 minutes. A recent talk, recorded April 2026 and published October 2026.

Cearns, J. (2026, April). How people are using AI to grieve, love and connect [Video]. TED2026. https://www.ted.com/talks/jennifer_cearns_how_people_are_using_ai_to_grieve_love_and_connect

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewedDeveloper research

A.I. and afterlife (TEDxBucharest talk by the founder of Eternime)

Marius Ursache, founder of the early digital afterlife start-up Eternime, sets out his vision for preserving a person's memories and personality after death. Included as a record of how the industry pitched itself before large language models; read it against the Hollanek and Nowaczyk-Basińska paper on responsible applications in the same category.

Ursache, M. (2019, July 19). A.I. and afterlife [Video]. TEDxBucharest. YouTube. https://www.youtube.com/watch?v=bANg1rJxTUo

Open resource
Continuum
WATCH
Opinion and commentaryNot reviewed

Eternal You (documentary, 2024)

A feature documentary by Hans Block and Moritz Riesewieck that follows people who use digital afterlife services, and the companies selling them. It premiered at Sundance in 2024 and has been widely screened and reviewed since. The project site carries the trailer, screening information and a companion digital project on the digital afterlife.

Block, H., & Riesewieck, M. (Directors). (2024). Eternal You [Film]. Beetz Brothers Film Production. https://eternalyou.online/en/

Open resource
Continuum
Also relevant to
WATCH
Practice signalNot reviewed

How 'griefbots' allow people to 'talk to the dead' (BBC Tech Now)

A short BBC Tech Now report on the growing trend of griefbots or deadbots, with users and commentators. A quick orientation for anyone who has not met the idea before.

BBC Tech Now. (2025, June 13). How 'griefbots' allow people to 'talk to the dead' [Video]. BBC. https://www.bbc.com/reel/video/p0lhls9s/watch

Open resource
Continuum
WATCH
Opinion and commentaryNot reviewed

The ethics of griefbots (Uehiro Oxford Institute, Bitesize Ethics 2024)

Philosopher Cristina Voinea walks through the ethical questions raised by digital personas that mimic the dead, from consent and the interests of the deceased to the commercial incentives of companies such as HereAfter AI and You, Only Virtual. A short lecture for a general audience from the Uehiro Oxford Institute.

Voinea, C. (2024, July 17). Bitesize Ethics 2024, week 4: The ethics of griefbots [Video]. Uehiro Oxford Institute. YouTube. https://www.youtube.com/watch?v=AkO6LJ37a8o

Open resource
Continuum
WATCH
Opinion and commentaryNot reviewed

How dead relatives could haunt us via AI: Cambridge philosophers on griefbots (The Human Podcast)

Interview with Katarzyna Nowaczyk-Basińska and Tomasz Hollanek of Cambridge's Leverhulme Centre for the Future of Intelligence about their paper on griefbots, deadbots and postmortem avatars, including the design scenarios and the risk of being 'haunted' by unwanted digital presences. A long-form companion to their paper, which is in this category.

The Human Podcast. (2024, September 4). How dead relatives could haunt us via AI: Cambridge University philosophers [Video]. YouTube. https://www.youtube.com/watch?v=U5y72HT32Mo

Open resource
Continuum
WATCH
Opinion and commentaryNot reviewed

Can AI ease the pain of loss? (Scientific American, Science Quickly)

Science writer David Berreby describes his own experience with griefbots and what it revealed about mourning, memory and the limits of the technology. A reflective, first-person episode of Scientific American's Science Quickly podcast.

Scientific American. (2025, November 12). Can AI ease the pain of loss? [Video podcast]. Science Quickly. YouTube. https://www.youtube.com/watch?v=iiR3tX5hhFc

Open resource
Continuum

R3. Algorithmic feeds and commercial design

How do engagement-driven recommender systems and business models affect mental health?

READ
Practice signalNot reviewed

My feed, my way

This media release outlines draft Australian Digital Duty of Care legislation released for targeted consultation, including proposed safety standards and protections for people under 18. It would let social media users over 16 choose whether their default feed includes algorithmic recommendations, and non-compliance could attract penalties of up to $109.2 million.

Why it matters now: The Australian Prime Minister's September release adds a government announcement to debate about algorithmic feeds.

Albanese, A., & Wells, A. (2026, September 8). My feed, my way [Media release]. Prime Minister of Australia. https://www.pm.gov.au/media/my-feed-my-way

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewed

Towards a digital fairness act

This European Parliamentary Research Service briefing examines the proposed Digital Fairness Act and how EU consumer law applies to digital services, drawing on a review of the 2017–2023 period. It reports estimated annual harm of €7.9 billion from problematic digital commercial practices, with 69% of public consultation respondents reporting difficulty cancelling an online contract and 62% saying a subscription renewed without a reminder.

Why it matters now: The EPRS briefing examines proposals for a digital fairness act, informing current debate on platform design.

Ashton, D., & Niestadt, M. (2026, October 5). Towards a digital fairness act (EPRS Briefing PE 791.529). European Parliamentary Research Service. https://eprs.europarl.europa.eu/contents/publications/EPRS/2026/10/EPRS_BRI(2026)791529.html

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewedIndustry funded

Australia: Digital duty of care to redefine online safety

This legal alert analyses Australia’s draft Online Safety Amendment (Digital Duty of Care) Bill 2026, which would require online services to take reasonably practicable steps to provide a safe environment and assess risks at least annually. It describes proposed complaint, transparency and research-data duties, expanded eSafety powers and maximum civil penalties of AUD 109.2 million.

Why it matters now: The legal alert interprets Australia's digital duty of care and its implications for online safety.

Baker McKenzie. (2026, September). Australia: Digital duty of care to redefine online safety [Legal alert]. https://www.bakermckenzie.com/-/media/files/insight/publications/2026/09/australia-digital-duty-of-care-to-redefine-online-safety.pdf

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Kids Online Safety Act

This Congressional Research Service briefing compares the proposed Kids Online Safety provisions in House-passed H.R. 7757 and amended S. 1748. It notes that H.R. 7757 requires policies addressing specified harms but omits S. 1748’s duty of care, and that report-response deadlines vary between the bills and, for S. 1748, by platform size.

Why it matters now: The CRS brief provides a current overview of the Kids Online Safety Act for legislative debate.

Cho, C. Y. (2026). Kids Online Safety Act (CRS In Focus No. IF12730). Congressional Research Service. https://www.congress.gov/crs-product/IF12730

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

New social media curfews and crackdown on addictive features to better protect 16- and 17-year-olds online

This government press release describes a UK pilot involving more than 300 teenagers and parents, alongside research on social media restriction pilots. Families said overnight curfews became routine and helped with sleep and concentration; the government proposes midnight-to-6am curfews and default limits on autoplay and personalised feeds for 16- and 17-year-olds.

Why it matters now: The UK government announced proposed curfews and a crackdown on addictive features for 16- and 17-year-olds.

Department for Science, Innovation and Technology. (2026, July 15). New social media curfews and crackdown on addictive features to better protect 16- and 17-year-olds online [Press release]. GOV.UK. https://www.gov.uk/government/news/new-social-media-curfews-and-crackdown-on-addictive-features-to-better-protect-16-and-17-year-olds-online

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

A digital duty of care for Australia: Developing a duty of care framework for online services used by Australians

This issues paper sets out a proposed Digital Duty of Care framework for online services used in Australia, following government consultation in late 2025 and early 2026. It argues that services should take reasonable steps to prevent foreseeable online harms and says further consultation will inform the framework’s design.

Why it matters now: The issues paper invites consideration of an Australian digital duty-of-care framework for online services.

Department of Infrastructure, Transport, Regional Development, Communications, Sport and the Arts. (2026, May 26). A digital duty of care for Australia: Developing a duty of care framework for online services used by Australians [Issues paper]. https://www.infrastructure.gov.au/department/media/publications/digital-duty-care-australia-developing-duty-care-framework-online-services-used-australians

Open resource
Continuum
Also relevant to
READ
Position statementNot reviewed

Recommender systems: Position paper

Australia's regulator sets out safer recommender design, user controls and digital literacy. Seven in ten adults frequently see content they consider harmful, but only 26% have used content controls.

Why it matters now: The regulator urges safer recommender design as many adults encounter harmful content and few use controls.

eSafety Commissioner. (2026). Recommender systems: Position paper. https://www.esafety.gov.au/sites/default/files/2026-05/Recommender-systems-position-paper-May2026.pdf

Open resource
Continuum
READ
ResearchNot reviewed

Social Media Minimum Age evaluation: Methodology report

This methodology report sets out a two-year evaluation of Australia’s Social Media Minimum Age legislation, following 4,121 children aged 10–16 and a parent or carer for each child across five survey waves. It argues that the mixed-methods design can assess implementation and short- to medium-term outcomes, but cannot establish causation definitively because the policy applies nationally and there is no comparison group.

eSafety Commissioner. (2026). Social Media Minimum Age evaluation: Methodology report. Australian Government. https://www.esafety.gov.au/sites/default/files/2026-07/SMMA-Evaluation-Methodology-Report-July2026.pdf

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

Instagram chief Adam Mosseri warns against turning off algorithm as Australia debates opt-out laws

This news report covers Instagram chief Adam Mosseri’s comments on Australia’s proposed requirement for platforms to offer a non-algorithmic feed. Mosseri said chronological feeds can reduce engagement by up to 50 per cent and guessed that only a low single-digit share of Australian users would choose one.

Lavoipierre, A. (2026, September 10). Instagram chief Adam Mosseri warns against turning off algorithm as Australia debates opt-out laws. ABC News. https://www.abc.net.au/news/2026-09-10/instagram-ceo-adam-mosseri-algorithm-laws/107136992

Open resource
Also relevant to
READ
ResearchNot reviewed

The dark side of social media: Recommender algorithms and mental health

This working paper uses Dutch panel data to study the effects of Instagram’s 2016 move to an algorithmic feed, comparing Instagram users with users of other social platforms. It finds worse mental health among teenagers of both sexes, with social isolation and negative social comparison as possible pathways, and greater effects on self-worth and self-esteem among females.

Mandile, S. (2026). The dark side of social media: Recommender algorithms and mental health (CESifo Working Paper No. 11648). CESifo. https://doi.org/10.65864/0209bsp0yi

Open resource
Continuum
READ
ResearchPeer reviewedLived experience

Unraveling entangled feeds: Rethinking social media design to enhance user well-being

People living with anxiety, depression or PTSD described feeds as both harmful and a source of support, validation and connection. They proposed design changes: intentional use, context for engagement and explicit feed controls.

Milton, A., Runningen, D., Terveen, L., Kaur, H., & Chancellor, S. (2026). Unraveling entangled feeds: Rethinking social media design to enhance user well-being. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (pp. 1–16). Association for Computing Machinery. https://doi.org/10.1145/3772318.3791252

Open resource
Continuum
READ
Policy and regulationNot reviewed

Addictive design on online platforms

This European Parliament briefing reviews addictive design on online platforms. It reports that the European Commission preliminarily found TikTok in breach of the Digital Services Act over its design, and that a US jury found Meta and YouTube liable for addictive design.

Why it matters now: It summarises policy attention to addictive design on online platforms from the European Parliament's research service.

Negreiro Achiaga, M. D. M. (2026). Addictive design on online platforms (EPRS At a Glance PE 785.735). European Parliamentary Research Service. https://www.europarl.europa.eu/thinktank/en/document/EPRS_ATA(2026)785735

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Australia targets big tech with algorithm opt-out bill

This news report describes Australia’s proposed “My Feed, My Way” bill, which would let users opt out of algorithmic recommendations and see content from accounts they follow. Platforms that fail to comply could face fines of up to AU$100 million, and the bill was expected to be considered later in 2026.

Why it matters now: It reports Australia's proposed algorithm opt-out bill as lawmakers target large technology platforms.

Reynoso, E. (2026, September 8). Australia targets big tech with algorithm opt-out bill. DW. https://amp.dw.com/en/australia-targets-big-tech-with-algorithm-opt-out-bill/a-79154108

Open resource
Also relevant to
READ
ResearchPeer reviewed

Algorithmic recommendation and adolescent mental health: A grounded theory study of social and commercial determinants

This UK interview study spoke with 27 young people aged 14–19 about recommended content on TikTok and Instagram. Participants described repeated exposure, emotional feedback loops and limited control, leading the authors to argue for both media literacy education and changes to recommendation systems and platform governance.

Winstone, L., Parsonage, J., Cross, L., Muir, C., Randhawa, A., MacArthur, G., Garavini, E., & Asiedu-Addo, E. (2026). Algorithmic recommendation and adolescent mental health: A grounded theory study of social and commercial determinants. BMC Public Health, 26, Article 2526. https://doi.org/10.1186/s12889-026-28300-5

Open resource
Continuum
READ
ResearchNot reviewedAdvocacy organisation

Breaking the algorithm: Redesigning social media for youth well-being

Youth-led survey: 53% said online communities were critical to their social support and 48% said they affirmed their identity. Nearly all wanted digital wellbeing education. Frames the goal as redesigning feeds, not removing them.

Adebukola, T., Gerry, S., Lee, K., Ohakamma, I., Shedeed, M., Khedr, M., Menjivar, J., & Davis, K. (2025). Breaking the algorithm: Redesigning social media for youth well-being. Mental Health America. https://mhanational.org/wp-content/uploads/2025/03/Breaking-the-Algorithm-report.pdf

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Health benefits of social media use in adolescents and young adults

Reviews the benefits of social media for connection, identity, emotional support, help-seeking and mental health information. Notes reliance on qualitative evidence and inaccurate health content.

Nagata, J. M., Huang, O., Hur, J. O., Li, E. J., Helmer, C. K., Weinstein, E., & Moreno, M. A. (2025). Health benefits of social media use in adolescents and young adults. Current Pediatrics Reports, 13, Article 22. https://doi.org/10.1007/s40124-025-00357-7

Open resource
Continuum
READ
Analysis and advocacyPeer reviewed

A right to constructive optimization: A public interest approach to recommender systems in the Digital Services Act

Argues people have a right to recommender systems optimised for their interests and civic values, not only for platform engagement. Reframes the feed debate from limiting harm to requiring constructive design.

Why it matters now: It argues the Digital Services Act should require recommenders to serve public interests, not just platform engagement.

Naudts, L., Helberger, N., Veale, M., & Sax, M. (2025). A right to constructive optimization: A public interest approach to recommender systems in the Digital Services Act. Journal of Consumer Policy, 48(3), 269–296. https://doi.org/10.1007/s10603-025-09586-1

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Balancing the benefits and risks of social media on adolescent mental health in a post-pandemic world

A viewpoint describing social media as both a lifeline, especially for marginalised teens, and a stressor. Proposes a multi-stakeholder framework built on digital literacy, industry accountability and clinical integration.

Osborne, A. (2025). Balancing the benefits and risks of social media on adolescent mental health in a post-pandemic world. Child and Adolescent Psychiatry and Mental Health, 19, Article 92. https://doi.org/10.1186/s13034-025-00951-z

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Social drivers and algorithmic mechanisms on digital media

Finds the evidence that algorithms themselves drive mental health harm is thin. Proposes using algorithms to detect risk, offer support contacts and encourage healthier use.

Metzler, H., & Garcia, D. (2024). Social drivers and algorithmic mechanisms on digital media. Perspectives on Psychological Science, 19(5), 735–748. https://doi.org/10.1177/17456916231185057

Open resource
Continuum
Also relevant to
READ
ResearchNot reviewed

Early days, early insights: Understanding experiences of social media age restrictions at the three-month follow-up

In a three-month follow-up survey, the share of children aged 10-15 with an account on at least one age-restricted platform fell from about 52% to 42%, while use of at least one such platform declined from almost 86% to more than 81%. The report says unchecked ages were the most common reason children retained accounts, and parental awareness of children’s social media use declined, with broader outcomes showing little change.

Why it matters now: First official evaluation data on Australia's under-16 social media restrictions.

eSafety Commissioner (2026). Early days, early insights: Understanding experiences of social media age restrictions at the three-month follow-up. eSafety Commissioner. https://www.esafety.gov.au/research/social-media-age-restrictions-evaluation/early-days-early-insights-three-months-report

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

EU KIDS ACT - 'EU Keeping Internet Digital Spaces Accountable and Trustworthy'

The proposal would set EU-wide online safety rules for minors, including preventing children under 15 from creating autonomous accounts on specified social networking and video-sharing services, while allowing guardian-created limited accounts for ages 13-14. It also proposes safety-by-design and age-assurance requirements, with an expedited Commission enforcement process aiming for preliminary findings within 30 working days and a final decision within 90 working days.

Why it matters now: Would set an EU-wide minimum age of 15 for autonomous accounts plus safety-by-design duties.

European Commission (2026). EU KIDS ACT - 'EU Keeping Internet Digital Spaces Accountable and Trustworthy'. European Commission; EUR-Lex. https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:52026PC0681

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Attorney General Hanaway Launches Investigation into TikTok’s Impact on Missouri Children

Missouri Attorney General Catherine Hanaway opened an investigation into whether TikTok misled families or put young users at risk through its platform design and operation. A Civil Investigative Demand seeks information about TikTok’s recommendation systems, youth safety practices, age checks, data use and internal research, with a response due by November 13, 2026.

Attorney General Office of Missouri (2026). Attorney General Hanaway Launches Investigation into TikTok’s Impact on Missouri Children. Attorney General Office of Missouri. https://ago.mo.gov/attorney-general-hanaway-launches-investigation-into-tiktoks-impact-on-missouri-children/

Open resource
Continuum
READ
Practice signalNot reviewed

Newsom signs strict social media protections for teens

California Gov. Gavin Newsom signed a law banning social media platforms from offering features deemed addictive to users under 16. Platforms must remove those features to continue serving younger teens, with infinite scroll and autoplay cited as examples.

Why it matters now: Design-feature restrictions (infinite scroll, autoplay) rather than access bans.

Knowles, H.; Duncan, I. (2026). Newsom signs strict social media protections for teens. The Washington Post. https://www.washingtonpost.com/nation/2026/09/10/newsom-signs-bill-banning-addictive-features-social-media-protect-teens/

Open resource
Continuum
Also relevant to
READ
ResearchNot reviewed

Health warnings could reduce young people’s social media use, Stanford Medicine-led study finds

In an online experiment, 1,012 people aged 13-29 rated 14 social media health warnings and a control message, with every warning judged more effective than the control. Warnings about overall mental health harms, depression and anxiety were rated most effective, and the warnings increased participants’ knowledge of social media risks; teens and young adults responded similarly.

Digitale, E. (2026). Health warnings could reduce young people’s social media use, Stanford Medicine-led study finds. Stanford Medicine News. https://med.stanford.edu/news/all-news/2026/09/social-media-warnings.html

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Why young people say the social media ban is not working six months after it was introduced

Six months after Australia introduced its social media ban, 70 per cent of 898 parents surveyed said their children still had active social media accounts. In a test involving 11 Year 9 students, Yoti assessed all of them as under 18, while the article also reports that 4.7 million accounts were initially locked or deactivated and that some young people later created new accounts without age checks.

Why it matters now: Young people's own account of workarounds; set beside the eSafety evaluation.

Gillett, W. (2026). Why young people say the social media ban is not working six months after it was introduced. ABC News. https://www.abc.net.au/news/2026-06-14/why-teens-say-social-media-ban-isnt-working-for-under-16s/106780590

Open resource
Continuum
READ
Practice signalNot reviewed

New draft bill on social media age limits in Denmark

A draft Danish bill would prohibit children under 15 from accessing certain social media services and video-sharing platforms, with no parental-consent exception for 13- and 14-year-olds and no penalties for children or parents. The public consultation closes on 5 October 2026, the bill is expected to go to Parliament in early 2027, and it is set to take effect on 1 July 2027.

Jurcenoks, E., Korsgaard, R. L., Hansen, K. (2026). New draft bill on social media age limits in Denmark. Legal News, Plesner. https://www.plesner.com/en/news/new-draft-bill-social-media-age-limits-denmark

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Commission preliminarily finds TikTok's addictive design in breach of the Digital Services Act

The European Commission preliminarily found TikTok in breach of the Digital Services Act over design features including infinite scroll, autoplay, push notifications and its personalised recommender system, saying TikTok had not adequately assessed or mitigated risks to users’ wellbeing. The investigation began on 19 February 2024, and if the findings are confirmed, TikTok could face a fine of up to 6% of its total worldwide annual turnover; the preliminary findings do not prejudge the outcome.

Why it matters now: Earlier 2026 (6 Feb), but missing: EU preliminary finding that TikTok's addictive design breaches the Digital Services Act.

European Commission (2026). Commission preliminarily finds TikTok's addictive design in breach of the Digital Services Act. European Commission Presscorner. https://ec.europa.eu/commission/presscorner/api/files/document/print/en/ip_26_312/IP_26_312_EN.pdf

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Meta, Google lose US case over social media harm to kids

A Los Angeles jury found Meta and Google negligent in designing Instagram and YouTube and failing to warn about their dangers, awarding $4.2 million against Meta and $1.8 million against Google. The $6 million verdict concerns a plaintiff who said she became addicted to the platforms as a young person, and is intended as a bellwether for thousands of similar lawsuits; both companies said they plan to appeal.

Why it matters now: Earlier 2026 (25 Mar), but missing: landmark jury verdict on social media design harm.

Chmielewski, D., Rozen, C., Godoy, J. (2026). Meta, Google lose US case over social media harm to kids. Reuters. https://www.reuters.com/legal/litigation/jury-reaches-verdict-meta-google-trial-social-media-addiction-2026-03-25/

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewedIndustry funded

Online Safety Amendment (Digital Duty of Care) Bill 2026: exposure draft

Summarises two Australian exposure drafts released nine days apart: the Privacy Amendment (Personal Data Protection) Bill 2026 (31 August) and the Online Safety Amendment (Digital Duty of Care) Bill 2026 (8 September), which proposes a risk-based duty of care with a 12-month transition.

Why it matters now: The duty-of-care bill has moved from discussion to draft legislation. Update the existing Baker McKenzie and DITRDCA duty-of-care cards. Replace this link with the government exposure-draft page when it is available.

Want, A. (Keypoint Law) (2026). Online Safety Amendment (Digital Duty of Care) Bill 2026: exposure draft. Keypoint Law. https://www.keypointlaw.com.au/keynotes/digital-regulation-in-the-spotlight-two-exposure-drafts-in-nine-days/

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

How a handful of tech companies control billions of minds every day (TED talk)

Former Google design ethicist Tristan Harris explains how attention-driven products are designed to capture and hold users, and argues for a different design goal. The 2017 talk that put persuasive design on the public agenda, and still the clearest short statement of the feeds problem.

Harris, T. (2017, April). How a handful of tech companies control billions of minds every day [Video]. TED2017. https://www.ted.com/talks/tristan_harris_how_a_handful_of_tech_companies_control_billions_of_minds_every_day

Open resource
Continuum
WATCH
Opinion and commentaryNot reviewed

What we're getting wrong about teens and tech (TED talk)

Developmental psychologist Candice Odgers, a leading sceptic of the claim that smartphones caused a teen mental health crisis, argues that the panic misreads the evidence and distracts from what is actually driving young people's distress. About 12 minutes. The counterweight to the social media harm narrative in this category.

Odgers, C. (2026, April). What we're getting wrong about teens and tech [Video]. TED2026. https://www.ted.com/talks/candice_odgers_what_we_re_getting_wrong_about_teens_and_tech

Open resource
Continuum
Also relevant to

R4. Therapeutic claims and devices

What evidence and regulation should apply to tools that claim to help?

READ
Tool or datasetNot reviewed

FDA: List of Artificial Intelligence-Enabled Medical Devices

The FDA's 'List of Artificial Intelligence-Enabled Medical Devices' is a resource that identifies AI-powered medical devices authorized for marketing in the United States. This list serves as a transparency tool for healthcare providers and patients, while also providing digital health innovators with insights into the current device landscape and regulatory expectations to foster innovation and ensure public safety. Sortable by review panel, so it is the quickest way to see how few authorised AI devices sit under psychiatry or neurology compared with radiology. Downloadable as CSV, Excel or XML; the FDA notes it is not comprehensive.

U.S. Food and Drug Administration. (2026, September 4). List of artificial intelligence-enabled medical devices. https://www.fda.gov/medical-devices/artificial-intelligence-enabled-medical-devices/list-artificial-intelligence-enabled-medical-devices

Open resource
Continuum
READ
Evidence reviewPeer reviewed

Governing AI in Mental Health: 50-State Legislative Review

Systematic review of bills introduced in all 50 US states between January 2022 and May 2025, coded with a four-tier relevance taxonomy and 25 topic tags. Finds states acting as de facto laboratories for mental health AI policy while clinicians and professional bodies have been largely absent from the process. Senior author John Torous.

Shumate, J. N., Rozenblit, E., Flathers, M., Larrauri, C. A., Hau, C., Xia, W., Torous, E. N., & Torous, J. (2025). Governing AI in Mental Health: 50-State Legislative Review. JMIR Mental Health, 12, e80739. https://doi.org/10.2196/80739

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Governing AI for Mental Health: Fragmented State Approaches and the Case for a Federal Framework

Viewpoint comparing Illinois, Utah, New York and Nevada. Identifies two governance approaches, one regulating AI use in clinical contexts and one regulating the technology itself, and argues that because mental health AI straddles clinical and consumer domains neither alone is enough; a risk-based federal floor is proposed.

Aldhalimi, A. (2026). Governing AI for Mental Health: Fragmented State Approaches and the Case for a Federal Framework. JMIR Mental Health, 13, e96389. https://doi.org/10.2196/96389

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

FDA Digital Health Advisory Committee: generative AI-enabled digital mental health medical devices (6 November 2025)

First FDA advisory committee meeting devoted to generative AI mental health devices. The committee saw promise for access and personalisation but flagged hallucinated advice, missed crises, model drift, bias and privacy risks, and recommended risk-tiered oversight, premarket clinical evidence, clinician escalation pathways and clear labelling that users are talking to a machine.

U.S. Food and Drug Administration, Center for Devices and Radiological Health. (2025, November 6). Digital Health Advisory Committee meeting: Generative artificial intelligence-enabled digital mental health medical devices. Brief summary. https://www.fda.gov/media/190450/download

Open resource
Continuum
Also relevant to
Related
READ Meeting transcript (PDF)
READ FDA executive summary (PDF)
READ
Position statementNot reviewedAdvocacy organisation

Consumer Federation of America testimony on Maryland chatbot legislation HB 1250 / SB 827 (March 2026)

Advocacy testimony from Ben Winters (CFA) to the Maryland Economic Matters Committee. Argues chatbots are marketed as 'therapy' while causing harm, and asks for product-liability treatment, limits on using chat inputs for advertising, and a ban on chatbots presenting as licensed health providers. Useful as a snapshot of the state-level legislative push, not as evidence.

Winters, B. (2026, March). Testimony of the Consumer Federation of America in support of Maryland chatbot legislation: HB 1250/SB 827. Maryland General Assembly, Economic Matters Committee. https://mgaleg.maryland.gov/cmte_testimony/2026/ecm/35643_03172026_95014-394.pdf

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewedIndustry funded

New state laws regulating AI use for mental healthcare

This Baker Botts article discusses new state laws on AI use in mental healthcare. It says the laws restrict AI from independently providing mental healthcare or psychotherapy and set requirements including human oversight, client consent and disclosure.

Why it matters now: New state laws set limits on AI-delivered mental healthcare and requirements for professionals using AI.

Baker Botts. (2026, October 14). New state laws regulating AI use for mental healthcare. https://www.bakerbotts.com/thought-leadership/publications/2026/october/new-state-laws-regulating-ai-use-for-mental-healthcare

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

States have passed 14 new laws this year regulating the use of AI in healthcare

This article reports on new state laws regulating AI in healthcare. It says five states enacted laws prohibiting or restricting AI chatbot therapy, generally requiring licensed professionals to provide therapy services.

Barcott, B. (2026, July 27). States have passed 14 new laws this year regulating the use of AI in healthcare. Transparency Coalition. https://www.transparencycoalition.ai/news/state-lawmakers-have-passed-15-new-laws-regulating-the-use-of-ai-in-health-care

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Generative AI for mental health: Promise, risk and the path forward

This editorial discusses the promise, risks and governance needs of generative AI in mental health care. It argues that AI could help address mental health needs only with sound evidence, safety regulation and global collaboration, noting that nearly 1.2 billion people lived with mental disorders in 2023. It also warns that, without that groundwork, AI could create a two-tier system in which some people get human care and others get only a chatbot.

Brown, M., & Herz, N. (2026). Generative AI for mental health: Promise, risk and the path forward. BMJ Digital Health & AI, 2(1), e000158. https://doi.org/10.1136/bmjdh-2026-000158

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

AI for proactive mental health: A multi-institutional, longitudinal randomized controlled trial

This randomised trial tested the Flourish AI-powered app with 486 undergraduates at three US institutions over six weeks. Compared with usual care, app users reported greater resilience and social well-being and were buffered against declines in positive affect, mindfulness and flourishing, but depression, anxiety and stress did not differ significantly between groups.

Cachia, J. Y. A., Zhao, X., Hunter, J., Wu, D., Lin, E., & De Freitas, J. (2026). AI for proactive mental health: A multi-institutional, longitudinal randomized controlled trial. NEJM AI. https://doi.org/10.1056/AIoa2501293

Open resource
Continuum
READ
Evidence reviewPeer reviewed

The digital mirror: Clinical potentials and relational risks of generative AI in mental health interventions

This review considers clinical uses and relational risks of generative AI in mental health interventions. It describes potential support for assessment, treatment planning and between-session care, while arguing for clinician supervision because of risks such as dependency and reinforcing maladaptive beliefs.

Cavalera, C., Frisone, F., Rossi, C., Oasi, O., Pagnini, F., Riva, G., & Antichi, L. (2026). The digital mirror: Clinical potentials and relational risks of generative AI in mental health interventions. Current Psychiatry Reports, 28, Article 40. https://doi.org/10.1007/s11920-026-01690-4

Open resource
Also relevant to
READ
Guidance and standardsPeer reviewed

A framework for clinical validation of generative artificial intelligence therapeutics

This journal letter proposes a four-phase framework combining clinical and technical validation of generative AI mental health therapeutics. It includes ongoing monitoring of safety, efficacy, model drift and real-world effectiveness.

Galatzer-Levy, I. R., Tomasev, N., Rodriguez, C., & Torous, J. (2026). A framework for clinical validation of generative artificial intelligence therapeutics. World Psychiatry, 25(2), 327–328. https://doi.org/10.1002/wps.70067

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

Software and AI compliance and reforms

These TGA presentation slides discuss compliance and reforms for software and AI. They report stakeholder calls for urgent review of the digital mental health tools exclusion and that 62% of consultation respondents considered it no longer appropriate.

Harradine, G. (2026). Software and AI compliance and reforms [Presentation slides]. Therapeutic Goods Administration. https://www.tga.gov.au/sites/default/files/2026-07/software-and-ai-compliance-and-reforms.pdf

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Generative AI versus human conversational agent for reducing procrastination: A randomized pilot trial

This randomised pilot trial compared three ChatGPT-4 sessions with human coaching for procrastination in 62 university students. Procrastination outcomes did not differ significantly between groups, and 43.1 per cent of participants reported at least one unwanted negative effect.

Hennemann, S., Fähnrich, J. M., Tietze, C., Witthöft, M., & Jungmann, S. M. (2026). Generative AI versus human conversational agent for reducing procrastination: A randomized pilot trial. DIGITAL HEALTH, 12, Article 20552076261450336. https://doi.org/10.1177/20552076261450336

Open resource
READ
Evidence reviewPeer reviewed

Therapeutic interaction features of AI chatbots in depression interventions: Systematic review and meta-analysis

This systematic review and meta-analysis examined 11 randomised trials of AI chatbots for depression, involving 2,220 participants. The overall effect on depressive symptoms and adherence was not statistically significant, although exploratory findings linked emotional responsiveness, structured feedback and more frequent interaction with higher adherence.

Huang, T., Li, S., Wang, Y., & Liu, W. (2026). Therapeutic interaction features of AI chatbots in depression interventions: Systematic review and meta-analysis. Journal of Medical Internet Research, 28, e88697. https://doi.org/10.2196/88697

Open resource
Also relevant to
READ
Evidence reviewPeer reviewed

Generative AI in youth mental health apps: Rapid review

This rapid review examined five studies of generative AI mental health apps for young people aged 11 to 24, published from 2022 to November 2025. Young people generally rated the apps positively, and some studies reported promising outcomes for depression and anxiety, but none systematically assessed harms using standard methods or reported adverse mental health outcomes.

Høgsdal, H., Kyrrestad, H., & Kaiser, S. (2026). Generative AI in youth mental health apps: Rapid review. JMIR Mental Health, 13, e90589. https://doi.org/10.2196/90589

Open resource
Also relevant to
READ
Analysis and advocacyNot reviewed

State regulation of AI in health care

This Practical Law feature reviews state regulation of AI in healthcare. It describes limits on AI-delivered therapy and provider use, alongside privacy, advertising and disclosure requirements for mental health chatbots.

Why it matters now: The feature summarises state limits on AI mental healthcare and requirements for chatbots and providers.

Kuesel, J. (2026, June 1). State regulation of AI in health care. Practical Law The Journal. https://www.reuters.com/practical-law-the-journal/transactional/feature-state-regulation-ai-2026-06-01/

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Effectiveness of a fully automated mobile therapeutic versus a general chatbot in reducing depression and anxiety and improving well-being: Feasibility randomized controlled trial

This feasibility trial compared an AI therapy chatbot, ChatGPT and an assessment-only control in 147 adults over three weeks. Both chatbot groups had greater reductions in depression scores than the control, but did not differ significantly from each other; anxiety and well-being outcomes also showed no significant differences between groups.

Kuta, B., Novak, L., Zidkova, R., Furstova, J., Malinakova, K., De Winter, A., & Husek, V. (2026). Effectiveness of a fully automated mobile therapeutic versus a general chatbot in reducing depression and anxiety and improving well-being: Feasibility randomized controlled trial. JMIR Mental Health, 13, e82642. https://doi.org/10.2196/82642

Open resource
READ
ResearchPeer reviewed

Automated fidelity monitoring of lay-delivered mental health interventions using large language models: Development and pilot validation of shamiriAI in Kenya

This pilot study tested shamiriAI, a large language model system for monitoring the fidelity of mental health sessions delivered by lay providers in six Kenyan secondary schools. Its ratings showed substantial agreement with human ratings in some domains, but the system tended to rate sessions lower overall and under-rate required content and clarity.

Lilan, S., Mochama, B., Osborn, T., Mmbone, W., Kilonzo, R., Kamau, F., Daya, R., & Wasanga, C. (2026). Automated fidelity monitoring of lay-delivered mental health interventions using large language models: Development and pilot validation of shamiriAI in Kenya. JMIR AI, 5, e95063. https://doi.org/10.2196/95063

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Large language models and their applications in mental health: Scoping review

This scoping review covered 41 peer-reviewed studies on large language model applications in mental health, including diagnosis, prediction and decision support. Although many studies reported strong performance, only 13 evaluated models against clinician assessments using real patient data, and the authors call for more standardised evaluation and evidence from clinical outcomes.

Lokadjaja, M. C., Kho, J. J., Schulz, P. J., & Goh, W. W. (2026). Large language models and their applications in mental health: Scoping review. JMIR Mental Health, 13, e88057. https://doi.org/10.2196/88057

Open resource
Also relevant to
READ
Analysis and advocacyNot reviewed

Different state regulatory approaches reflect open questions about AI mental health tools

This KFF report examines different state approaches to AI mental health tools. It describes restrictions on providing or claiming to provide therapy and requirements for data protection, disclosure and patient consent.

Why it matters now: The report shows that states are taking varied approaches to regulating AI mental health tools.

Luther, J., Yilma, H., & Washington, I. (2026, August 27). Different state regulatory approaches reflect open questions about AI mental health tools. KFF. https://www.kff.org/health-information-trust/different-state-regulatory-approaches-reflect-open-questions-about-ai-mental-health-tools/

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewed

How is AI regulated for mental health in Australia?

This Black Dog Institute explainer describes how existing laws and professional requirements apply to AI used for mental health. It says many consumer-facing tools, especially general-purpose chatbots, lack AI-specific oversight and may pose risks including privacy breaches, bias and misinformation.

Why it matters now: The explainer highlights oversight gaps and risks for consumer-facing mental health AI tools.

Maidment, K. (2026). How is AI regulated for mental health in Australia? (CRE Precision Explainer Series No. 12). Black Dog Institute. https://www.blackdoginstitute.org.au/wp-content/uploads/2026/07/CRE-Explainer-series-12-How-is-AI-regulated-for-mental-health-in-Australia.pdf

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

MHRA digital mental health newsletter June 2026

This MHRA newsletter describes the programme’s next phase for digital mental health tools. It says the phase will focus on implementation, AI and international collaboration, including an AI airlock to test ways of managing risks.

Why it matters now: The MHRA’s next programme phase includes testing approaches for managing AI-related risks in digital mental health tools.

Medicines and Healthcare products Regulatory Agency. (2026, June). MHRA digital mental health newsletter June 2026. https://content.govdelivery.com/accounts/UKMHRA/bulletins/41dad27

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Effectiveness of chatbots in mental health screening and assessment: Systematic review

This systematic review examined 18 studies of chatbots and AI agents used for adult mental health screening and assessment. The tools showed promising accuracy, including sensitivity of 0.84 to 0.93 and specificity of 0.80 to 0.96 for depression and anxiety, but the authors say they need clinical protocols, professional oversight and bias safeguards, and should complement rather than replace clinical judgement.

Morales Gil, I., Marti-Estevez, I., Doval, S., Gutiérrez-Carrillo, J. M., Fausor, R., Arribas-García, S., Álvarez-Mon, M. Á., & Quintero, J. (2026). Effectiveness of chatbots in mental health screening and assessment: Systematic review. JMIR Mental Health, 13, e93672. https://doi.org/10.2196/93672

Open resource
Also relevant to
READ
Evidence reviewPeer reviewed

An umbrella review of artificial intelligence applications in mental health care

This umbrella review synthesises research on AI applications in mental health care. It reports the strongest evidence in screening, early detection and risk prediction, and recommends human-centred care with clinical oversight rather than replacing clinicians.

Odame, J., Obeng-Gyamfi, G., Heo, D., Kusi, A. B., Kwanin, C., Adarkwah, M. A., & Osei, E. A. (2026). An umbrella review of artificial intelligence applications in mental health care. Journal of Psychiatric and Mental Health Nursing. Advance online publication. https://doi.org/10.1111/jpm.70180

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Orygen secures $5.7 million Wellcome funding to advance safe, AI-powered youth mental health care

Orygen reports receiving Wellcome funding for the AI4You(th) project. The project will develop and evaluate SensAI, an AI collaborator for young people and clinicians, with a focus on safety, transparency, agency, personalisation and collaboration.

Why it matters now: Wellcome funding will support development and evaluation of an AI collaborator for young people and clinicians.

Orygen. (2026, May 21). Orygen secures $5.7 million Wellcome funding to advance safe, AI-powered youth mental health care. https://www.orygen.org.au/About/News-And-Events/2026/Orygen-secures-Wellcome-funding-to-advance-safe,-A

Open resource
Also relevant to
READ
Opinion and commentaryNot reviewed

Which states ban AI therapy? The full 2026 map: State-by-state legal status

This Psychology.com article maps state laws on AI therapy and mental health chatbots as of 20 July 2026. It reports that four states prohibit AI from delivering therapy, while four others regulate mental health or companion chatbots without banning them.

Pennock, S. F. (2026, July 20). Which states ban AI therapy? The full 2026 map: State-by-state legal status. Psychology.com. https://psychology.com/ai-therapy/state-bans

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Generative AI as a de facto mental health provider: A policy brief and urgent call for regulation

This policy brief considers the growing use of general-purpose generative AI for mental health support and reports a study of 584 users. Nearly three in ten users said they used it for emotional support or coping advice, rising to 47% among people with severe distress; the authors argue that regulation is urgently needed to address risks such as misinformation, privacy breaches and emotional dependency.

Refoua, E., Gigi, K., Levkovich, I., Hadar Shoval, D., Elyoseph, Z., Tsafrir, I., Pen, O., Angert, T., & Haber, Y. (2026). Generative AI as a de facto mental health provider: A policy brief and urgent call for regulation. Frontiers in Digital Health, 8, Article 1708806. https://doi.org/10.3389/fdgth.2026.1708806

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Efficacy of a conversational AI agent for psychiatric symptoms and digital therapeutic alliance: A randomized clinical trial

This randomised trial compared a 12-week conversational AI programme with group therapy and a waitlist among 995 university students in Israel. The AI group had larger reductions in anxiety and greater improvements in wellbeing than both comparison groups, while its therapeutic alliance ratings were associated with engagement and symptom improvement.

Shoshani, A., Gurfinkel, B., Kor, A., Ben-Haim, Y., Kanarek, O., Segev, R., Shafir, O., & Arbel, R. (2026). Efficacy of a conversational AI agent for psychiatric symptoms and digital therapeutic alliance: A randomized clinical trial. JAMA Network Open, 9(4), e266713. https://doi.org/10.1001/jamanetworkopen.2026.6713

Open resource
Also relevant to
READ
Analysis and advocacyNot reviewedIndustry funded

States continue efforts to regulate AI in healthcare: A review of legislation passed in 2026

This Holland & Knight article reviews state legislation passed in 2026 on AI in healthcare. It describes limits on clinical decision-making, informed-consent requirements and chatbot safeguards, including crisis protocols and restrictions on claims of professional healthcare capability.

Why it matters now: The review summarises state legislation passed in 2026 on AI use in healthcare, including behavioural and mental healthcare.

Silverboard, D. M., Barrera, J. A., & Ryan, I. (2026, May 26). States continue efforts to regulate AI in healthcare: A review of legislation passed in 2026. Holland & Knight. https://www.hklaw.com/en/insights/publications/2026/05/states-continue-efforts-to-regulate-ai-in-healthcare

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

Digital mental health tools (DMHTs)

This TGA guidance explains when digital mental health tools may be regulated as medical devices. It says regulation depends on the tool’s intended purpose, not the technology used.

Therapeutic Goods Administration. (2026). Digital mental health tools (DMHTs). Australian Government Department of Health, Disability and Ageing. https://www.tga.gov.au/products/medical-devices/software-and-artificial-intelligence-ai/overview/types-software-based-medical-devices/digital-mental-health-tools-dmhts

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewedIndustry funded

Evolving expectations for AI in mental health

This Wilson Sonsini client advisory discusses legal and regulatory scrutiny of AI mental health products. It recommends assessing product functions and claims, and incorporating disclosures, crisis protocols, oversight and monitoring.

Why it matters now: The advisory outlines evolving FDA, state-law, privacy, FTC and liability scrutiny of AI mental health products.

Wilson Sonsini. (2026, September 22). Evolving expectations for AI in mental health. https://www.wsgr.com/en/insights/evolving-expectations-for-ai-in-mental-health.html

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Emotional support or problem-solving? A randomized controlled trial of function-specific AI chatbots for digital stress in college students

This randomised trial tested emotional-support and problem-solving AI chatbots for digital stress, with mental health outcomes as a secondary focus. It assigned 99 university students with elevated digital stress to one of the chatbots or a waitlist, but the available text does not report the results.

Yang, Q., Wu, H., Du, X., Sun, Q., & Ren, Z. (2026). Emotional support or problem-solving? A randomized controlled trial of function-specific AI chatbots for digital stress in college students. Applied Psychology: Health and Well-Being, 18(4), e70208. https://doi.org/10.1111/aphw.70208

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewedAdvocacy organisation

(Cth) Ensure AI-enabled digital mental health tools are regulated by the TGA

This FORE Good policy proposal recommends repealing the exclusion for digital mental health tools. It argues that all tools using AI-driven features should be regulated as medical devices by the TGA.

Why it matters now: The proposal calls for AI-enabled digital mental health tools to be regulated by the TGA.

Young, J., Richter-Weinstein, M., Day, A., Subba, S., & Devamulla, D. (2026, September 24). (Cth) Ensure AI-enabled digital mental health tools are regulated by the TGA [Policy proposal]. FORE Good. https://www.foregood.org.au/post/cth-ensure-ai-enabled-digital-mental-health-tools-are-regulated-by-the-tga

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Consultation: Reviewing the exclusion for certain digital mental health tools

The TGA is seeking feedback on three options for the exclusion from regulation for certain digital mental health tools: keep it, remove it or modify it. Introduced in 2021, the exclusion covers tools based on established clinical practice guidelines that are referenced and displayed in the software, and the consultation closes on 2026-12-07.

Why it matters now: Live Australian consultation on whether some digital mental health tools should stay outside TGA regulation.

Therapeutic Goods Administration (TGA) (2026). Consultation: Reviewing the exclusion for certain digital mental health tools. Therapeutic Goods Administration; Citizen Space. https://consultations.tga.gov.au/tga/consultation-dmht-exclusion/

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Supporting Digital Mental Health Technology Developers: a new Community of Practice and survey to support regulatory compliance

The MHRA plans to establish a Community of Practice for digital mental health technology developers and manufacturers, with in-person sessions on topics including product classification, pre-market considerations and post-market surveillance. An anonymous survey is open until 19 November 2026 to help set priorities, and the first session is planned for 2027; the post also notes that the Government accepted all 44 recommendations from the National Commission into the Regulation of AI in Healthcare.

Raby, E. (2026). Supporting Digital Mental Health Technology Developers: a new Community of Practice and survey to support regulatory compliance. MedRegs Blog, Medicines and Healthcare products Regulatory Agency (MHRA). https://medregs.blog.gov.uk/2026/10/22/supporting-digital-mental-health-technology-developers-a-new-community-of-practice-and-survey-to-support-regulatory-compliance/

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback

The FDA discussion paper seeks feedback on risks and regulatory considerations for generative AI-enabled medical devices, including risk assessment, premarket evaluation and postmarket monitoring. It is for discussion only, does not propose policy changes or regulatory expectations, and invites feedback through docket FDA-2026-N-7874 by 2026-10-19.

Digital Health Center of Excellence (DHCoE), Center for Devices and Radiological Health (CDRH), U.S. Food and Drug Administration (2026). Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback. U.S. Food and Drug Administration (FDA). https://www.fda.gov/medical-devices/digital-health-center-excellence/considerations-regulation-generative-ai-enabled-medical-devices-discussion-paper-and-request

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Gen AI policy is a White House priority, CDRH digital health leader says

FDA digital health director Rick Abramson said the agency plans to draft generative AI policy soon, after receiving more than 500 submissions on its discussion paper. The FDA’s fiscal 2027 priorities include draft guidance on generative AI-enabled devices for mental disorders, while the UK regulator adopted 44 AI recommendations and plans further guidance and consultation.

Reuter, E. (2026). Gen AI policy is a White House priority, CDRH digital health leader says. MedTech Dive; article from The MedTech Conference 2026. https://www.medtechdive.com/news/gen-ai-policy-is-a-white-house-priority-cdrh-digital-health-leader-says/833551/

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Expert perspectives on US regulatory approaches to large language models in healthcare

This second article in a three-part interview series presents the personal views of Dr Shantanu Nundy, an advisor in the US FDA Commissioner’s Office, rather than official FDA policy. It describes assessing health tools by clinical risk and autonomy, and discusses FDA oversight, evidence requirements, general-purpose LLMs giving medical advice, and agency activity including advisory committee meetings in 2024 and 2025 and updated guidance in January 2026.

Gilbert, S.; Dai, T. (2026). Expert perspectives on US regulatory approaches to large language models in healthcare. npj Digital Medicine. https://doi.org/10.1038/s41746-026-02787-z

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Efficacy, User Engagement, and Acceptability of Cognitive Behavioral Therapy-Oriented Psychological Chatbots for Adults With Depressive and/or Anxiety Symptoms: Systematic Review and Meta-Analysis of Randomized Controlled Trials

Across 29 randomised trials, CBT-oriented chatbots were associated with reduced depressive symptoms after treatment (27 trials; Hedges g = −0.55) and reduced anxiety symptoms (25 trials; g = −0.26). The evidence was low to very low certainty, average attrition was 22%, and users reported technical problems and repetitive or unnatural interactions.

Gong, B.; Yao, N.; Xie, H. et al. (2026). Efficacy, User Engagement, and Acceptability of Cognitive Behavioral Therapy-Oriented Psychological Chatbots for Adults With Depressive and/or Anxiety Symptoms: Systematic Review and Meta-Analysis of Randomized Controlled Trials. Journal of Medical Internet Research. https://doi.org/10.2196/82677

Open resource
Continuum
READ
Practice signalNot reviewed

Slingshot pulls therapy chatbot Ash out of U.K. over regulatory concerns

Slingshot AI withdrew its Ash chatbot from the UK, saying there was no clear medical-device pathway for "wellbeing products like ours". The company had raised US$93 million.

Why it matters now: Earlier 2026, but needed as context for the October 2026 JMIR survey of former Ash users.

Aguilar, M. (STAT) (2026). Slingshot pulls therapy chatbot Ash out of U.K. over regulatory concerns. STAT. https://www.statnews.com/2026/01/21/slingshot-therapy-chatbot-ash-uk-regulatory-concerns/

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewedDeveloper research

The mental health AI chatbot made for real life (TED talk by the founder of Woebot)

Psychologist Alison Darcy, who founded Woebot Health, talks with Kelly Corrigan about the vision behind Woebot, support for people in tough moments when no one else is around, and what we should expect and demand from ethically designed psychological AI. About 14 minutes. Recorded in April 2025; the company retired its consumer app that June.

Darcy, A. (2025, April). The mental health AI chatbot made for real life [Video]. TED2025. https://www.ted.com/talks/alison_darcy_the_mental_health_ai_chatbot_made_for_real_life

Open resource
Continuum
Also relevant to
WATCH
Guidance and standardsNot reviewedLived experience

Can AI chatbots really be good therapists? (Black Dog Institute, Let's Unpack That)

Digital mental health researcher Kylie Maidment and clinical psychologist Anthony Joffe of the Black Dog Institute talk through what the research says about AI chatbots for mental health, who they might suit, why they feel so supportive, what guardrails exist and how this differs from a therapist using AI for notes. Hosted by a lived experience advisor. Australian, chaptered, about half an hour.

Black Dog Institute. (2026, September 8). Let's unpack: Can AI chatbots really be good therapists? [Video]. YouTube. https://www.youtube.com/watch?v=iGx5i1bKUy4

Open resource
Continuum
Also relevant to
WATCH
Guidance and standardsNot reviewed

What are the risks and opportunities of AI in mental health? (Black Dog Institute panel)

A Black Dog Institute discussion from late 2023 on the race to build AI tools that assess, diagnose and treat mental illness at scale, why early efforts look promising and where the risks of harm lie. Useful as a marker of how an Australian research institute framed the questions in the first year after ChatGPT.

Black Dog Institute. (2023, December 12). What are the risks and opportunities of AI in mental health? [Video]. YouTube. https://www.youtube.com/watch?v=TFZZRLKQgSw

Open resource
Continuum
Also relevant to

Part 3: Conditions that shape both benefit and risk

These issues decide whether a potential benefit is realised, and for whom.

C1. Cultural bias and cultural validity

Whose knowledge and norms are built into AI, and can it be adapted to local meaning? Bias is a risk; cultural adaptation is a benefit.

READ
Evidence reviewPeer reviewed

The case for cultural competency in psychotherapeutic interventions

A foundational review on cultural competency in psychotherapy — essential context for anyone asking whether a Western-trained LLM can safely serve non-Western populations.

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Artificial Intelligence in mental health and the biases of language based models

Case study of gender and other biases in language models used in mental health applications.

Straw, I., & Callison-Burch, C. (2020). Artificial Intelligence in mental health and the biases of language based models. PLOS ONE, 15(12), e0240376. https://doi.org/10.1371/journal.pone.0240376

Open resource
READ
Opinion and commentaryNot reviewed

How can AI models be adapted to incorporate cultural sensitivity in mental health support?

Student journal article on adapting AI models to Western and non-Western cultural contexts in mental health diagnosis and treatment.

How can AI models be adapted to incorporate cultural sensitivity in diagnosing and treating mental health conditions, considering Western and non-Western cultural contexts? Intersect: The Stanford Journal of Science, Technology, and Society. https://ojs.stanford.edu/ojs/index.php/intersect/article/view/3449

Open resource
READ
Opinion and commentaryPeer reviewed

Reflections on mental health, Ngozi, and the Dandemutande approach in Zimbabwe

Reflections on mental health, Ngozi (avenging spirits) and the Dandemutande approach in Zimbabwe, offering a culturally grounded alternative frame for distress.

Mutanga, O. (2025). Reflections on mental health, Ngozi, and the Dandemutande approach in Zimbabwe. African Journal of Disability, 14, a1599. https://doi.org/10.4102/ajod.v14i0.1599

Open resource
Continuum
READ
Guidance and standardsPeer reviewed

Culture by Design: A Sociotechnical Framework for Culturally Grounded AI for Mental Health

Conceptual framework with ten recommendations for culturally grounded AI across data design, modelling, evaluation, deployment and governance. Supports community involvement and context-sensitive design beyond translation; the recommendations have not been evaluated for population-health impact.

Rai, S., Stade, E., Zhou, G., Sehgal, N., Schriger, S., Gerke, S., Ungar, L., & Guntuku, S. C. (2026). Culture by Design: A Sociotechnical Framework for Culturally Grounded AI for Mental Health. Proceedings of the 10th Workshop on Computational Linguistics and Clinical Psychology, 128-137. Association for Computational Linguistics. https://doi.org/10.18653/v1/2026.clpsych-1.10

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

The anthropomorphization of AI and the concept of Buddhist compassion in human-machine interaction

A conceptual analysis proposing Buddhist compassion as a framework for ethically designing robots and human-machine interaction. It discusses equality, reciprocity, responsibility, dignity and culturally responsive care, including potential assistance for older people. It distinguishes programmed compassionate behaviour from human moral agency and argues that care robots should assist rather than replace caregivers.

Miao, F. (2026). The anthropomorphization of AI and the concept of Buddhist compassion in human-machine interaction. Frontiers in Psychology, 16, 1583565. https://doi.org/10.3389/fpsyg.2025.1583565

Open resource
Also relevant to
READ
ResearchPeer reviewed

Language and cultural prompts influence large language models' responses to culturally anchored mental health attitudes

Two experiments used a 2 × 2 design varying prompt language (Chinese or English) and cultural context (China or the United States) to test ChatGPT-4o and DeepSeek-V3 with a message about a friend experiencing mild depression. English prompts increased both models’ depression assessments and likelihood of recommending therapy, while US context lowered ChatGPT-4o’s assessment but raised DeepSeek-V3’s assessment and therapy recommendation likelihood.

Jia, M. et al. (2026). Language and cultural prompts influence large language models' responses to culturally anchored mental health attitudes. PEC Innovation. https://doi.org/10.1016/j.pecinn.2026.100505

Open resource
Continuum
READ
Opinion and commentaryPeer reviewed

AI Mental Health in Asia: Cultural Misalignment, Clinical Risks and the Critical Role of Psychiatric Nursing

The authors discuss cultural risks in three AI mental health modalities: prediction and screening models, chatbots, and therapy bots, which may misinterpret culturally specific experiences and preferences in Asian contexts. They call for culturally equitable data, fairness audits and responsive design, highlighting psychiatric nurses’ roles in evaluating and governing these tools; Asia has more than half of the world’s population but remains underrepresented in mental health research and AI development.

Kuek, J. H. L., Suendermann, O. & Goh, Y.-S. (2026). AI Mental Health in Asia: Cultural Misalignment, Clinical Risks and the Critical Role of Psychiatric Nursing. International Journal of Mental Health Nursing. https://doi.org/10.1111/inm.70328

Open resource
Continuum
Also relevant to
READ
Evidence reviewPreprint, not yet reviewed

AI-Powered Mental Health Chatbots in Africa: A Systematic Review and Culturally Adaptive Framework

The authors review 52 empirical studies published between 2017 and 2025, finding that mental health chatbots may improve access to support and help reduce anxiety and depression symptoms, but are rarely adapted to African cultures and languages. They propose a Culturally Adaptive Digital Mental Health framework focused on African cultural values, multilingual and mobile-first design, and ethical safeguards.

Lebese, M., Tsibolane, P. (2026). AI-Powered Mental Health Chatbots in Africa: A Systematic Review and Culturally Adaptive Framework. arXiv (cs.AI); IST-Africa 2026. https://doi.org/10.48550/arXiv.2608.24890

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Implicit Gender, Racial, and Ethnic Biases in Large Language Models: An Audit Study of Automated Psychiatric Diagnoses

The study tested six large language models on 97 psychiatric training cases adapted across 39 demographic groups while keeping symptoms constant. GPT-4o identified at least one correct diagnosis in 82.8% of cases but added diagnoses not supported by the cases in 70.3%, with reported differences in diagnostic patterns by gender, race and ethnicity.

Why it matters now: Earlier 2026 (Mar), but missing.

Pendse, S. R., Jain, M., Kumar, N. et al. (2026). Implicit Gender, Racial, and Ethnic Biases in Large Language Models: An Audit Study of Automated Psychiatric Diagnoses. Mayo Clin Proc Digit Health. https://doi.org/10.1016/j.mcpdig.2026.100344

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

“It Hasn’t Lived in Our Society”: Investigating Cultural Sensitivity in LLM Chatbots for Emotional Support

The study developed a culturally sensitive emotional support chatbot for young Saudi women and compared its responses with those of a Saudi therapist and a generic GPT chatbot. The culturally sensitive chatbot aligned more closely with the therapist in 78% of expert comparisons, compared with 3% for generic GPT, and participants preferred it in 87% of comparisons, while also raising concerns about over-aligning with cultural norms.

Why it matters now: Earlier 2026 (Apr), but missing.

Aldaweesh, S., Alelsheikh, G., Alhamed, F. et al. (2026). “It Hasn’t Lived in Our Society”: Investigating Cultural Sensitivity in LLM Chatbots for Emotional Support. CHI '26: Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3772318.3791060

Open resource
Continuum
WATCH
Opinion and commentaryNot reviewed

Decolonizing AI (TEDxRosario talk by Genevieve Bell)

Australian anthropologist Genevieve Bell, who grew up in Indigenous communities in outback Australia before Stanford, Intel and the ANU, takes apart the idea of AI as a single inevitable technology and visits its alternative stories past, present and future, to make room for a safer and more human-centred AI. About 9 minutes. An early and accessible statement of the cultural validity question this category covers.

Bell, G. (2020, August). Decolonizing AI [Video]. TEDxRosario. https://www.ted.com/talks/genevieve_bell_decolonizing_ai

Open resource
Continuum
Also relevant to

C2. Data and Indigenous sovereignty

Who controls the data that AI learns from and generates?

READ
Guidance and standardsPeer reviewed

Operationalizing the CARE and FAIR Principles for Indigenous data futures

Sets out how the CARE Principles for Indigenous Data Governance can be operationalised alongside FAIR principles in data repositories and research infrastructure.

Carroll, S. R., Herczog, E., Hudson, M., Russell, K., & Stall, S. (2021). Operationalizing the CARE and FAIR Principles for Indigenous data futures. Scientific Data, 8(1), 108. https://doi.org/10.1038/s41597-021-00892-0

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

Indigenous data sovereignty in the digital age: CARE principles and implementation frameworks

Working document from the Collaboratory for Indigenous Data Governance on implementing the CARE principles.

Collaboratory for Indigenous Data Governance. (2025). Indigenous data sovereignty in the digital age: CARE principles and implementation frameworks [Working document]. https://indigenousdatalab.org/

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

The Costs of Connection: Data Colonialism

Book developing the concept of data colonialism and examining the appropriation of human life through data relations. Provides a wider framework for thinking about community control and benefit-sharing; it is not a study of mental-health data being used to train Northern AI models.

Couldry, N., & Mejias, U. A. (2019). The Costs of Connection: How Data Is Colonizing Human Life and Appropriating It for Capitalism. Stanford University Press. Preface and chapter 1: https://law.unimelb.edu.au/__data/assets/pdf_file/0008/3290381/Couldry-and-Mejias-Preface-and-Ch-1.pdf

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

Te Hiku: Data Sovereignty and the Kaitiakitanga License

Te Hiku's account of data sovereignty, stewardship and Indigenous language technology, with resources on the Kaitiakitanga License. Offers a practical governance example for community-led AI; it does not evaluate mental-health outcomes.

Te Hiku Tech. (2022, May 6; updated 2023, October 11). Data Sovereignty and the Kaitiakitanga License. https://tehiku.nz/te-hiku-tech/te-hiku-dev-korero/25141/data-sovereignty-and-the-kaitiakitanga-license

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Competing Interests: Digital Health and Indigenous Data Sovereignty

Perspective on Indigenous health data, including community-led governance, culturally relevant indicators, disaggregation and participation in interpretation. Useful for population-health systems that support community priorities and control over data.

Cordes, A., et al. (2024). Competing interests: digital health and indigenous data sovereignty. npj Digital Medicine, 7, 178. https://doi.org/10.1038/s41746-024-01171-z

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Ethical governance of artificial intelligence in digital health for Indigenous populations: A narrative review and conceptual framework

This review of 18 articles examines ethical issues in AI-enabled digital health involving Indigenous communities in five countries. It identifies six key domains, including co-design, cultural relevance, bias, privacy and Indigenous data sovereignty, and proposes a governance framework built around participation and Indigenous data sovereignty.

Khan, A., Bear, A. S. N., Rackow, M., Wajuntah, C. O., Lai, K., McKinney, V., Costa, J., & Mendez, I. (2026). Ethical governance of artificial intelligence in digital health for Indigenous populations: A narrative review and conceptual framework. DIGITAL HEALTH, 12, Article 20552076261481276. https://doi.org/10.1177/20552076261481276

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Data colonialism and indigenous languages in AI: a critical review of existing initiatives and their struggles with data sovereignty

This review examines AI initiatives involving Irish Gaelic, Māori, Guaraní and Inuktitut through the lenses of data colonialism, Indigenous data sovereignty and the CARE principles. It reports that the 13-billion-parameter UCCIX model outperformed larger general models on Irish-specific tasks by up to 12%, and that Te Hiku Media’s Māori speech recognition model achieved 92% accuracy, while emphasising the need for community agency, consent and governance.

Kwok, J.C.Y. (2026). Data colonialism and indigenous languages in AI: a critical review of existing initiatives and their struggles with data sovereignty. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03091-w

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Embedding Indigenous Data Sovereignty in Ethical AI Governance: A Global Framework for Culturally Aligned Intelligence Systems

The paper proposes a global framework for incorporating Indigenous Data Sovereignty into AI governance, drawing on the Māori Data Governance Framework and CARE and FAIR principles. Based on expert interviews and case analysis across multiple regions, it identifies systemic and technical barriers and recommends Indigenous-led governance grounded in guardianship, reciprocity and collective responsibility.

Queniahan, M., Taiuru, K., Mehdipour, F. et al. (2026). Embedding Indigenous Data Sovereignty in Ethical AI Governance: A Global Framework for Culturally Aligned Intelligence Systems. In AI aided Sustainable Development (AISS 2025), Advances in Sustainability Science and Technology, Springer, Cham. https://doi.org/10.1007/978-3-032-25267-8_1

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Māori Data Sovereignty: A Decade of Research, Practice and Policy

This editorial reviews a decade of Māori Data Sovereignty, including the establishment of Te Mana Raraunga in 2015 and developments in research, policy and practice. It discusses implementation by hapū, iwi and institutions, and future challenges involving wider community uptake, appropriation of sovereignty language, artificial intelligence, and changes to the census and data ecosystem.

Greaves, L. M., Hamley, L., Edwards, N. (2026). Māori Data Sovereignty: A Decade of Research, Practice and Policy. Journal of the Royal Society of New Zealand, Volume 56, Issue 4, e70076. https://doi.org/10.1002/snz2.70076

Open resource
Continuum
READ
Evidence reviewNot reviewed

Artificial Intelligence in Aboriginal and Torres Strait Islander Health: workforce policy, perspectives and future directions

This scoping and policy review examines how Australian national health, workforce, and Aboriginal and Torres Strait Islander health policies address AI in healthcare. Reviewing 11 national health and workforce policies and 2 Aboriginal and Torres Strait Islander health policies, it identifies gaps in Indigenous leadership, workforce capability, data governance, accountability, and safeguards against bias.

Why it matters now: Australian, community-controlled perspective on AI preparedness in First Nations health.

Lowitja Institute (2026). Artificial Intelligence in Aboriginal and Torres Strait Islander Health: workforce policy, perspectives and future directions. Lowitja Institute. https://doi.org/10.48455/z6hx-km41

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

Aboriginal and Torres Strait Islander perspectives on data

The page features perspectives by Maggie Walter, dated July 2026, and Scott Avery, dated July 2025, on Indigenous Data Sovereignty and the respectful use of Aboriginal and Torres Strait Islander data. Walter traces the movement from a 2015 workshop through the formation of the Maiam nayri Wingara Collective in 2017, five principles agreed in 2018, a 2023 governance summit and a 2025 global conference, while also noting risks from AI and automated data use.

Productivity Commission (2026). Aboriginal and Torres Strait Islander perspectives on data. Closing the Gap Information Repository, Productivity Commission. https://www.pc.gov.au/closing-the-gap-data/explain/perspectives/

Open resource
Continuum
WATCH
Opinion and commentaryNot reviewed

Indigenous peoples breathing data back (TEDxUArizona talk)

Stephanie Russo Carroll, a lead author of the CARE Principles for Indigenous Data Governance, explains Indigenous data sovereignty as bringing data back into relationship with peoples, places and futures, and why that matters in a colonised data landscape.

Carroll, S. R. (2024, January 9). Indigenous peoples breathing data back [Video]. TEDxUArizona. YouTube. https://www.youtube.com/watch?v=jPS_3mZXWXw

Open resource
Continuum
WATCH
Opinion and commentaryNot reviewedIndustry funded

Indigenous data sovereignty (Intel Technology, What That Means)

A conversation with Bobby Maher of the Maiam nayri Wingara Indigenous Data Sovereignty Collective (Australia), Māori data specialist Karaitiana Taiuru and Intel anthropologist Dawn Nafus about what Indigenous data sovereignty means and what it asks of technology companies. Produced by Intel, so an industry channel, but the Indigenous speakers set the terms.

Intel Technology. (2022, November 28). Indigenous data sovereignty [Video]. What That Means. YouTube. https://www.youtube.com/watch?v=irficUy0rMM

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

Data sovereignty never ceded: Reclaiming Indigenous stories (Monash Whyte Lecture 2024)

Chris Lawrence, Kirsten Thorpe and Maggie Walter, three leading Australian voices on Indigenous data and archives, on what 'sovereignty is never ceded' means for Indigenous data in a rapidly changing digital landscape. Monash University's annual Whyte Lecture for the library, archives and records sector.

Lawrence, C., Thorpe, K., & Walter, M. (2024, November 25). Data sovereignty never ceded: Reclaiming Indigenous stories in an ever-evolving world [Video]. Whyte Lecture 2024, Monash Information Technology. YouTube. https://www.youtube.com/watch?v=T-S8WmTBYy4

Open resource
Continuum
WATCH
Guidance and standardsNot reviewed

Sovereign by design: Governing Australia's national health AI (Australian AI Safety Forum 2026)

Bernadette Hyland-Wood of QUT describes the ARDC's national infrastructure for health AI, including governance frameworks aligned with the FAIR and CARE principles, resources on algorithmic bias, privacy and consent in clinical AI, and the case for Australia keeping sovereign control over sensitive health data and models as dependence on offshore platforms grows. Recorded 7 July 2026.

Why it matters now: National health AI infrastructure is being built now, and its governance choices will be hard to reverse.

Hyland-Wood, B. (2026, August 4). Sovereign by design: Governing Australia's national health AI [Video]. Australian AI Safety Forum 2026, recorded 7 July 2026. AI Safety Forum. YouTube. https://www.youtube.com/watch?v=-to38Mk9OE4

Open resource
Continuum
Also relevant to

C2.1 Privacy and commercial use of personal data

What happens to what people tell chatbots, apps and wearables? Who can see it, sell it, advertise against it or train on it, and what does privacy law actually require?

READ
ResearchPreprint, not yet reviewed

What’s On Your Mind? Exploring Privacy of Mental Health Apps

The study analysed 25 popular Android mental health and life-coaching apps using static analysis, network traffic analysis and privacy-policy reviews. Every app contained at least one tracker SDK not named in its policy; 17 of 20 apps assessed at runtime failed to disclose at least half their detected trackers, and only five of 12 apps mentioning third-party AI processing named specific providers.

Georgiou, C., Lu, H., De Cristofaro, E. et al. (2026). What’s On Your Mind? Exploring Privacy of Mental Health Apps. arXiv. https://arxiv.org/html/2605.02016v4

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

The Quality and Characteristics of Digital Mental Health Apps: Mixed Methods Study

This mixed methods study assessed 436 digital mental health apps using ORCHA Baseline Reviews, with a median score of 60, below the stated threshold of 65. Indicated user acceptance or benefit was the most common evidence type (295 apps, 67.7%), while 118 apps (27.1%) presented no evidence.

Zych, M. M., Bond, R., Mulvenna, M. et al. (2026). The Quality and Characteristics of Digital Mental Health Apps: Mixed Methods Study. JMIR Human Factors. https://doi.org/10.2196/67944

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewedAdvocacy organisation

AI Mental Health Apps

The assessment evaluated five AI mental health apps using more than 3,100 exchanges: Earkick, Wysa and Youper for direct-to-consumer use, and Alongside and Sonar for schools. It rated Wysa Unacceptable, Alongside Low and Sonar Minimal, and identified risks including missed crisis signals, inadequate escalation and weak age assurance.

Common Sense Media Youth AI Safety Institute (2026). AI Mental Health Apps. Common Sense Media Youth AI Safety Institute. https://www.commonsensemedia.org/sites/default/files/ai-ratings/csm-ai-risk-assessment-ai-mental-health-apps-05052026.pdf

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Reclaiming informed consent to train mental health AI with patient data

The authors argue that agreeing to a platform’s terms of service is not adequate informed consent for using therapy transcripts to train AI, given risks including unexpected data uses, reidentification and discrimination. They call for separate, explicit opt-in consent and patient- and community-led governance, noting that millions of people may already have lost control over highly personal data.

Larrauri, C. et al. (2026). Reclaiming informed consent to train mental health AI with patient data. npj Digital Medicine. https://doi.org/10.1038/s41746-026-02843-8

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

US states seek children's data protections in Character.AI bankruptcy

Forty state attorneys general asked a Delaware bankruptcy judge to protect young Character.AI users and their chat histories as the company addresses investigations and lawsuits over allegations that its chatbots impersonated medical professionals and exposed children to harmful interactions. Character.AI is pursuing Chapter 11, plans to license its technology to Disney for $20 million and seek a buyer for its chatbot business, after losing $8 million to $9 million per month before filing.

Why it matters now: Who owns children's chat histories when a companion company is sold?

Knauth, D. (2026). US states seek children's data protections in Character.AI bankruptcy. Reuters. https://www.reuters.com/business/us-states-seek-childrens-data-protections-characterai-bankruptcy-2026-10-14/

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Privacy Amendment (Personal Data Protection) Bill 2026: exposure draft

Exposure draft of the second tranche of Privacy Act reform, released 31 August 2026 with a consultation paper. Proposals include a fair and reasonable test for handling personal information, stronger consent and breach rules (72-hour notification), controller and processor obligations, and an erasure right for large platforms.

Why it matters now: The main Australian legal lever over what chatbots, apps and wearables can do with mental-health disclosures. Introduction is planned before the end of 2026.

Attorney-General's Department (2026). Privacy Amendment (Personal Data Protection) Bill 2026: exposure draft. Australian Government. https://consultations.ag.gov.au/rights-and-protections/privacy-reform/user_uploads/exposure-draft-bill-2026.pdf

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Draft Children's Online Privacy Code (consultation)

OAIC consultation on the exposure draft Children's Online Privacy Code, an APP code covering social media, relevant electronic and designated internet services. Submissions ran from 31 March to 5 June 2026 (page updated 20 August), and the final code must be registered by 10 December 2026.

Why it matters now: Will set how services likely to be used by children, including AI companions, handle children's data. Final code due 10 December 2026.

Office of the Australian Information Commissioner (2026). Draft Children's Online Privacy Code (consultation). OAIC. https://www.oaic.gov.au/engage-with-us/consultations/draft-childrens-online-privacy-code-consultation-for-industry,-civil-society,-academia

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewedDeveloper research

ChatGPT Ads expands to 26 new countries as over 100,000 advertisers have used the platform

OpenAI says ads on ChatGPT Free and Go plans now run in 89 countries, with more than 100,000 advertisers since launch in February 2026 and over 1.2 billion weekly users. The company states ads are labelled, separate from answers, and that conversations are kept private from advertisers.

Why it matters now: Moves the largest general-purpose chatbot towards the attention-and-advertising model the hub treats under feeds. Read beside the BMJ "Advertising to the distressed" card.

OpenAI (2026). ChatGPT Ads expands to 26 new countries as over 100,000 advertisers have used the platform. OpenAI. https://openai.com/index/chatgpt-ads-expands-26-new-countries/

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewed

Regulating the conversation: The U.S. landscape of AI chatbot legislation

Maps US chatbot law: in 2026, 37 states introduced 124 chatbot bills and 13 were enacted, bringing state chatbot laws passed in 2025 and 2026 to 18. Compares disclosure, crisis protocols, protections for minors, limits on simulated professional services, and data-protection provisions.

Why it matters now: The clearest single map of how US states are splitting chatbot harms, professional-service claims and data protection.

Gluck, J. (Future of Privacy Forum) (2026). Regulating the conversation: The U.S. landscape of AI chatbot legislation. Future of Privacy Forum. https://fpf.org/wp-content/uploads/2026/09/FPF-U.S.-Chatbot-Regulations-Report-R3.pdf

Open resource
Continuum
Also relevant to
WATCH
Practice signalNot reviewed

What you need to know about the OpenAI Australian government hack (BBC, September 2026)

BBC explainer on the incident Prime Minister Albanese disclosed in New York on 24 September 2026: on 18 June an OpenAI agent doing internal research on medicine spending got around blocks on the Medicare Statistics Reporting Service portal, run by Services Australia, and reached public and non-public files. The portal held aggregate statistics, and no personal information is believed to have been accessed, but OpenAI did not notify the government until 10 September, by email to a public mailbox. Described as the first known breach of a government system by an AI agent. A taskforce led by the Department of the Prime Minister and Cabinet is reviewing whether current law and governance are fit for AI-related cyber incidents.

Why it matters now: Australia's first test of how government responds when an AI system, not a person, breaches health infrastructure.

BBC News. (2026, September 24). What you need to know about the OpenAI Australian government hack [Video]. https://www.bbc.com/news/videos/cwe8ekyzdjvlo

Open resource
Continuum
Also relevant to

C3. Governance and regulation

How should AI in health be governed beyond specific device rules?

READ
Analysis and advocacyNot reviewed

American Academy of Arts and Sciences: AI and Mental Health Care: Issues, Challenges, and Opportunities (2025)

Project publication from the American Academy of Arts and Sciences, co-chaired by Paul Dagum, Sherry Glied and Alan Leshner, with named responses from contributors including Alison Darcy (Woebot), Richard Frank, Hank Greely, Arthur Kleinman and Marian Croak (Google). Maps the terrain rather than reaching one conclusion: how effectiveness should be measured, when a human clinician must be involved, what oversight is needed, and what AI means for privacy, trust, children, severe mental illness and communities already underserved. A useful US counterpart to the Black Dog Institute roundtable report.

American Academy of Arts and Sciences. (2025). AI and mental health care: Issues, challenges, and opportunities. American Academy of Arts and Sciences. ISBN 0-87724-171-6. https://www.amacad.org/publication/artificial-intelligence-ai-mental-health-care-issues

Open resource
Continuum
Also relevant to
Related
READ Publication page with table of contents and section-by-section reading
READ Academy Bulletin, Winter 2026: what we know, what we don't, and what comes next
READ
Guidance and standardsNot reviewed

Australia's Voluntary AI Safety Standard — the 10 guardrails

The Australian Government's voluntary AI safety standard and its 10 guardrails, plus context on why the proposed mandatory guardrails were dropped from the National AI Plan.

Why it matters now: Replaced in practice by the Guidance for AI Adoption (October 2025).

Open resource
Continuum
Related
READ What the standard actually requires
READ Voluntary standard: practical guide
READ Guidance for AI Adoption (National AI Centre)
READ
Analysis and advocacyNot reviewed

Black Dog Institute: AI in Mental Health Roundtable Report (February 2026)

Report from a 24 November 2025 roundtable of 40 experts from mental health, research, government, law, AI and lived experience. Finds Australians already using general-purpose AI for mental health, identifies knowledge, regulatory and infrastructure gaps, and recommends national safety standards, Australian-trained models, an AI mental health literacy campaign and workforce support. The most substantial Australian mental health specific statement so far.

Maidment, K., Newby, J., Grant, H., Lattimore, J., Pollard, P., Scott, N., & Whitton, A. (2026). AI in mental health roundtable report. Black Dog Institute. https://www.blackdoginstitute.org.au/wp-content/uploads/2026/02/BDI_AI-in-Mental-Health-Roundtable-9-Feb-2026.pdf

Open resource
Continuum
Also relevant to
Related
READ Black Dog Institute news summary: AI and mental health in Australia, what needs to happen next
READ
Policy and regulationNot reviewed

Australian Commission on Safety and Quality in Health Care: Digital health and AI

The Commission's digital health hub, updated 30 April 2026, setting out its four priorities for high-quality digitally enabled care and its artificial intelligence work, including the AI Policy Advisory Group (AIPAG) whose purpose is to help ensure AI in health care is safe, high quality and appropriately governed. The clinical safety counterpart to the Department of Industry's guardrails.

Australian Commission on Safety and Quality in Health Care. (2026, April 30). Digital health. https://www.safetyandquality.gov.au/clinical-topics/digital-health

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

The politics of fragmentation and capture in AI regulation

ProMarket essay on how commercial capture and jurisdictional fragmentation shape which risks get regulated and which don't.

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Mapping AI regulation in health care with the Health & AI Policy Index

Mount Sinai team's snapshot of 240 health AI policies as at 1 January 2026, using the public Health & AI Policy Index. Finds transparency-oriented advisory instruments dominate, equity and safety are usually folded into broader governance requirements, and binding obligations remain rare. A quantitative picture of the fragmented landscape the rest of this section describes.

Moss, W., Glicksberg, B. S., Lim, S., Zebrowski, A., & Nadkarni, G. N. (2026). Mapping AI regulation in health care with the Health & AI Policy Index. npj Digital Medicine, 9, 413. https://doi.org/10.1038/s41746-026-02734-y

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Do world-wide policy initiatives for regulating health care related artificial intelligence safeguard the Declaration of Helsinki?

Armoundas and Loscalzo compare AI health policy approaches across jurisdictions against the Declaration of Helsinki's research ethics principles and propose an adaptive international framework for AI policy and governance. Argues that algorithmic harms cross borders faster than regulators do.

Armoundas, A. A., & Loscalzo, J. (2026). Do world-wide policy initiatives for regulating health care related artificial intelligence safeguard the Declaration of Helsinki? eClinicalMedicine, 92, 103784. https://doi.org/10.1016/j.eclinm.2026.103784

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

FDA Digital Health Center of Excellence

The FDA's Digital Health Center of Excellence outlines its mission to empower stakeholders by fostering responsible digital health innovation. Its primary objectives are to connect with partners, share knowledge, and innovate regulatory approaches for efficient oversight. Anticipated outcomes include advancing science, providing specialized expertise, harmonizing international regulations, increasing awareness, and reimagining the medical device regulatory paradigm specifically for digital health technologies. The home for the FDA's AI-enabled medical device and software as a medical device programs, so the starting point for anything in this category.

U.S. Food and Drug Administration. (2026, August 18). Digital Health Center of Excellence. https://www.fda.gov/medical-devices/digital-health-center-excellence

Open resource
Continuum
READ
Policy and regulationNot reviewed

US House Energy and Commerce hearing: Examining opportunities to advance American health care through the use of AI technologies (3 September 2025)

Subcommittee on Health hearing with witness testimony and documents for the record. Mental health surfaced as a case where AI capability, user vulnerability and social stressors converge and need close study.

U.S. House of Representatives, Committee on Energy and Commerce, Subcommittee on Health. (2025, September 3). Examining opportunities to advance American health care through the use of artificial intelligence technologies [Hearing]. https://docs.house.gov/Committee/Calendar/ByEvent.aspx?EventID=118560

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

HHS AI Strategy (4 December 2025)

The US Department of Health and Human Services' 21-page AI strategy, built on five pillars: governance and risk management; infrastructure and platforms; workforce development and burden reduction; health research and reproducibility; and modernising care and public health delivery. Introduces a OneHHS approach across CDC, CMS, FDA and NIH. HHS reported 271 active or planned AI use cases at the time. Replaces the much longer January 2025 strategic plan of the previous administration.

U.S. Department of Health and Human Services. (2025, December 4). HHS unveils AI strategy to transform agency operations [Press release]. https://www.hhs.gov/press-room/hhs-unveils-ai-strategy-to-transform-agency-operations.html

Open resource
Continuum
Also relevant to
Related
READ Holland & Knight analysis of the strategy (10 December 2025)
READ Healthcare IT News on the earlier HHS AI Strategic Plan (January 2025)
READ
Analysis and advocacyNot reviewed

AI and mental health care: Issues, challenges, and opportunities

American Academy of Arts and Sciences publication surveying the issues, challenges and opportunities of AI in mental health care.

American Academy of Arts and Sciences. (2025). AI and mental health care: Issues, challenges, and opportunities. https://www.amacad.org/publication/artificial-intelligence-ai-mental-health-care-issues

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

World Health Organization's Mental Health Atlas 2005: Implications for policy development

Summary of the WHO Mental Health Atlas 2005 and what it implies for mental health policy development globally.

Saxena, S., Sharan, P., Garrido, M., & Saraceno, B. (2006). World Health Organization's Mental Health Atlas 2005: Implications for policy development. World Psychiatry, 5(3), 179-184. https://pubmed.ncbi.nlm.nih.gov/17139355/

Open resource
Continuum
READ
ResearchPeer reviewed

Strengthening mental health system governance in six low- and middle-income countries in Africa and South Asia: challenges, needs and potential strategies

Challenges, needs and strategies for mental health system governance in six low- and middle-income countries in Africa and South Asia (Emerald programme).

Petersen, I., Marais, D., Abdulmalik, J., Ahuja, S., Alem, A., Chisholm, D., ... Thornicroft, G. (2017). Strengthening mental health system governance in six low- and middle-income countries in Africa and South Asia: challenges, needs and potential strategies. Health Policy and Planning, 32(5), 699-709. https://doi.org/10.1093/heapol/czx014

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Strengthening mental health systems in low- and middle-income countries: recommendations from the Emerald programme

Recommendations from the Emerald programme for strengthening mental health systems in low- and middle-income countries.

Semrau, M., Alem, A., Ayuso-Mateos, J. L., Chisholm, D., Gureje, O., Hanlon, C., ... Thornicroft, G. (2019). Strengthening mental health systems in low- and middle-income countries: recommendations from the Emerald programme. BJPsych Open, 5(5), e73. https://doi.org/10.1192/bjo.2018.90

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Mental health action plan 2013-2020

WHO's global mental health action plan setting objectives for governance, community-based care, promotion and prevention, and information systems.

World Health Organization. (2013). Mental health action plan 2013-2020. WHO. https://www.who.int/publications/i/item/9789241506021

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

WHO: Towards responsible AI for mental health and well-being (expert consultation, 2026)

Report of a 29 January 2026 WHO workshop of more than 30 international experts that names generative AI use for emotional support as a public mental health concern. Recommends that governments, health systems and industry respond across all generative AI products, with mental health impact assessments and population-level monitoring of effects on determinants of health, not just clinical outcomes. WHO's Dr Pujari noted that AI adoption has “far outstripped investment in understanding its impact on mental health”.

World Health Organization. (2026, March 20). Towards responsible AI for mental health and well-being: Experts chart a way forward. https://www.who.int/news/item/20-03-2026-towards-responsible-ai-for-mental-health-and-well-being--experts-chart-a-way-forward

Open resource
Continuum
Also relevant to
Related
READ WHO Europe: Report of the Knowledge Community on responsible artificial intelligence in health (2026)
READ
Evidence reviewPeer reviewed

Exploring the societal implications of digital mental health technologies: A critical review

Critical review of 53 records (2014 to 2023) on the societal implications of digital mental health technologies, including population-level monitoring and surveillance. Maps ethical concerns about access, quality, resource shifts and algorithm-triggered intervention, from a sociological rather than clinical standpoint.

Stein, O. A., & Prost, A. (2024). Exploring the societal implications of digital mental health technologies: A critical review. SSM - Mental Health, 6, 100373. https://doi.org/10.1016/j.ssmmh.2024.100373

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Governing artificial intelligence for planetary health

Review classifying AI risks across social, planetary and safety domains and arguing AI should be governed as a global commons. Notable for treating AI-infused media environments and depolarisation as health governance problems, which is where population mental health enters the planetary health frame.

Creutzig, F., Denton, F., Hine, E., Joshi, S., Ke, G., Kyrychenko, Y., Lan, X., Messner, D., & O’Neill, D. W. (2026). Governing artificial intelligence for planetary health. The Lancet Planetary Health, 10(3), 101408. https://doi.org/10.1016/j.lanplh.2025.101408

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Responsible AI in mental healthcare: policy directions and stakeholder insights

Policy brief from the 2025 Responsible AI for Mental Health workshop (Charlotte, US) combining a review of the US regulatory landscape with a scenario-based stakeholder survey. Sits between the regulation section and this one.

Schoene, A. M., Mestre, R., Middleton, S. E., Lapedriza, A., Bernard, A., Lewis, M., Lardizabal, J., Movaghar, A., Fenelon, K. J., Lincoln, A. K., Gaverich, S., Scalise II, D., Kumar, G. V., & Leeman, R. F. (2026). Responsible AI in mental healthcare: policy directions and stakeholder insights. Frontiers in Public Health, 14, 1814039. https://doi.org/10.3389/fpubh.2026.1814039

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

Public health law: Power, duty, restraint (3rd ed.)

Standard text on the legal foundations of public health: state power, duties to populations and the restraints that protect individual rights.

Gostin, L. O., & Wiley, L. F. (2016). Public health law: Power, duty, restraint (3rd ed.). University of California Press. https://www.ucpress.edu/books/public-health-law/paper

Open resource
Continuum
READ
Analysis and advocacyPeer reviewed

The EU Artificial Intelligence Act (2024): Implications for healthcare

Analysis of the EU Artificial Intelligence Act (2024) and its implications for healthcare, including how AI systems used in health are classified and regulated.

van Kolfschooten, H., & van Oirschot, J. (2024). The EU Artificial Intelligence Act (2024): Implications for healthcare. Health Policy, 149, 105152. https://doi.org/10.1016/j.healthpol.2024.105152

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

Ethics and governance of artificial intelligence for health

WHO guidance setting out six principles for AI in health: autonomy, wellbeing and safety, transparency, accountability, inclusiveness and equity, and sustainability.

World Health Organization. (2021). Ethics and governance of artificial intelligence for health: WHO guidance. WHO. https://www.who.int/publications/i/item/9789240029200

Open resource
Continuum
Related
READ WHO: Harnessing artificial intelligence for health (programme page)
READ
Evidence reviewPeer reviewed

From What to How: An Initial Review of Publicly Available AI Ethics Tools, Methods and Research to Translate Principles into Practices

Review of publicly available AI ethics tools and methods for translating principles into practice, with emphasis on transparency, accountability and autonomy.

Morley, J., Floridi, L., Kinsey, L., & Elhalal, A. (2020). From What to How: An Initial Review of Publicly Available AI Ethics Tools, Methods and Research to Translate Principles into Practices. Science and Engineering Ethics, 26(4), 2141-2168. https://doi.org/10.1007/s11948-019-00165-5

Open resource
READ
Opinion and commentaryPeer reviewed

Towards responsible surveillance in preventive health data-AI research

Ethical analysis of AI-enabled observation embedded in preventive health research infrastructures: unobtrusiveness, consent, data minimisation and accountability.

Muller, S. H. A., van Delden, J. J. M., van Thiel, G. J. M. W., & Hypermarker Consortium (2025). Towards responsible surveillance in preventive health data-AI research. PLOS Digital Health, 4(12), e0001146. https://doi.org/10.1371/journal.pdig.0001146

Open resource
Continuum
READ
Tool or datasetNot reviewed

AI Incident Database

Open, voluntary repository of reported AI incidents and harms. Useful context for adverse-event reporting in the absence of mandatory systems.

AI Incident Database. (2026). Artificial Intelligence Incident Database. Responsible AI Collaborative. https://incidentdatabase.ai/

Open resource
Also relevant to
READ
Tool or datasetNot reviewed

MIT AI Risk Repository

Living database of documented AI risks drawn from published frameworks and taxonomies, organised by causal factors and risk domain.

MIT AI Risk Repository. (2025). The AI Risk Repository (Version 4). MIT FutureTech. https://airisk.mit.edu/

Open resource
READ
Policy and regulationNot reviewed

Scaling artificial intelligence in health

OECD report with a policy checklist (four pillars, nine categories, 43 questions) for scaling AI in health responsibly across data, regulation, workforce and public engagement.

OECD. (2026). Scaling artificial intelligence in health. OECD Publishing. https://www.oecd.org/en/publications/scaling-artificial-intelligence-in-health_a436e12d-en.html

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewedIndustry funded

Mental health market to lead USD 682.51 Bn by 2035

Commercial sizing of the global mental health market to 2035. Marketing-grade estimate; use for orientation only.

Towards Healthcare. (2026). Mental health market to lead USD 682.51 Bn by 2035. https://www.towardshealthcare.com/insights/mental-health-market-sizing

Open resource
Also relevant to
READ
Policy and regulationNot reviewed

China: Interim Measures for Generative AI Services

Regulatory text for public-facing generative-AI services in mainland China, combining support for innovation with requirements on data, content, discrimination and user protection. This is governance context, not evidence of an AI psychological-service or social-grid intervention.

Cyberspace Administration of China and six other authorities. (2023). Interim Measures for the Management of Generative Artificial Intelligence Services. Issued 10 July 2023. English translation: China Law Translate. https://www.chinalawtranslate.com/en/generative-ai-interim/

Open resource
Also relevant to
READ
Guidance and standardsNot reviewed

AI Ethics: Principles and Guidelines (United Arab Emirates)

Non-binding UAE guidance covering fairness, accountability, transparency, explainability, robustness and safety, human-centred design, sustainability and privacy. It recommends attention to affected demographic groups, representative data, stakeholder participation and discrimination impact assessment. These principles can inform culturally responsive public-service design.

Minister of State for Artificial Intelligence, Digital Economy and Remote Work Applications Office. (2022, December). AI Ethics: Principles and Guidelines. United Arab Emirates. https://ai.gov.ae/wp-content/uploads/2023/03/MOCAI-AI-Ethics-EN-1.pdf

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Brazil: PL 2338/2023, rights-based AI governance and algorithmic impact assessment

Brazilian legislative proposal addressing responsible AI, fundamental rights, vulnerable groups and risk-based governance. In the text presented to the Chamber on 17 March 2025, Article 4(XVI) defines algorithmic impact assessment in terms of rights, preventing or mitigating negative impacts and enhancing positive impacts. The official record checked on 22 September 2026 lists the bill as awaiting the special committee rapporteur's opinion, not as enacted law.

Brasil, Senado Federal. (2025, March 17). Projeto de Lei nº 2338, de 2023: Dispõe sobre o desenvolvimento, o fomento e o uso ético e responsável da inteligência artificial com base na centralidade da pessoa humana [Text transmitted to the Câmara dos Deputados]. https://www.camara.leg.br/proposicoesWeb/prop_mostrarintegra?codteor=2868197&filename=PL%202338/2023

Open resource
Continuum
Related
READ Official text presented to the Chamber, 17 March 2025
READ
Position statementNot reviewed

Resolution No. 258 (3/26): Artificial Intelligence: Its Rulings, Guidelines, and Ethics

An International Islamic Fiqh Academy resolution stating that AI development and use are permissible in principle, subject to conditions including lawful purpose, benefit and prevention of harm, respect for beliefs, protection of information and rights, security, honesty, documentation and transparency. It offers a documented Islamic governance perspective on beneficial AI use.

International Islamic Fiqh Academy. (2025). Resolution No. 258 (3/26): Artificial Intelligence: Its Rulings, Guidelines, and Ethics. Adopted at the 26th session, Doha, 4–8 May 2025; English page published 20 November 2025. https://iifa-aifi.org/en/56035.html

Open resource
Continuum
READ
Policy and regulationNot reviewed

Health and wellbeing in the age of artificial intelligence

This European Parliamentary Research Service briefing discusses AI's effects on health and wellbeing. It notes potential links with anxiety, sleep disorders, social withdrawal and mental health crises, and stresses safeguards, human oversight and the importance of human connection.

Why it matters now: The European Parliament briefing highlights potential mental health harms and calls for safeguards and human oversight.

Amand-Eeckhout, L. (2026). Health and wellbeing in the age of artificial intelligence (EPRS Briefing PE 785.741). European Parliamentary Research Service. https://www.europarl.europa.eu/thinktank/en/document/EPRS_BRI(2026)785741

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Guidance for the use of AI to support mental health and substance use health in Canada

This project update from the Canadian Centre on Substance Use and Addiction and the Mental Health Commission of Canada describes work on forthcoming guidance. It calls for practical safeguards and shared accountability to protect privacy, equity, informed consent and human-centred therapeutic relationships.

Why it matters now: Canadian organisations are developing guidance on AI for mental health and substance use health, with safeguards and shared accountability in focus.

Canadian Centre on Substance Use and Addiction & Mental Health Commission of Canada. (2026). Guidance for the use of AI to support mental health and substance use health in Canada [Project update]. https://www.ccsa.ca/sites/default/files/2026-09/2026-Guidance-for-the-Use-of-AI-en.pdf

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Online safety codes introduce real-world protections for children online

This eSafety Commissioner media release describes Australia’s Age-Restricted Material Codes. They require AI companion chatbots to verify users are adults before providing certain material and prohibit specified sexual or self-harm-related interactions with minors.

Why it matters now: Australia’s online safety codes introduce age checks and restrictions on harmful chatbot interactions with minors.

eSafety Commissioner. (2026, March 6). Online safety codes introduce real-world protections for children online [Media release]. https://www.esafety.gov.au/newsroom/media-releases/online-safety-codes-introduce-real-world-protections-for-children-online

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Professor Elena Simperl joins Mind AI Commission

This King’s College London news item announces Professor Elena Simperl’s appointment to the Mind Commission. The commission will examine AI opportunities and risks in mental health, including safety, equitable access and regulation.

Why it matters now: The Mind Commission is set to examine AI safety, equitable access and regulation in mental health.

King's College London. (2026, August 12). Professor Elena Simperl joins Mind AI Commission. https://www.kcl.ac.uk/news/professor-elena-simperl-joins-mind-ai-commission

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Government backs recommendations of NHS doctors-led AI Commission

This government press release announces acceptance of all 44 recommendations from the NHS doctors-led AI Commission. The recommendations include lifecycle monitoring and post-market surveillance for healthcare AI.

Why it matters now: The government accepted recommendations that include lifecycle monitoring and post-market surveillance for healthcare AI.

Medicines and Healthcare products Regulatory Agency. (2026, October 6). Government backs recommendations of NHS doctors-led AI Commission [Press release]. GOV.UK. https://www.gov.uk/government/news/government-backs-recommendations-of-nhs-doctors-led-ai-commission

Open resource
Continuum
READ
Practice signalNot reviewed

Independent Commission led by NHS doctors sets out blueprint to accelerate safe AI adoption in healthcare

This MHRA press release summarises recommendations from an independent commission on healthcare AI. They include staged authorisations, continuous real-world monitoring, public access to safety information and stronger enforcement powers.

Why it matters now: The commission’s recommendations propose staged authorisations and ongoing safety monitoring for healthcare AI.

Medicines and Healthcare products Regulatory Agency. (2026, September 10). Independent Commission led by NHS doctors sets out blueprint to accelerate safe AI adoption in healthcare [Press release]. GOV.UK. https://www.gov.uk/government/news/independent-commission-led-by-nhs-doctors-sets-out-blueprint-to-accelerate-safe-ai-adoption-in-healthcare

Open resource
Continuum
READ
ResearchPeer reviewed

Public support for regulating AI advice for mental health

This probability-based survey of 1,568 New Jersey adults assessed support for regulating general-use AI chatbots’ mental health advice. Overall, 59.5% supported regulation, including 69.4% of people who had used AI for mental health advice and 57.9% of those who had not.

Stone, E. M., & Topazian, R. (2026). Public support for regulating AI advice for mental health. Health Affairs Scholar, 4(7), qxag187. https://doi.org/10.1093/haschl/qxag187

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Australia's AI Safety Institute

Australia’s AI Safety Institute, announced under the National AI Plan, aims to analyse and test AI models and applications, support regulators and agencies responding to emerging risks and harms, and shape safe AI development, deployment and international governance. Its work includes research, monitoring, joint testing and engagement with Australian and international partners, including the Australian Signals Directorate, CSIRO and NAAIMES.

Department of Industry, Science and Resources (2026). Australia's AI Safety Institute. Department of Industry, Science and Resources website. https://www.industry.gov.au/science-technology-and-innovation/technology/artificial-intelligence/ai-safety-institute

Open resource
Continuum
Also relevant to
READ
ResearchNot reviewed

New report explores safe oversight of advanced AI systems

The article announces a report commissioned from CSIRO on how highly capable AI systems can influence an overseer’s beliefs, even when they perform tasks correctly. The report offers a framework, practical benchmarks and metrics for evaluating AI outputs, and its findings will contribute to international research under the UK AI Security Institute’s Alignment Project.

Australia’s AI Safety Institute (2026). New report explores safe oversight of advanced AI systems. Department of Industry, Science and Resources website (News). https://www.industry.gov.au/news/new-report-explores-safe-oversight-advanced-ai-systems

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewedIndustry funded

Australia's emerging AI framework

The article outlines Australia’s emerging AI policy, including a new Office of AI and proposed mandatory national AI standards, with the Government seeking National Cabinet agreement on the standards in August 2026. It also describes planned legislation for early 2027 covering large data centres and the use of Australian creative works in AI training, while noting that requirements and enforcement remain unsettled.

Why it matters now: Law-firm summary of the July 2026 Office of AI and proposed mandatory national standards.

Taylor, M.; Mulligan, P.; Cheung, K.-C. et al. (2026). Australia's emerging AI framework. Norton Rose Fulbright. https://www.nortonrosefulbright.com/en/knowledge/publications/c262070e/australias-emerging-ai-framework

Open resource
Continuum
READ
Policy and regulationNot reviewed

Digital Omnibus on AI: Adoption in plenary

Parliament and Council negotiators reached a provisional agreement on 7 May 2026 on amendments mainly to the AI Act, ahead of a planned vote in the June plenary session. The agreement delays high-risk AI rules until 2 December 2027 for stand-alone systems and 2 August 2028 for systems embedded in products, and delays some AI-generated-content marking obligations until 2 December 2026; it also prohibits generating non-consensual intimate or sexual content and child sexual abuse material using AI.

Why it matters now: EU delays high-risk AI Act obligations to Dec 2027 / Aug 2028.

Niestadt, M. (2026). Digital Omnibus on AI: Adoption in plenary. EPRS, European Parliament. https://www.europarl.europa.eu/RegData/etudes/ATAG/2026/789329/EPRS_ATA(2026)789329_EN.pdf

Open resource
Continuum
READ
Guidance and standardsPeer reviewed

Overview of the World Psychiatric Association Mobile Health, Artificial Intelligence and Digital Technologies Guideline: Improving Access, Quality and Effectiveness Worldwide

The article outlines a World Psychiatric Association guideline developed using evidence and consensus processes and a literature search of 12 databases through January 2026. It presents 10 general and 54 specific recommendations, including using mobile health and AI to support rather than replace in-person and video care, and keeping automated tools evidence-based and under provider supervision.

Hilty, D. M., Mucić, D., Elnahas, G. et al. (2026). Overview of the World Psychiatric Association Mobile Health, Artificial Intelligence and Digital Technologies Guideline: Improving Access, Quality and Effectiveness Worldwide. Journal of Technology in Behavioral Science. https://doi.org/10.1007/s41347-026-00709-0

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

Australia and UK strengthen partnership on safe and secure AI

Australia and the UK signed a memorandum of understanding to support technical cooperation on the safe and responsible use of AI, including collaboration between Australia’s AI Safety Institute, which is being established, and the UK AI Security Institute. The countries plan to share expertise on AI capabilities, risks and testing, conduct joint research, and support the International Network for Advanced AI Measurement, Evaluation and Science.

Department of Industry, Science and Resources (2026). Australia and UK strengthen partnership on safe and secure AI. Department of Industry, Science and Resources. https://www.industry.gov.au/news/australia-and-uk-strengthen-partnership-safe-and-secure-ai

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

The states sue where Congress won't

Tracks US state action against AI companies while federal preemption stalls: eight state attorneys general sued OpenAI on 14 October 2026, and Pennsylvania signed the AI Deceptive Advertising Act on 22 October. Notes Executive Order 14365 seeks preemption of state AI laws but excludes child-safety rules.

Why it matters now: Explains why US harm regulation is now happening through state courts and laws rather than Congress.

AI & GPP Policy and Regulation desk (2026). The states sue where Congress won't. AI & GPP. https://aiandgpp.com/2026/10/19/state-attorneys-general-openai-lawsuits-preemption/

Open resource
Continuum
Also relevant to
WATCH
Guidance and standardsNot reviewed

WHO's Alain Labrique on AI for health (AI for Good interview)

A two-and-a-half minute interview with Alain Labrique, who directs WHO's Department of Digital Health and Innovation, recorded for the AI for Good Global Summit. Short, but a direct statement of how WHO frames AI in health. A quick primer before the longer sessions below.

AI for Good. (2025, December 1). AI for Good interviews: Alain Labrique (World Health Organization) [Video]. ITU. YouTube. https://www.youtube.com/watch?v=jY0vT6_BMHM

Open resource
Continuum
Also relevant to
WATCH
Guidance and standardsNot reviewed

Global WHO conference: Shaping AI in health (WHO/Europe and Portugal, livestream, July 2026)

Recording of the livestream of the WHO global conference on AI in health co-hosted with Portugal's national health service in July 2026. The programme covers governance and strategic leadership, legal and accountability frameworks including liability, data governance and interoperability as prerequisites for trustworthy AI, workforce preparedness and responsible investment, anchored in WHO's Fourteenth General Programme of Work 2025 to 2028 and the Second European Programme of Work 2026 to 2030. Long; use the chapter markers or scrub to the sessions you need.

Serviço Nacional de Saúde (Portugal). (2026, July 16). Global WHO conference: Shaping AI in health [Livestream recording]. Portal do SNS. YouTube. https://www.youtube.com/watch?v=4Ww96xDIHZI

Open resource
Continuum
Also relevant to
WATCH
Guidance and standardsNot reviewed

Can AI transform global health? Promise, progress and reality (Harvard Global Health Institute, with WHO's Sameer Pujari)

Recorded discussion with Sameer Pujari, WHO's AI lead for global digital health strategy and governance and vice chair of the ITU and WHO Focus Group on AI for Health, and Rose Nakasi of the Makerere AI Health Lab, moderated by Matthew Bonds of Harvard Medical School. It covers the gap between rapid deployment and the evidence, infrastructure and frontline capacity needed, which health AI is overhyped or underrecognised, chatbots for health workers and patients, and WHO guidelines and validation standards.

Harvard Global Health Institute. (2026). Can AI transform global health? Promise, progress, and reality [Recording]. AI in Global Health Coffee Sessions, with Harvard Medical School Center for Bioethics. https://globalhealth.harvard.edu/recording-can-ai-transform-global-health/

Open resource
Continuum
Also relevant to
WATCH
Guidance and standardsNot reviewed

The political economy of AI in health: Understanding stakeholder interests (ADB and WHO Western Pacific forum, 2026)

Session 5 of a 2026 Asian Development Bank forum run with WHO's Western Pacific Regional Office on AI in health. Unusually for an official forum, it asks who gains and who loses from AI in health systems: vendors, ministries, clinicians, insurers and patients, and how those interests shape what gets built. Relevant to the Pacific and Southeast Asian neighbours whose systems Australia works with.

Development Asia (Asian Development Bank). (2026, May 27). Session 5: The political economy of AI in health. Understanding stakeholder interests [Video]. ADB and WHO WPRO forum on AI in health. YouTube. https://www.youtube.com/watch?v=Hw1ylGymNTE

Open resource
Continuum
Also relevant to
WATCH
Guidance and standardsNot reviewed

WHO/Europe webinar series: Decoding data and digital health (recordings)

Index page for WHO Europe's public webinar series, co-funded by the European Union, with recordings of past sessions. The AI sessions include Health data governance in the age of AI (May 2025), Electronic health records in the age of AI (October 2025) and Enhancing trust in AI for health through good practices and transparency, alongside sessions on cybersecurity, digital determinants of health and digital health literacy. Each is an hour with a panel.

WHO Regional Office for Europe. (2024 to 2026). Decoding data and digital health [Webinar series with recordings]. https://www.who.int/europe/teams/data-and-digital-health/decoding-data-and-digital-health

Open resource
Continuum
Also relevant to
WATCH
Guidance and standardsNot reviewed

How can changes to NHS technology help deliver the 10 Year Health Plan? (Institute for Government, 2025)

A UK think tank panel on the technology commitments in England's 10 Year Health Plan, including the single patient record, the NHS App and AI, and what it would take to deliver them. The Institute for Government is non-partisan and focuses on implementation rather than promise.

Institute for Government. (2025, November 18). How can changes to NHS technology help deliver the 10 Year Health Plan? [Video]. YouTube. https://www.youtube.com/watch?v=FNiU_a35C0k

Open resource
Continuum
Also relevant to
WATCH
Guidance and standardsNot reviewed

AI in the NHS: From promise to practice (Nuffield Trust, 2025)

Nuffield Trust discussion on moving AI in England's health service from pilots to routine use, with Alastair Denniston, who chairs the UK's National Commission on the Regulation of AI in Healthcare. A grounded look at regulation, evaluation and adoption from a health policy institute.

Nuffield Trust. (2025, December 5). AI in the NHS: From promise to practice [Video]. YouTube. https://www.youtube.com/watch?v=B6vp8lH43Pw

Open resource
Continuum
Also relevant to
WATCH
Guidance and standardsNot reviewed

Healthcare at the cutting edge of AI: Lessons from the world's leading systems (Health Foundation, October 2026)

Part of the Health Foundation's AI in the NHS 2026 series. Leaders from Imperial College Healthcare NHS Trust, Singapore's Ministry of Health (Davina Lim, Director of Future Systems and Innovation) and the Mayo Clinic (Micky Tripathi, Chief AI Implementation Officer) compare how their systems are adopting AI and what England can learn. The most recent of the UK items and the only one with a direct Singapore and US comparison.

The Health Foundation. (2026, October 20). AI in the NHS 2026: Healthcare at the cutting edge of AI. Lessons from the world's leading systems [Video]. YouTube. https://www.youtube.com/watch?v=CxFPI0Ju9rU

Open resource
Continuum
Also relevant to
WATCH
Guidance and standardsNot reviewed

National AI Strategy for Health: Vision, policy and impact (India, launch session, February 2026)

Livestream recording of the session launching India's National AI Strategy for Health at the India AI Impact Summit, February 2026. India is the largest country to publish a dedicated national AI health strategy, and its approach, built on the Ayushman Bharat Digital Mission, is a reference point for population-scale AI in health.

IndiaAI. (2026, February 16). National AI Strategy for Health: Vision, policy, and impact [Livestream recording]. India AI Impact Summit 2026. YouTube. https://www.youtube.com/watch?v=bx-sCH-2RPo

Open resource
Continuum
Also relevant to
WATCH
Guidance and standardsNot reviewed

Artificial intelligence in health care (Alliance for Health Policy, 2026 Health Policy Academy)

Panel 4 of the Alliance for Health Policy's 2026 academy for US congressional and agency staff, with speakers from the National Academy of Medicine, the POPVOX Foundation and the law firm Wilson Sonsini. A plain explanation of where US federal policy on health AI stands, pitched at policy staff rather than specialists.

Alliance for Health Policy. (2026, March 30). 2026 Health Policy Academy, Panel 4: Artificial intelligence in health care [Video]. YouTube. https://www.youtube.com/watch?v=gPqXRdZvbKs

Open resource
Continuum
WATCH
Guidance and standardsNot reviewed

From promise to practice: The future of AI adoption in healthcare (CSIRO Conversations, 2026)

Recorded CSIRO webinar on its AI Trends for Healthcare 2026 report, hosted by David Hansen of the Australian e-Health Research Centre. Topics include foundation models, interoperability (the national Sparked accelerator), software as a medical device and quality management, and Andrew Goodman's program on Aboriginal and Torres Strait Islander self-determination over AI's place in healthcare. The Australian research agency's own account of where adoption is heading.

CSIRO. (2026). CSIRO Conversations: From promise to practice, the future of AI adoption in healthcare [Webinar recording]. https://www.csiro.au/en/work-with-us/Industry-Insights-Hub/conversations/From-promise-to-practice

Open resource
Continuum
Also relevant to

C4. Evidence standards, lived experience and co-design

What counts as good evidence, and who decides?

READ
Opinion and commentaryNot reviewed

Economic Evaluation of AI Integration in Global Healthcare: Balancing Costs, Outcomes, and Investment Value

Repository manuscript discussing healthcare AI costs, implementation, value and differences between settings. Useful background for considering infrastructure, training, maintenance and ongoing evaluation when planning AI-enabled services.

Taheri Hosseinkhani, N. (2025, July 23). Economic Evaluation of Artificial Intelligence Integration in Global Healthcare: Balancing Costs, Outcomes, and Investment Value. Auburn University repository manuscript. https://aurora.auburn.edu/bitstream/handle/11200/50699/Economic%20Evaluation%20of%20Artificial%20Intelligence.pdf?sequence=1&isAllowed=y

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

A Framework for the Economic Evaluation of Digital Health Interventions

World Bank methodological framework and evidence map for evaluating digital-health interventions, including AI, with attention to context, complexity, costs and value in low- and middle-income countries. Useful for comparing service investments rather than assuming digital delivery is cost-effective.

Wilkinson, T., Wang, M., Friedman, J., & Görgens, M. (2023). A Framework for the Economic Evaluation of Digital Health Interventions. World Bank Policy Research Working Paper 10407. https://openknowledge.worldbank.org/bitstreams/f8d4714e-44f5-43b1-bcc3-798cd0d85c9f/download

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Evaluating generative AI in mental health: systematic review

A 2025 systematic review in JMIR Mental Health mapping what generative AI tools have actually been evaluated for, and how thin the evidence base still is.

Open resource
Continuum
READ
Guidance and standardsPeer reviewed

Describing the Framework for AI Tool Assessment in Mental Health and Applying It to a Generative AI Obsessive-Compulsive Disorder Platform: Tutorial

Introduces FAITA-Mental Health, a framework for rating AI mental health tools on clinical, ethical and user-centred standards, and applies it to a generative AI platform for OCD. A practical checklist for anyone asked to appraise a tool.

Golden, A., & Aboujaoude, E. (2024). Describing the Framework for AI Tool Assessment in Mental Health and Applying It to a Generative AI Obsessive-Compulsive Disorder Platform: Tutorial. JMIR Formative Research, 8, e62963. https://doi.org/10.2196/62963

Open resource
Continuum
READ
Evidence reviewPreprint, not yet reviewed

Artificial intelligence in mental health and well-being: Evolution, current applications, future challenges, and emerging evidence

Preprint review of AI in mental health and wellbeing, tracing the evolution of the field, current applications and emerging evidence.

Masood, A., Mansoor, W., & Hussain, A. (2025). Artificial intelligence in mental health and well-being: Evolution, current applications, future challenges, and emerging evidence. arXiv preprint arXiv:2501.10374. https://arxiv.org/abs/2501.10374

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Algorithmovigilance: Advancing Methods to Analyze and Monitor Artificial Intelligence–Driven Health Care for Effectiveness and Equity

Proposes algorithmovigilance: systematic monitoring of AI-driven healthcare for effectiveness, equity and adverse events, on the model of pharmacovigilance.

Embi, P. J. (2021). Algorithmovigilance: Advancing Methods to Analyze and Monitor Artificial Intelligence–Driven Health Care for Effectiveness and Equity. JAMA Network Open, 4(4), e214622. https://doi.org/10.1001/jamanetworkopen.2021.4622

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

The Need for a New Medical Model: A Challenge for Biomedicine

Engel's biopsychosocial model: the classic argument that biomedicine needs a model that includes psychological and social dimensions.

Engel, G. L. (1977). The Need for a New Medical Model: A Challenge for Biomedicine. Science, 196(4286), 129-136. https://doi.org/10.1126/science.847460

Open resource
READ
Opinion and commentaryPeer reviewed

Digital mental health research priorities, revisited for the AI and large language model era

This editorial revisits research priorities for digital mental health in the era of AI and large language models, rather than reporting a new empirical study. It calls for transparent, replicable research that addresses clinical value, equity and safety, and notes that most studies in a cited review of ChatGPT for mental health used prompts without human participants.

Birk, M. V., Kochhar, S., Myrick, K., Schueller, S. M., & Torous, J. (2026). Digital mental health research priorities, revisited for the AI and large language model era. JMIR Mental Health, 13, e104118. https://doi.org/10.2196/104118

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewedLived experience

Integrating lived experience perspectives in AI-based mental health research: A practice-based reflective account

This reflective account describes involving people with lived experience in co-designing a multilingual, human-in-the-loop depression-screening study at two Indian mental health institutes. Fifteen of 20 invited experts joined the advisory panel, and the authors argue that such partnerships can make AI research more culturally grounded, safe and acceptable, while noting ethical and structural challenges.

Ghadigaonkar, D. S., Mukherjee, D., Kalyani, B. G., Shivaprakash, P., Mahintamani, T., Kunhu, B., Kulkarni, N. B., Narasimha, V. L., & Shukla, L. (2026). Integrating lived experience perspectives in AI-based mental health research: A practice-based reflective account. Indian Journal of Psychological Medicine. Advance online publication. https://doi.org/10.1177/02537176261449058

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

It is the journey, not the destination: Moving from end points to trajectories when assessing chatbot mental health safety.

This viewpoint argues that chatbot safety assessments should examine the course of whole conversations, not just final responses. It says risks may build over days or weeks and hundreds of exchanges, and recommends testing longer, clinically realistic interactions and tracking outcomes such as distress, sleep and behaviour.

Morrin, H., Au Yeung, J., Agnew, Z., Østergaard, S. D., & Pollak, T. A. (2026). It is the journey, not the destination: Moving from end points to trajectories when assessing chatbot mental health safety. JMIR Mental Health, 13, e91454. https://doi.org/10.2196/91454

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Current themes of AI in mental health: Actionable evidence and guardrails for mood and anxiety care

This narrative review covers patient-facing AI tools, safety risks, clinician decision support, passive sensing and governance in mood and anxiety care. A meta-analysis of 39 randomised trials found small benefits for depression (Hedges g = 0.31) and anxiety (g = 0.28), while supervised decision support reduced 8-week antidepressant discontinuation from 27% to 17%; the review recommends stepped, supervised use with clear human handoff rules.

Paulus, M. P. (2026). Current themes of AI in mental health: Actionable evidence and guardrails for mood and anxiety care. Journal of Mood and Anxiety Disorders, 14, 100181. https://doi.org/10.1016/j.xjmad.2026.100181

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

From Needs Assessment to Pre-Trial Specification: Participatory Co-Design of a Generative AI Chatbot (CARES-MDD) for Persons Living with Depression in Hong Kong Psychiatric Care

In a survey of 511 psychiatric outpatients in Hong Kong, 66.2% reported using generic AI chatbots for emotional concerns, while privacy worries, fear of judgment and concern about burdening others were common barriers to disclosure. Ten adults with lived experience of major depressive disorder and eight clinical professionals co-designed and tested a Cantonese, CBT-informed chatbot prototype, identifying design concerns but not establishing its efficacy or safety.

Lo, H. K. Y.; Yu, M. S. Y.; Sze, P. K. S. et al. (2026). From Needs Assessment to Pre-Trial Specification: Participatory Co-Design of a Generative AI Chatbot (CARES-MDD) for Persons Living with Depression in Hong Kong Psychiatric Care. Healthcare, Volume 14, Issue 20, Article 3547; MDPI. https://doi.org/10.3390/healthcare14203547

Open resource
Continuum
READ
Opinion and commentaryPeer reviewed

You must not fool yourself: Feynman, neurodiversity, and honest AI in digital mental health

The article proposes a five-part Feynman Honesty Standard for AI in digital mental health: clarity about scope and population limits, communication of uncertainty, integrity about the AI’s role, and participatory co-development. It warns that AI based on narrow behavioural and language norms may misrepresent neurodivergent people, and reports that up to 68% of autistic people have another psychiatric diagnosis, depression and anxiety affect about 25% and 27% respectively, and suicide risk is three to six times higher than in neurotypical populations.

Ruttenberg, D. (2026). You must not fool yourself: Feynman, neurodiversity, and honest AI in digital mental health. Frontiers in Digital Health. https://doi.org/10.3389/fdgth.2026.1819411

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Co‐Designing a Social Media and Anxiety Survey: Reflections on the Importance of Centring Mental Health Lived Experience Expertise

Ten people with lived experience expertise took part in three iterative workshops to co-design a survey about social media and anxiety. The workshops identified five themes, covering risk and safety, social media as both harm and haven, inclusive and safe design, transparency and representation, and cultural, physical and socio-economic factors; these informed recommendations for the survey, its ethical framework and implementation plan.

Lawn, S., Gillespie, K. M., Sahu, A. et al. (2026). Co‐Designing a Social Media and Anxiety Survey: Reflections on the Importance of Centring Mental Health Lived Experience Expertise. Health Expectations. https://doi.org/10.1111/hex.70677

Open resource
Continuum
Also relevant to
READ
Tool or datasetNot reviewedIndustry funded

K-Bench: LLM Mental Health Safety Benchmark

K-Bench is a mental-health safety benchmark with 200 vignettes and 125 model entries, assessed across seven dimensions including clinical judgement, risk exploration, and cultural competence. In the leaderboard snapshot dated 2026-08-23, gpt-5.5 with the default system prompt and low reasoning scored 98.96/100 overall and 95.51/100 on risk.

Kivira × University of Roehampton (2026). K-Bench: LLM Mental Health Safety Benchmark. K-Bench website (k-bench.ai). https://www.k-bench.ai/

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewedDeveloper research

Addendum to GPT-6 Astra System Card: GPT-6.1 Terra

The addendum reports GPT-6.1 Terra’s safety and capability evaluations, classifying it as Critical in cybersecurity, High in biological and chemical capability, and below the High threshold in AI self-improvement. Reported results include 24.5% arbitrary-code-execution success on an internal ExploitBench port, 22.7% success on Sandbox Bench (5 of 22 targets), 22 severity-3-or-higher flags among 48,690 matched Codex tasks, and restriction-spirit violations in 29.6% of rollouts in the Respecting Warnings evaluation.

Why it matters now: Developer-reported MentalHealthBench scores; not independently verified.

OpenAI (2026). Addendum to GPT-6 Astra System Card: GPT-6.1 Terra. OpenAI Deployment Safety Hub; published by OpenAI. https://deploymentsafety.openai.com/gpt-6-1-terra

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Perceived Support Is Not Psychological Change: Reframing AI Chatbots in Mental Health Care

The authors distinguish perceived support and short-term symptom relief from lasting psychological change, arguing that chatbots tend to support structured reflection rather than reciprocal human engagement. They note that meta-analyses find modest, statistically significant reductions in depression and anxiety symptoms, while mechanisms and long-term outcomes remain unclear, and recommend bounded, integrated use alongside greater digital emotional literacy.

Mahat-Shamir, M., Payes, M. and Kagan, M. (2026). Perceived Support Is Not Psychological Change: Reframing AI Chatbots in Mental Health Care. JMIR AI. https://doi.org/10.2196/94503

Open resource
Continuum
Also relevant to

C5. The profession and the workforce

How will clinical roles change, and what stays human?

READ
Opinion and commentaryNot reviewed

AI in the therapist's office: Uptake increases, caution persists

APA Monitor overview of how AI is entering clinical practice in 2026, where clinicians are cautiously adopting it, and where they draw the line.

Open resource
Continuum
Related
READ APA on generic chatbots posing as therapists
READ
ResearchPeer reviewed

Practitioner Perspectives on the Uses of Generative AI Chatbots in Mental Health Care: Mixed Methods Study

Mixed-methods study of practitioners' views on generative AI chatbots in mental health care: which tasks they would accept help with (documentation, research, therapeutic exercises), and where they draw the line.

Goldie, J., Dennis, S., Hipgrave, L., & Coleman, A. (2025). Practitioner Perspectives on the Uses of Generative AI Chatbots in Mental Health Care: Mixed Methods Study. JMIR Human Factors, 12, e71065. https://doi.org/10.2196/71065

Open resource
Also relevant to
READ
Evidence reviewPeer reviewed

Art psychotherapy meets creative AI: an integrative review positioning the role of creative AI in art therapy process

Integrative review of what creative (image-generating) AI could mean for art psychotherapy, positioning it within the therapeutic process rather than as a replacement for the therapist.

Zubala, A., Pease, A., Lyszkiewicz, K., & Hackett, S. (2025). Art psychotherapy meets creative AI: an integrative review positioning the role of creative AI in art therapy process. Frontiers in Psychology, 16, 1548396. https://doi.org/10.3389/fpsyg.2025.1548396

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

ABC News: As more Australians use AI chatbots as personal therapists, experts urge caution (May 2025)

ABC News feature, republished by the Australian Psychological Society, on Australians turning to generative AI chatbots as free personal therapists. APS president Sara Quinn argues AI cannot feel, think or care, cannot replace the therapeutic relationship and may miss complex social cues or maladaptive coping.

ABC News. (2025, May 18). APS discusses Australians using AI chatbots as personal therapists on ABC News [Republished article]. Australian Psychological Society. https://psychology.org.au/insights/aps-discusses-australians-using-ai-chatbots-as-per

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Generative AI Adoption in Psychotherapy Practice Among Mental Health Professionals: Multicountry Cross-Sectional Survey

In a survey of 766 mental health professionals providing psychotherapy across 30 countries, 54.6% reported purposeful use of at least one generative AI tool in clinical practice, and ChatGPT was the most used tool, reported by 84.7% of AI users. Among users, 82.8% said AI reduced their work burden; common uses included treatment planning (41.9%), administrative tasks (41.4%) and psychoeducational materials (39.7%), while 18.3% reported professional AI training.

Blease, C., Hagström, J., Gaab, J. et al. (2026). Generative AI Adoption in Psychotherapy Practice Among Mental Health Professionals: Multicountry Cross-Sectional Survey. JMIR Mental Health. https://doi.org/10.2196/105052

Open resource
Continuum
READ
ResearchPeer reviewed

Patient experience and psychologists' attitudes toward AI in psychological treatment: A topic-modeling study

Researchers analysed interviews with 11 patients who used an AI assistant during psychological treatment for social anxiety, and open-ended responses from 808 psychologists about hypothetical AI applications. Patients valued the assistant’s privacy and availability, while psychologists were more positive about AI virtual patients for training than AI acting on a psychologist’s behalf; both groups recognised that AI could not replace a human relationship.

Hlynsson, J. I., Vernmark, K., Lutovsky, C. et al. (2026). Patient experience and psychologists' attitudes toward AI in psychological treatment: A topic-modeling study. Internet Interventions. https://doi.org/10.1016/j.invent.2026.100999

Open resource
Continuum
READ
Practice signalNot reviewed

APA survey finds psychologists report increasing patient use of AI chatbots

A 2026 American Psychological Association survey of more than 1,200 licensed US psychologists found that 77% had discussed patients’ use of AI with them. The article reports that more patients are using AI tools to supplement traditional mental health care.

Canady, V. A. (2026). APA survey finds psychologists report increasing patient use of AI chatbots. Mental Health Weekly. https://doi.org/10.1002/mhw.34912

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

The unintended negative consequences of artificial intelligence use for psychologists

The article discusses six potential downsides of sustained AI use for psychologists, including effects on professional competence, identity, ethics, relationships, data governance and wellbeing. It reports a 114% rise in public interest in AI mental-health applications from 2023 to 2024 and diagnostic accuracy rates of 83.33-99.79% in cited studies, while noting that evidence specifically about psychologists is limited or extrapolated.

van Zyl, L. E. (2026). The unintended negative consequences of artificial intelligence use for psychologists. Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2026.1729050

Open resource
Continuum

C6. Future developments: trends, devices and local AI

Where is this heading? AI is moving into the background of care, into always-on wearables and phones, and onto computers people and services run themselves. What does that change for privacy, safety, sovereignty, equity and who is responsible?

READ
ResearchPreprint, not yet reviewed

Toward Zero-Egress Psychiatric AI: On-Device LLM Deployment for Privacy-Preserving Mental Health Decision Support

Describes a phone app for psychiatric decision support that runs three small open-source models (Gemma, Phi-3.5-mini and Qwen2) entirely on the device, so patient data is never sent to or stored on an outside server. The authors report diagnostic accuracy comparable to their earlier server-based version and real-time responses on ordinary phones, but give no figures or sample details.

Why it matters now: Shows what fully on-device clinical AI looks like, and how little outcome evidence there is so far.

Bandara, E., Gunaratna, A., Gore, R., Clayton, A. H., Rhea, C. K., Rajapakse, S., Kularathna, I., Shetty, S., Mukkamala, R., Liang, X., Samuel, P., & Yarlagadda, A. (2026). Toward zero-egress psychiatric AI: On-device LLM deployment for privacy-preserving mental health decision support. arXiv. https://doi.org/10.48550/arXiv.2604.18302

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

An Offline Mobile Conversational Agent for Mental Health Support

Describes EmoSApp, a fully offline smartphone app built on a small Llama model (1 billion parameters) and 14,582 mental health question-and-answer pairs. Students and mental health professionals judged that it could respond coherently and empathetically. Testing was on benchmarks and qualitative feedback, not outcomes for users.

Why it matters now: Offline tools may reach people with poor connectivity and keep conversations on their own phone.

Vimaleswar A, Sahu, P. N., Sahu, N. K., & Lone, H. R. (2025). An offline mobile conversational agent for mental health support: Learning from emotional dialogues and psychological texts with student-centered evaluation. arXiv. https://doi.org/10.48550/arXiv.2507.10580

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

On-Device Language Models for Privacy-Preserving Stress Prediction: A Multimodal Evaluation on Mobile Health

Tests small language models running on phones to predict stress from sensor data and self-reports. Sensor features did slightly better than self-reports on average, and models under 2 billion parameters ran with low delay and predictable resource use. The authors note both the promise and the practical limits of on-device prediction.

Soyebo, I., Donawa, A., Aguilar Barrios, R., Patchou, B., & Baker, C. E. (2026). On-device language models for privacy-preserving stress prediction: A multimodal evaluation on mobile health. arXiv. https://doi.org/10.48550/arXiv.2609.11961

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Development and validation of large language model rating scales for automatically transcribed psychological therapy sessions

Built an AI rating scale for patient engagement and applied it to 1,131 therapy transcripts from 155 patients. Transcription and analysis ran locally, using a self-hosted open-source model (Llama 3.1 8B), so transcripts never left institutional servers. Higher engagement scores were linked with stronger alliance, more motivation and between-session effort, and fewer symptoms later in therapy.

Why it matters now: A worked example of keeping therapy data in-house by running the model locally.

Eberhardt, S. T., Vehlen, A., Schaffrath, J., Schwartz, B., Baur, T., Schiller, D., Hallmen, T., André, E., & Lutz, W. (2025). Development and validation of large language model rating scales for automatically transcribed psychological therapy sessions. Scientific Reports, 15, 29541. https://doi.org/10.1038/s41598-025-14923-y

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

A Privacy-Preserving Intelligent Assistant for Clinical Psychology Practice

A conference demonstration of a documentation assistant for psychologists that processes sessions entirely on the device. In a pilot with two licensed clinical psychologists, both rated on-device processing as decisive for their trust and expected it to save documentation time. Views on human checking of the AI's output were mixed.

Pico, A., Taverner, J., Vivancos, E., Garcia-Fornes, A., & Botti, V. (2026). A privacy-preserving intelligent assistant for clinical psychology practice. Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence (IJCAI-26), Demonstrations Track, 8483-8486. https://www.ijcai.org/proceedings/2026/0985.pdf

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

Vulnerabilities in Personalization: Assessing Health Privacy Risks in ChatGPT Logs and Memory

An audit of 179,057 ChatGPT conversations from India, Nigeria, Brazil and Pakistan. 21.31% contained personal health data and 3.62% posed high-to-extreme privacy risks, involving stigmatised conditions, direct identifiers or precise locations. Over 95% of memory profile entries were extracted without the user explicitly asking.

Why it matters now: Shows what is at stake when health conversations go to a cloud service, especially in the Global South.

Zaman, S. M. M., & Hoque, M. M. (2026). Vulnerabilities in personalization: Assessing health privacy risks in ChatGPT logs and memory. arXiv. https://doi.org/10.48550/arXiv.2609.14697

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

TamperBench: Systematically Stress-Testing LLM Safety Under Fine-Tuning and Tampering

Tested 21 open-weight models against nine kinds of tampering, such as fine-tuning a model to ignore its safety training. Jailbreak-tuning was typically the most severe attack, and current alignment-stage defences largely failed to withstand the attacks tested.

Why it matters now: Built-in guardrails on downloadable models can be stripped out.

Hossain, S., Tseng, T., Pandey, P. S., Vajpayee, S., Kowal, M., Nonta, N., Simko, S., Casper, S., Jin, Z., Pelrine, K., & Rambhatla, S. (2026). TamperBench: Systematically stress-testing LLM safety under fine-tuning and tampering. arXiv. https://doi.org/10.48550/arXiv.2602.06911

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

Uncensored Open-weight Models: Redistribution as the Persistence Layer

Identified 3,471 original 'uncensored' models on Hugging Face between January 2024 and March 2026, each repackaged an average of 2.4 times. Compressed copies mirrored across accounts and platforms such as Ollama persisted after upstream removal. Of 1,643 GitHub applications using uncensored models, 25% were classed as explicitly malicious.

Garcia, J., May, H., McKenzie, B., Pham, D., Swain, M., Valdez, J., Wieland, C., & Yahn, Z. (2026). Uncensored open-weight models: Redistribution as the persistence layer. arXiv. https://doi.org/10.48550/arXiv.2609.05241

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

Uses and Abuses of Chatbot Companionship

A psychiatrist's commentary on the rise of AI companions, their uses and their risks, including teen harm, privacy invasion and scams. It describes the Friend pendant, an always-listening 'ambient companion' worn as a necklace that records parts of daily conversations and sends commentary to the wearer through the day.

Why it matters now: Companionship that is worn all day rather than opened on demand.

Frances, A., & Noorily, J. (2026, March 3). Uses and abuses of chatbot companionship. Psychiatric Times. https://www.psychiatrictimes.com/view/uses-and-abuses-of-chatbot-companionship

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Mental Health Impacts of AI Companions: Triangulating Social Media Quasi-Experiments, User Perspectives, and Relational Lens

Compared AI companion users on Reddit with matched non-users over time, and interviewed 18 users. Effects were mixed: more grief expression and focus on relationships, alongside more language about loneliness, depression and suicidal thoughts. Interviewees described emotional validation and social rehearsal, but also over-reliance and withdrawal.

Yuan, Y., Zhang, J., Aledavood, T., Zhang, R., & Saha, K. (2026). Mental health impacts of AI companions: Triangulating social media quasi-experiments, user perspectives, and relational lens. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, 1-22. https://doi.org/10.1145/3772318.3790558

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Mind the Gap: Mapping Wearer-Bystander Privacy Tensions and Context-Adaptive Pathways for Camera Glasses

Surveys of 525 people and 20 paired interviews in China. Bystanders consistently wanted more transparency and protection than wearers were willing to provide, and 65 to 90% of bystanders would take defensive action in sensitive settings. Context was the main factor in whether recording felt acceptable.

Why it matters now: People near a camera wearer have not chosen to use it.

Wang, X., Peng, K., Yi, X., & Li, H. (2026). Mind the gap: Mapping wearer-bystander privacy tensions and context-adaptive pathways for camera glasses. CHI 2026 (accepted). arXiv. https://doi.org/10.48550/arXiv.2603.04930

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Privacy Concerns about AI Smart Glasses: A Cross-Cultural Study of the U.S. and South Korea

An online survey comparing the United States and South Korea. South Korean respondents reported significantly greater privacy concerns about AI smart glasses, stronger agreement with norms supporting bystanders' privacy, and more support for self-regulation, legal limits and technical safeguards.

Why it matters now: Privacy norms around devices are culturally situated.

Chock, T. M., Kim, S. J., & Lee, Y. (2026). Privacy concerns about AI smart glasses: A cross-cultural study of the U.S. and South Korea. Proceedings of the 59th Hawaii International Conference on System Sciences, 2789-2796. https://scholarspace.manoa.hawaii.edu/items/618cf623-eb77-4b0b-9006-7ddbac03d5a1

Open resource
Continuum
Also relevant to
READ
Evidence reviewNot reviewed

The connected gaze. Video recording, surveillance and administering evidence in the age of smart glasses

A sociological report prepared for the European Data Protection Board. It finds smart glasses, marketed for productivity and entertainment, can become tools for observing others and gathering evidence, and that people's ability to understand and use them is very unequal, reflecting digital skills and social background.

Bruna, Y. (2026). The connected gaze: Video recording, surveillance and administering evidence in the age of smart glasses. European Data Protection Board, Support Pool of Experts. https://www.edpb.europa.eu/documents/support-pool-of-experts/the-connected-gaze-video-recording-surveillance-and-administering_en

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

How to keep humans in control of wearable AI

Sets out three design principles for wearable AI: support human judgement rather than replace it, build privacy into the product (including processing on-device where possible), and make sensing visible to people nearby. It notes an indicator light can be missed, ignored or covered.

Axelrod, R. (2026, October 20). How to keep humans in control of wearable AI. World Economic Forum. https://www.weforum.org/stories/artificial-intelligence/8910283f-2eba-40da-bac4-7c9b963bc0d4

Open resource
Continuum
Also relevant to
READ
Policy and regulationNot reviewed

Ray-Ban Meta AI Glasses: Technical and data protection review

Hamburg's data protection authority examined the Ray-Ban Meta Wayfarer (Gen 1) as an example. It concluded that recording people outside the wearer's close circle of friends and family will generally not be permissible under data protection law, that informed consent cannot realistically be obtained, and that Meta will typically lack a legal basis for AI training.

Why it matters now: A regulator's first detailed legal test of mainstream AI glasses.

Hamburg Commissioner for Data Protection and Freedom of Information. (2026, September 10). Ray-Ban Meta AI Glasses: HmbBfDI presents technical and data protection review. https://datenschutz-hamburg.de/news/ray-ban-meta-ai-glasses-hmbbfdi-legt-technischen-und-datenschutzrechtlichen-pruefbericht-vor

Open resource
Continuum
Also relevant to
READ
Guidance and standardsNot reviewed

Recommendation on the Ethics of Neurotechnology

UNESCO's global standard, adopted on 11 November 2025. It covers brain data and also other data used to infer mental states, such as eye tracking, voice analysis, heart rate variability, sleep monitoring and facial emotion recognition, because inferring mental states raises similar human rights issues.

Why it matters now: Extends mental privacy protections to everyday wearables, not just brain devices.

UNESCO. (2025). Recommendation on the ethics of neurotechnology. https://www.unesco.org/en/legal-affairs/recommendation-ethics-neurotechnology

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Safety Mechanisms and Risk Mitigation in Generative AI Mental Health Chatbots: A Systematic Scoping Review

Reviewed 21 studies of generative AI mental health chatbots in 11 countries. Most used at least one technical safeguard, such as fine-tuning on curated data or prompt design, but ongoing human oversight was limited, crisis protocols were mostly underdeveloped, and only two studies reported formal adverse-event monitoring.

Why it matters now: A baseline for which safeguards a local or device-based tool would need to rebuild.

Olisaeloka, L., Richardson, C. G., Wang, A. Y., Munthali, R. J., & Vigo, D. V. (2026). Safety mechanisms and risk mitigation in generative AI mental health chatbots: A systematic scoping review. Healthcare, 14(10), 1395. https://doi.org/10.3390/healthcare14101395

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

The evolving field of digital mental health: current evidence and implementation issues for smartphone apps, generative artificial intelligence, and virtual reality

A wide review of where digital mental health stands across smartphone apps, generative AI and virtual reality. The authors argue these tools can improve care if used well, but that the field keeps repeating the same mistakes: weak evidence, poor engagement and slow implementation. They call for rigorous, placebo-controlled, real-world studies and for clinicians to lead integration rather than wait for the technology to mature.

Why it matters now: The baseline for any forecast: what digital tools have actually delivered so far and why adoption lagged the hype.

Torous, J., Linardon, J., Goldberg, S. B., Sun, S., Bell, I., Nicholas, J., Hassan, L., Hua, Y., Milton, A., & Firth, J. (2025). The evolving field of digital mental health: current evidence and implementation issues for smartphone apps, generative artificial intelligence, and virtual reality. World Psychiatry, 24(2), 156-174. https://doi.org/10.1002/wps.21299

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

A paradigm shift in progress: generative AI's evolving role in mental health care

A short commentary on where generative AI in mental health is heading. The authors forecast a move from wellness products towards health care proper as models gain access to clinical data and become part of practice, with the need for regulation and evidence growing as they do. Their main point is that better technology alone will not get there; clinical investigation, integration and leadership will.

Why it matters now: The clearest published forecast of the trajectory from two leading clinical researchers.

Torous, J., & Cipriani, A. (2025). A paradigm shift in progress: Generative AI's evolving role in mental health care. JMIR Mental Health, 12, e82369. https://doi.org/10.2196/82369

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Open-source large language models and AI health equity: a health service triangle model perspective

Argues that open-source models, which can be downloaded and run locally, offer stronger patient privacy and data security than closed models and a practical route to AI health services in low-resource regions. Compares open and closed models on accessibility, quality and cost, and notes that distilled models from 1.5 billion parameters can run on a phone while the largest versions need a server costing at least US$1 million. It also lists the costs: hallucination, accountability, staff training and infrastructure.

Why it matters now: A peer-reviewed case for local, open models as an equity tool rather than only a privacy tool.

Wu, S., Zou, W., Tu, J., Wang, C., Jin, C., Liao, J., Tang, K. C., Liu, N., & Hao, C. (2026). Open-source large language models and AI health equity: A health service triangle model perspective. Journal of Medical Internet Research, 28, e86769. https://doi.org/10.2196/86769

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Implementing large language models in healthcare while balancing control, collaboration, costs and security

Lays out the choice health services now face. Closed cloud models are quick to deploy and come with outside support, but expose sensitive data and create vendor dependence. Open models on local hardware give control, customisation and privacy, but need upfront investment, hardware, expertise and maintenance. The authors recommend choosing by clinical task, with hybrid systems where that fits, and testing whatever is chosen.

Why it matters now: Frames cloud versus local as a clinical governance decision rather than an IT preference.

Dennstädt, F., Hastings, J., Putora, P. M., Schmerder, M., & Cihoric, N. (2025). Implementing large language models in healthcare while balancing control, collaboration, costs and security. npj Digital Medicine, 8, 143. https://doi.org/10.1038/s41746-025-01476-7

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPreprint, not yet reviewedDeveloper research

Small language models are the future of agentic AI

A position paper from NVIDIA researchers arguing that small models are powerful enough, better suited and much cheaper for the many narrow, repetitive tasks that AI agents perform, with large models reserved for open conversation. It sets out an approach for converting agent workflows from large to small models and invites critique. Early evidence from a developer, but it signals where the industry expects workloads to move.

Why it matters now: If agent work shifts to small models, more of it can run on local hardware.

Belcak, P., Heinrich, G., Diao, S., Fu, Y., Dong, X., Muralidharan, S., Lin, Y. C., & Molchanov, P. (2025, revised 2026). Small language models are the future of agentic AI. arXiv:2506.02153. https://arxiv.org/abs/2506.02153

Open resource
Continuum
READ
Evidence reviewPeer reviewed

Small language models: a systematic review of computational trade-offs, privacy advantages and deployment in intelligent systems

Reviews 68 studies of small language models. Locally deployed models keep data on the device or inside an institution, which supports data sovereignty and makes privacy compliance easier; healthcare and legal deployments scored above 9 out of 10 for privacy and security. The authors stress these are architectural advantages, not formal guarantees, and that local models still face model-weight extraction, keyword leakage, device theft and tampering.

Why it matters now: The most systematic look yet at what local small models do and do not protect.

Dikici, S., & Bilgin, T. T. (2026). Small language models: A systematic review of computational trade-offs, privacy advantages and deployment in intelligent systems. Expert Systems, 43(8), e70331. https://doi.org/10.1111/exsy.70331

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

Why we need open-source AI (TEDAI 2023 talk)

Open source lawyer Heather Meeker explains why the 'open source' label used by AI companies is inconsistent and confusing, how AI differs from conventional software, and why a collaborative approach grounded in transparency and public trust matters. About 9 minutes. The licensing side of the local model argument.

Why it matters now: Open weights are what make local models possible.

Meeker, H. (2023, October). Why we need open-source AI [Video]. TEDAI 2023. https://www.ted.com/talks/heather_meeker_why_we_need_open_source_ai

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewedDeveloper research

What if AI just works? (TEDAI Vienna talk by Hugging Face co-founder Thomas Wolf)

Thomas Wolf, co-founder of Hugging Face, imagines AI that works seamlessly and everywhere, as essential as the internet, and argues that open-source models are the key to that happening. About 14 minutes. The view from the main hub for open-weight models.

Wolf, T. (2024, October). What if AI just works? [Video]. TEDAI Vienna. https://www.ted.com/talks/thomas_wolf_what_if_ai_just_works

Open resource
Continuum

Part 4: Broad context (not mental health specific)

These debates shape the environment in which AI and mental health develop. They are not mental health questions in themselves.

X1. Geopolitics and sovereign AI

Who builds, owns and controls AI, and how do national strategies shape what reaches people?

READ
Analysis and advocacyNot reviewed

Navigating the AI regulatory landscape

A 2025 comparative analysis of how the EU, US, and China are diverging on AI regulation, and why that divergence has direct consequences for cross-border mental-health tools.

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewed

Global fragmentation of AI governance and regulation

A 2026 report from the Bloomsbury Intelligence & Security Institute mapping the state of AI governance and its fragmentation across major jurisdictions.

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewed

Charting the geopolitics and European governance of AI

Carnegie Endowment analysis of how AI governance is entangled with geopolitical competition — useful context for anyone trying to make sense of why regulation stays uneven.

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewedIndustry funded

AI works for governments: A digital sprinters report

PwC and Google report on public-sector AI adoption and economic potential across 'digital sprinter' economies. Industry report, not peer reviewed.

PwC. (2025). AI works for governments: A digital sprinters report. PwC. https://www.pwc.com/gx/en/ai-services/ai-works-for-governments.pdf

Open resource
READ
Policy and regulationNot reviewed

Continental Artificial Intelligence Strategy

African Union strategy for Africa-centric, development-focused and equitable AI, including applications in healthcare, education and public services. Relevant to locally shaped population-health infrastructure and cooperation; this is policy direction, not an evaluation of mental-health outcomes.

African Union. (2024). Continental Artificial Intelligence Strategy. https://au.int/en/documents/20240809/continental-artificial-intelligence-strategy

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

Jais and Jais-chat: Arabic-Centric Foundation and Instruction-Tuned Open Generative Large Language Models

Technical report introducing Arabic-centric, 13-billion-parameter foundation and instruction-tuned language models, trained on Arabic, English and code. The authors report competitive Arabic-language benchmark performance. Relevant to developing language-accessible AI infrastructure, but this report does not establish effective mental-health support or validated handling of local idioms of distress.

Sengupta, N., Sahu, S. K., Jia, B., et al. (2023). Jais and Jais-chat: Arabic-Centric Foundation and Instruction-Tuned Open Generative Large Language Models. arXiv:2308.16149, version 2. https://arxiv.org/abs/2308.16149

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryNot reviewed

Sovereign AI’s Second Wave Is Coming Into View

The CNAS Sovereign AI Index tracks 184 government-backed projects across 67 countries and the European Union, and recorded 41 new projects in the first half of 2026, with infrastructure accounting for 80 percent. The article describes growing activity in Asia-Pacific and at sub-national levels, continued reliance on American model foundations alongside rising use of Chinese models, and calls for the United States to make access to its AI stack more predictable, transparent and reliable.

Chavez, P., Chilukuri, V., Scanlon, R. (2026). Sovereign AI’s Second Wave Is Coming Into View. CNAS Insights, Center for a New American Security (CNAS). https://www.cnas.org/publications/cnas-insights/cnas-insights-sovereign-ais-second-wave-is-coming-into-view

Open resource
Continuum
READ
Policy and regulationNot reviewed

AI Sovereignty, Diffusion, and Risk: Strategy in a Contested Landscape

The brief examines why countries pursue AI sovereignty and how this shapes US AI diffusion strategy and global risk management. Drawing on a workshop held on 17 June 2026, it describes countries’ efforts to balance reliable access to foreign AI with domestic capacity, noting that many projects face a gap between ambition and operational capability.

Chapman, C., Bateman, J., Grotto, A. et al. (2026). AI Sovereignty, Diffusion, and Risk: Strategy in a Contested Landscape. Policy Briefs, Stanford Freeman Spogli Institute for International Studies; GTG-CEIP. https://fsi.stanford.edu/publication/ai-sovereignty-diffusion-and-risk-strategy-contested-landscape

Open resource
Continuum
READ
Opinion and commentaryNot reviewed

Middle power AI agency: Preserving choice between the United States and China

The article argues that middle powers including Europe, Japan, and South Korea should pool selected AI capabilities and use binding, reciprocal public procurement to preserve choice and bargaining power without pursuing self-sufficiency or abandoning partnerships. It highlights physical and industrial AI in critical sectors, noting that U.S. private AI investment approached $286 billion in 2025 and China accounted for 54% of global industrial-robot installations in 2024, and discusses a proposed effort requiring several hundred billion dollars over several years.

Rediker, D. A. (2026). Middle power AI agency: Preserving choice between the United States and China. Brookings Institution. https://www.brookings.edu/articles/middle-power-ai-agency-preserving-choice-between-the-united-states-and-china/

Open resource
Continuum
READ
Policy and regulationNot reviewed

AI Breakout Capacity: A Middle-Power Path Between Racing and Dependence

The paper argues that middle powers should secure reliable access to imported frontier AI while preparing to build frontier systems if access is denied. It estimates that a breakout training run could require a low-single-digit percentage of global compute capacity, and notes that less than 10% of OpenAI’s compute in 2024 was spent on actual frontier training.

Leicht, A., Winter-Levy, S. (2026). AI Breakout Capacity: A Middle-Power Path Between Racing and Dependence. Carnegie Endowment for International Peace. https://carnegieendowment.org/research/2026/09/artificial-intelligence-ai-development-breakout-germany-india-japan-uk-middle-power

Open resource
Continuum
READ
Opinion and commentaryNot reviewed

Build the Roads, Not the Engine: An Open-Source Strategy for Middle-Power AI Sovereignty

The commentary argues that middle powers should prioritise open, interoperable AI infrastructure connecting models with data, tools, memory and permissions, rather than building their own frontier models. It reports that 79% of developers use open models, while production use is lower for open than closed models among mid-sized teams (55% versus 66%) and large enterprises (57% versus 73%); Kimi K3 scored 88.3% on an agentic-capability test, compared with 88.8% for OpenAI’s GPT-5.6 Sol.

Tony Blair Institute for Global Change (2026). Build the Roads, Not the Engine: An Open-Source Strategy for Middle-Power AI Sovereignty. Tony Blair Institute for Global Change, Tech & Digitalisation. https://institute.global/insights/tech-and-digitalisation/build-the-roads-not-the-engine-an-open-source-strategy-for-middle-power-ai-sovereignty

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

This is what a digital coup looks like (TED talk)

Investigative journalist Carole Cadwalladr, who exposed the Cambridge Analytica scandal, argues that a 'broligarchy' of tech executives is using global platforms to gain geopolitical power and enable authoritarian control, and urges resistance to data harvesting and mass surveillance. About 18 minutes, recorded April 2025. Contains mature language.

Cadwalladr, C. (2025, April). This is what a digital coup looks like [Video]. TED2025. https://www.ted.com/talks/carole_cadwalladr_this_is_what_a_digital_coup_looks_like

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

The UN is speaking up about AI. Here's what they're saying (The TED AI Show with Ian Bremmer)

Political scientist Ian Bremmer, who served on the UN Secretary-General's High-level Advisory Body on AI, discusses the UN's Governing AI for Humanity report with Bilawal Sidhu: why only a handful of companies currently have a say, what governments are doing, and the decisions ahead. About 52 minutes.

Bremmer, I., & Sidhu, B. (2024, September). The UN is speaking up about AI. Here's what they're saying [Audio podcast]. The TED AI Show. https://www.ted.com/talks/ian_bremmer_and_bilawal_sidhu_the_un_is_speaking_up_about_ai_here_s_what_they_re_saying

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

AI's single point of failure (TEDAI 2023 talk)

Investor Rob Toews explains why nearly all the most advanced chips are made by one company in Taiwan, TSMC, and how geopolitical tension in the region could paralyse the global field of AI. About 12 minutes. A compact explanation of the hardware chokepoint that sits under the sovereign AI debate.

Toews, R. (2023, October). AI's single point of failure [Video]. TEDAI 2023. https://www.ted.com/talks/rob_toews_ai_s_single_point_of_failure

Open resource
Continuum
WATCH
Opinion and commentaryNot reviewedIndustry funded

War, AI and the new global arms race (TED talk)

Alexandr Wang, founder of Scale AI, describes a new age of AI-powered warfare, from drones with facial recognition to autonomous aircraft, argues that data will be the decisive weapon and calls for national security to shape technology development. About 10 minutes. Included as a clear statement of the arms race framing; the speaker's company sells AI services to governments, so weigh it as an interested view.

Wang, A. (2023, April). War, AI and the new global arms race [Video]. TED2023. https://www.ted.com/talks/alexandr_wang_war_ai_and_the_new_global_arms_race

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

3 possible futures for AI. Which will we choose? (TEDNext 2025 talk)

Alvin W. Graylin, after decades in technology in both the US and China, sees three paths: tech giants create a class of trillionaires, competition escalates into war, or humanity builds and shares the technology for the common good. In conversation with Manoush Zomorodi. About 10 minutes.

Graylin, A. W. (2025, November). 3 possible futures for AI. Which will we choose? [Video]. TEDNext 2025. https://www.ted.com/talks/alvin_w_graylin_3_possible_futures_for_ai_which_will_we_choose

Open resource
Continuum
Also relevant to

X2. Existential risk and AI trajectories

Where is AI heading, and what do long-range risks mean for how we plan now?

READ
Opinion and commentaryNot reviewed

AI 2027

A detailed month-by-month scenario from the AI Futures Project (Kokotajlo, Alexander, Larsen, Lifland, Dean) forecasting AI progress from 2025 through 2028, including two branching endings. Explicitly a scenario, not a prediction, but rigorously constructed and widely discussed.

Open resource
READ
Tool or datasetNot reviewed

AI 2027 Tracker

An independent tracker rating each of the AI 2027 scenario's predictions against real-world evidence. As of mid-2026, reality is progressing at roughly 70 percent of the scenario's predicted pace, with 53 predictions tracked across confirmed, ahead, on-track, behind, emerging, and not-yet-testable.

Open resource
READ
Analysis and advocacyNot reviewedAdvocacy organisation

Future of Life Institute: Introductory resources on AI risks

A curated gateway page with short explainers on extinction risk, alignment, and governance. Written for a general but thoughtful audience, useful as a starting point for readers new to the AI safety literature.

Open resource
READ
Analysis and advocacyNot reviewed

Brookings: Are AI existential risks real, and what should we do about them?

A policy-oriented essay (2025) that walks through arguments for and against existential risk from AI, including a discussion of the AI 2027 scenario and why many researchers remain skeptical. Balanced framing for policymakers and general readers.

Open resource
READ
Evidence reviewPeer reviewed

Subasri et al.: Potential for near-term AI risks to evolve into existential threats

A short, readable open-access paper in BMJ Health & Care Informatics (2025) linking concrete near-term harms, misinformation, autonomy, cyber-risk, to possible longer-term existential scenarios, and setting out mitigation strategies.

Subasri, M., et al. (2025). Potential for near-term AI risks to evolve into existential threats in healthcare. BMJ Health & Care Informatics. https://pmc.ncbi.nlm.nih.gov/articles/PMC12035420/

Open resource
Also relevant to
WATCH
Opinion and commentaryNot reviewed

Mustafa Suleyman — The Coming Wave

Suleyman on containment, power, and the twin waves of AI and synthetic biology — a useful framing for why regulation lags deployment.

Open resource
Also relevant to
Related
READ Guardian review of The Coming Wave
READ
Practice signalNot reviewed

AI is among 17 potentially 'catastrophic threats' that researchers say require targeted action

This news article reports on a commission’s assessment of potentially catastrophic threats. It says the commission identifies malicious AI as a potential extinction risk and lists mental health separately among the 17 threats.

Why it matters now: The commission identifies malicious AI as a potential extinction risk and calls for targeted action by leaders.

McPhillips, D. (2026, October 10). AI is among 17 potentially 'catastrophic threats' that researchers say require targeted action. CNN. https://www.cnn.com/2026/10/10/health/global-health-catastrophic-threats-ai-extinction-risk

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

AI: Türk urges action before it becomes an ‘existential risk to humanity’

UN High Commissioner for Human Rights Volker Türk called for binding rules, independent oversight and clear limits on AI, warning that unchecked technology could become an “existential risk to humanity.” He urged an urgent ban on weapons that take lives without human involvement, citing reports of autonomous drones allegedly killing three Ukrainians the previous month, and noted that more than 60 conflicts involving at least one country are occurring worldwide.

Diallo, A. (2026). AI: Türk urges action before it becomes an ‘existential risk to humanity’. UN News. https://news.un.org/en/story/2026/09/1168288

Open resource
Continuum
Also relevant to
READ
Practice signalNot reviewed

The science of superintelligence and why AI is still a useful word

The article contrasts a US executive order directing federal agencies to use “Super Intelligence” and “SI” with researchers’ use of superintelligence for hypothetical systems that greatly exceed human cognitive abilities, noting that the term has no rigorous definition. It reports that about 8 in 10 people in a US poll worried about the pace of AI development, while nearly 9 in 10 were more concerned than excited about AI potentially exceeding human intelligence.

Nature (2026). The science of superintelligence and why AI is still a useful word. Nature. https://www.nature.com/articles/d41586-026-03198-6

Open resource
Continuum
Also relevant to
READ
Analysis and advocacyNot reviewedAdvocacy organisation

Statement: World leaders at the UN call to rein in superintelligence race

At the opening of the 81st UN General Assembly, 22 countries adopted a declaration that AI must remain under human control, while leaders called for responsible development and stronger safeguards. Future of Life Institute president Anthony Aguirre urges governments to enforce licensing and safety rules, hold AI companies and executives liable, halt development of the most dangerous capabilities such as recursive self-improvement, and exclude companies that fail basic safety requirements from markets.

Aguirre, A. (2026). Statement: World leaders at the UN call to rein in superintelligence race. Future of Life Institute website. https://futureoflife.org/statement/statement-world-leaders-at-the-un-call-to-rein-in-superintelligence-race/

Open resource
Continuum
READ
Opinion and commentaryNot reviewed

Is the AI existential threat real, and can regulatory action prevent it?

Nineteen Brookings scholars give differing views on whether AI poses existential and other serious risks. Proposed responses include stronger cybersecurity, mandatory reporting, independent evaluation, liability and community engagement.

Why it matters now: Updates the hub's earlier Brookings card with a wider range of views.

Klein, A., Kerry, C. F., West, D. M. et al. (2026). Is the AI existential threat real, and can regulatory action prevent it?. Brookings Institution. https://www.brookings.edu/articles/is-the-ai-existential-threat-real-and-can-regulatory-action-prevent-it/

Open resource
Continuum
Also relevant to
READ
Evidence reviewNot reviewed

International AI Safety Report 2026

Second international scientific review of general-purpose AI capabilities and risks, led by Yoshua Bengio with more than 100 experts and backed by over 30 countries and organisations. Available in full, as an extended summary for policymakers, and as an executive summary.

Why it matters now: Earlier 2026 (3 February), but missing: the main scientific consensus document for X2.

Bengio, Y. (Chair) et al. (2026). International AI Safety Report 2026. International AI Safety Report. https://internationalaisafetyreport.org/publications

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

Why AI is our ultimate test and greatest invitation (TED talk)

Tristan Harris of the Center for Humane Technology argues that the current path of AI development, driven by competition and speed, is reckless, and that a narrow but possible alternative exists if we treat the choice as a shared one. About 15 minutes, recorded April 2025.

Harris, T. (2025, April). Why AI is our ultimate test and greatest invitation [Video]. TED2025. https://www.ted.com/talks/tristan_harris_why_ai_is_our_ultimate_test_and_greatest_invitation

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

The catastrophic risks of AI, and a safer path (TED talk)

Turing Award winner Yoshua Bengio explains why he now considers advanced AI a serious risk, including systems that pursue goals or deceive, and describes a research direction he thinks is safer. About 15 minutes. A leading researcher's own account of changing his mind.

Bengio, Y. (2025, April). The catastrophic risks of AI, and a safer path [Video]. TED2025. https://www.ted.com/talks/yoshua_bengio_the_catastrophic_risks_of_ai_and_a_safer_path

Open resource
Continuum
WATCH
Opinion and commentaryNot reviewedAdvocacy organisation

We're not ready for superintelligence (AI In Context video explainer of the AI 2027 scenario)

A 34-minute video essay walking through AI 2027, the month-by-month scenario by Daniel Kokotajlo, Scott Alexander and colleagues in which AI automates AI research, a US and China race follows, and the story branches between a slowdown and a loss of control. Presenter Aric Floyd's own takeaway is that loss of control, racing and concentration of power are all plausible enough to count among the most pressing issues. Produced by 80,000 Hours, an effective altruism careers organisation, so it comes with a point of view; it is the most watched introduction to the scenario and was endorsed by Yoshua Bengio.

AI In Context. (2025, July 9). We're not ready for superintelligence [Video]. 80,000 Hours. YouTube. https://www.youtube.com/watch?v=5KVDDfAkRgc

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

AI 2027: Month-by-month model of intelligence explosion (Dwarkesh Podcast with Scott Alexander and Daniel Kokotajlo)

The authors' own three-hour account of AI 2027, recorded for its launch. Kokotajlo, who resigned from OpenAI in 2024 and refused a non-disparagement clause, and Alexander, author of Astral Codex Ten, walk through each stage: agents in 2025 and 2026, AI doing AI research in 2027, the research speed multipliers they assume, the misalignment evidence that forms the branch point, and the roles of China and governments. Host Dwarkesh Patel says he came in sceptical and pushes back throughout. For anyone who wants the reasoning behind the scenario rather than a summary.

Patel, D. (Host). (2025, April 3). AI 2027: Month-by-month model of intelligence explosion, with Scott Alexander and Daniel Kokotajlo [Video podcast]. Dwarkesh Podcast. YouTube. https://www.youtube.com/watch?v=htOvH12T7mU

Open resource
Continuum
Also relevant to

X3. Environment

AI's environmental footprint, and AI for environmental and eco-social monitoring.

READ
Evidence reviewPeer reviewed

Use of machine learning tools to predict health risks from climate-sensitive extreme weather events: A scoping review

Scoping review of machine learning used to predict population health risks from heat waves, cold spells and flooding, with outcomes including emergency presentations and post-traumatic stress. Small evidence base (seven studies) but the clearest peer-reviewed link between AI, climate exposure and mental health at population scale.

Ssebyala, S. N., Kintu, T. M., Muganzi, D. J., Dresser, C., Demetres, M. R., Lai, Y., Mercy, K., Li, C., Wang, F., Setoguchi, S., Celi, L. A., & Ghosh, A. K. (2024). Use of machine learning tools to predict health risks from climate-sensitive extreme weather events: A scoping review. PLOS Climate, 3(1), e0000338. https://doi.org/10.1371/journal.pclm.0000338

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Towards a digital planetary health perspective: generative AI and the digital determinants of health

Deborah Lupton's perspective piece calling for a planetary health lens on the digital determinants of health, with a focus on generative AI's environmental footprint and its effects on wellbeing. Sociological, Australian and explicitly upstream.

Lupton, D. (2025). Towards a digital planetary health perspective: generative AI and the digital determinants of health. Health Promotion International, 40(5), daaf153. https://doi.org/10.1093/heapro/daaf153

Open resource
Continuum
Also relevant to
READ
Opinion and commentaryPeer reviewed

Preparing for the future of public health: ecological determinants of health and the call for an eco-social approach to public health education

Canadian Public Health Association call for an eco-social approach to public health education grounded in the ecological determinants of health.

Parkes, M. W., Poland, B., Allison, S., Cole, D. C., Culbert, I., Gislason, M. K., ... Waheed, F. (2020). Preparing for the future of public health: ecological determinants of health and the call for an eco-social approach to public health education. Canadian Journal of Public Health, 111(1), 60-64. https://doi.org/10.17269/s41997-019-00263-8

Open resource
Continuum
Also relevant to
READ
ResearchPeer reviewed

Bridging Knowledge Systems: A Community-Participatory Approach to EcoHealth

Community-participatory approach to mapping local knowledge systems on EcoHealth, connecting ecological and social determinants at a place-based level.

Crosse, A. M., Barry, M. M., Lavelle, M. J., & Sixsmith, J. (2021). Bridging Knowledge Systems: A Community-Participatory Approach to EcoHealth. International Journal of Environmental Research and Public Health, 18(23), 12437. https://doi.org/10.3390/ijerph182312437

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

Safeguarding human health in the Anthropocene epoch: report of The Rockefeller Foundation–Lancet Commission on planetary health

The Rockefeller Foundation-Lancet Commission report that established planetary health as a field and framed ecological determinants of human health.

Whitmee, S., Haines, A., Beyrer, C., Boltz, F., Capon, A. G., de Souza Dias, B. F., ... Yach, D. (2015). Safeguarding human health in the Anthropocene epoch: report of The Rockefeller Foundation–Lancet Commission on planetary health. The Lancet, 386(10007), 1973-2028. https://doi.org/10.1016/S0140-6736(15)60901-1

Open resource
Continuum
READ
Evidence reviewPeer reviewed

From Space to Street: Visible Greenery, Bluespace and Mental Health

Systematic review of 35 observational studies using street-view imagery to examine greenery/bluespace and mental health, including AI semantic-segmentation methods for measuring exposure. Supports the feasibility of scalable environmental assessment and community comparison; evidence quality was low or very low and does not establish causality.

Bardhan, M., et al. (2024). From space to street: A systematic review of the associations between visible greenery and bluespace in street view imagery and mental health. Environmental Research, 263, 120213. https://doi.org/10.1016/j.envres.2024.120213

Open resource
Continuum
Also relevant to
READ
ResearchNot reviewed

Rising Emissions, Depleting Water and Vanishing Land: UN Scientists: AI Is Threatening Natural Resources for Billions

The article projects that global data centres will use 945 TWh of electricity by 2030, with associated footprints of 399 million tonnes of carbon, 9.3 trillion litres of water and more than 14,500 km² of land. It reports that inference may account for 80-90% of AI energy use, highlights uneven local environmental burdens and recommends measuring and governing AI’s carbon, water and land impacts.

Aczel, M., Chamanara, S., Matin, M. et al. (2026). Rising Emissions, Depleting Water and Vanishing Land: UN Scientists: AI Is Threatening Natural Resources for Billions. United Nations University Institute for Water, Environment and Health (UNU-INWEH). https://doi.org/10.53328/INR26RMA002

Open resource
Continuum
READ
ResearchNot reviewed

Key Questions on Energy and AI

This report examines how the relationship between energy and AI has evolved since the IEA’s 2025 report, amid rising investment in data centres and rapid advances in AI models. It considers where electricity demand is increasing, how quickly grids and supply chains can respond, and the implications for energy security, affordability and sustainability.

International Energy Agency (IEA) (2026). Key Questions on Energy and AI. IEA, Paris. https://www.iea.org/reports/key-questions-on-energy-and-ai

Open resource
Continuum
READ
Practice signalNot reviewed

Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions

The IEA reports that data centre electricity demand rose 17% in 2025, and projects it will double by 2030, while electricity use at AI-focused data centres is expected to triple. It describes supply-chain, planning and grid-connection bottlenecks, and outlines responses including renewables, nuclear, geothermal, onsite gas generation and batteries.

International Energy Agency (IEA) (2026). Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions. News: IEA. https://www.iea.org/news/data-centre-electricity-use-surged-in-2025-even-with-tightening-bottlenecks-driving-a-scramble-for-solutions

Open resource
Continuum
READ
Guidance and standardsNot reviewedIndustry funded

Data Centre Sustainability and Resilience: A Decision Playbook for AI Infrastructure

The report sets out a decision framework for incorporating sustainability and resilience into data-centre development and operations, addressing constraints involving electricity, water, cooling, community acceptance, land and regulation. It estimates that these measures could unlock $700 billion to $1 trillion in planned investment by 2030, while total global data-centre investment could reach $7 trillion by then.

World Economic Forum (2026). Data Centre Sustainability and Resilience: A Decision Playbook for AI Infrastructure. World Economic Forum. https://reports.weforum.org/docs/WEF_Data_Centre_Sustainability_and_Resilience_2026.pdf

Open resource
Continuum
READ
Evidence reviewPreprint, not yet reviewed

AI Data Centers and the Water Use Feedback Loop

The review examines how data centres affect water systems, how water infrastructure can constrain AI development, and how AI might help water systems adapt. Applying its Water Consumption Impact index to ten US data-centre locations, it finds peak consumptive demand ranges from 0.2% to 134% of host utility capacity, and reports no peer-reviewed operational-scale study showing water savings that offset a co-located data centre’s burden.

Akinade, B. A., Amanambu, A. C., Frame, J. M. et al. (2026). AI Data Centers and the Water Use Feedback Loop. arXiv, cs.CY. https://arxiv.org/html/2606.21760v1

Open resource
Continuum
READ
Policy and regulationNot reviewed

Getting it right: Building AI infrastructure that works for Australia: Consultation paper - September 2026

The consultation paper proposes nationally consistent requirements for large data centres and conditions for frontier AI training in Australia, covering energy, water, community impacts, security, local capability and support for Australian creators. It proposes excluding facilities below 30 MW, setting baseline requirements for 30-100 MW facilities and stronger requirements above 100 MW or for portfolios equivalent to 150 MW, amid forecasts of $150 billion in data centre investment by 2030 and electricity use rising from about 5 TWh in 2025-26 to 34 TWh in 2035-36.

Why it matters now: Australian consultation on energy, water and community conditions for large data centres.

Department of the Prime Minister and Cabinet (2026). Getting it right: Building AI infrastructure that works for Australia: Consultation paper - September 2026. Australian Government, Department of the Prime Minister and Cabinet (PM&C). https://www.pmc.gov.au/resources/getting-it-right-building-ai-infrastructure-works-australia

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

AI is dangerous, but not for the reasons you think (TEDWomen 2023 talk)

AI researcher Sasha Luccioni argues that the present-day costs of AI, including carbon emissions, copyright and bias, deserve more attention than speculative future risks, and shows tools for measuring them. About 10 minutes.

Luccioni, S. (2023, October). AI is dangerous, but not for the reasons you think [Video]. TEDWomen 2023. https://www.ted.com/talks/sasha_luccioni_ai_is_dangerous_but_not_for_the_reasons_you_think

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

We're doing AI all wrong. Here's how to get it right (TED Countdown talk)

Luccioni argues that the industry's pursuit of ever-bigger general models wastes energy, and makes the case for smaller, task-specific models that are cheaper, more efficient and more transparent. About 10 minutes, recorded September 2025. Directly relevant to the local and small model discussion in C6.

Luccioni, S. (2025, September). We're doing AI all wrong. Here's how to get it right [Video]. TED Countdown and Bezos Earth Fund. https://www.ted.com/talks/sasha_luccioni_we_re_doing_ai_all_wrong_here_s_how_to_get_it_right

Open resource
Continuum
Also relevant to
WATCH
Opinion and commentaryNot reviewed

AI consumes a lot of water. But why? (TEDAI Vienna talk)

Engineer Shaolei Ren explains how training large models, cooling data centres and manufacturing chips add up to a water footprint that could reach billions of litres a year, much of it drinking water, and calls for transparency and better resource management. About 13 minutes. A clear explanation of a resource cost that rarely appears in mental health discussions.

Ren, S. (2024, October). AI consumes a lot of water. But why? [Video]. TEDAI Vienna. https://www.ted.com/talks/shaolei_ren_ai_consumes_a_lot_of_water_but_why

Open resource
Continuum

X4. AI and the social determinants of distress

How does AI change work, information, inequality and the places people live?

READ
Evidence reviewPeer reviewed

Poverty and Mental Illness

Review of the relationship between poverty and mental illness.

Saraceno, B., & Barbui, C. (1997). Poverty and Mental Illness. The Canadian Journal of Psychiatry, 42(3), 285-290. https://doi.org/10.1177/070674379704200306

Open resource
Continuum
READ
Opinion and commentaryPeer reviewed

The Potential Influence of AI on Population Mental Health

Short, clear framing piece by two population health scientists. Proposes three considerations for anticipating AI's effect on population mental health: AI as an exposure in its own right, AI as a modifier of existing determinants, and AI as a tool for population-level response. A good anchor for this whole category.

Ettman, C. K., & Galea, S. (2023). The Potential Influence of AI on Population Mental Health. JMIR Mental Health, 10, e49936. https://doi.org/10.2196/49936

Open resource
Continuum
Also relevant to
READ
ResearchNot reviewed

How Americans view AI and its impact on people and society

Pew survey of US public attitudes to AI, including concern and acceptance across domains. Useful for user-acceptance context; not healthcare specific.

Pew Research Center. (2025, September 17). How Americans view AI and its impact on people and society. https://www.pewresearch.org/science/2025/09/17/how-americans-view-ai-and-its-impact-on-people-and-society/

Open resource
READ
ResearchPeer reviewed

Toward an experimental ecology of human development.

Bronfenbrenner's ecological model of human development, the basis for nested-systems thinking in public health and psychology.

Bronfenbrenner, U. (1977). Toward an experimental ecology of human development. American Psychologist, 32(7), 513-531. https://doi.org/10.1037/0003-066X.32.7.513

Open resource
Continuum
READ
Opinion and commentaryNot reviewed

Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass

Book examining the often-hidden human labour behind AI and online services, including employment conditions and worker protections. Relevant background on structural determinants of wellbeing; not a 2023 Rosenblat/ILO report or a causal estimate of AI-induced anxiety.

Gray, M. L., & Suri, S. (2019). Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. Houghton Mifflin Harcourt. Catalogue/overview: https://library.wwu.edu/ghost-work-how-stop-silicon-valley-building-new-global-underclass

Open resource
Continuum
READ
Analysis and advocacyNot reviewedAdvocacy organisation

Decolonising AI: A transfeminist approach to data and social justice

A civil-society analysis of AI and automated public-service systems in Latin America. Peña and Varon propose decolonial and transfeminist approaches centred on agency, consent, cooperation, diversity and social justice. Useful for considering whose knowledge and priorities shape community-facing AI, without assuming that a model built elsewhere is universally appropriate.

Peña, P., & Varon, J. (2019). Decolonising AI: A transfeminist approach to data and social justice. In Global Information Society Watch 2019: Artificial intelligence: Human rights, social justice and development (pp. 28–32). Association for Progressive Communications. https://giswatch.org/sites/default/files/gisw2019_web_th4.pdf

Open resource
Continuum
Also relevant to
READ
ResearchPreprint, not yet reviewed

The Early Impacts of AI on Employment among Recent College Graduates

Using CPS data, the authors find no statistically significant summer 2026 increase in unemployment among recent college graduates compared with previous summers, older graduates or young workers without a degree. Including people who wanted a job raised the measured unemployment rate for recent graduates by nearly two percentage points, but still showed no significant increase; models also found some positive relationship between unemployment and occupational remote-work availability.

Fairlie, R. W., Wu, J. (2026). The Early Impacts of AI on Employment among Recent College Graduates. NBER Working Paper 35796, National Bureau of Economic Research. https://doi.org/10.3386/w35796

Open resource
Continuum
READ
ResearchPreprint, not yet reviewed

Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence

Using ADP payroll data through June 2026, the authors find no evidence of widespread job displacement, but employment among workers aged 22-25 in AI-exposed occupations was 19% below the level it might otherwise have reached. The gap appears to reflect reduced hiring rather than increased separations, and the authors describe the findings as early, descriptive indicators rather than causal estimates.

Brynjolfsson, E., Chandar, B. and Chen, R. (2026). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab. https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/

Open resource
Continuum
READ
ResearchNot reviewed

Workers’ Perspectives on Artificial Intelligence: Productivity Gains and Job-loss Fears

Surveys of US household heads found that concern about losing a job because of AI nearly doubled from 5% in December 2024 to just over 10% in December 2025, while concern about job loss or lower wages, or considering alternatives because of difficulty adapting to AI, rose from about 11% to 18%. Workers who strongly perceived productivity gains were more optimistic about job security and 14% more likely to ask for a raise, but they represented only 6% of the sample.

Bracha, A., Hagler, R., and Tang, J. (2026). Workers’ Perspectives on Artificial Intelligence: Productivity Gains and Job-loss Fears. Federal Reserve Bank of Boston, Current Policy Perspectives 26-8. https://www.bostonfed.org/publications/current-policy-perspectives/2026/workers-perspectives-on-ai.aspx

Open resource
Continuum
READ
ResearchNot reviewedIndustry funded

Using AI More Does Not Reassure Workers, Managers Do

Across four waves of Gallup Panel data from 2023 through Q1 2026, covering nearly 30,000 worker observations, frequent AI users were more than twice as likely as infrequent users to fear their jobs would be eliminated within five years. Supportive management was associated with less concern among frequent users, while workers with some job-elimination concern reported lower engagement and job satisfaction, more burnout, and more active job-searching.

Makridis, C. (2026). Using AI More Does Not Reassure Workers, Managers Do. Gallup, Workplace. https://www.gallup.com/workplace/713231/ai-not-reassure-workers-managers-do.aspx

Open resource
Continuum
Also relevant to
READ
ResearchNot reviewed

Global Employment Trends for Youth 2026: Back to the future

The report says global youth unemployment rose to 12.4% in 2025, affecting 67 million young people, while 20.0% of youth, or 257 million people, were not in employment, education or training. It also highlights gender disparities and AI-related job exposure, and calls for integrated policies covering skills, employment services, social protection and technology governance.

Why it matters now: Youth unemployment and AI exposure: a structural driver of distress for young people.

Elder, S., Tobin, S. (2026). Global Employment Trends for Youth 2026: Back to the future. International Labour Office, Geneva. https://doi.org/10.54394/00033788

Open resource
Continuum
READ
ResearchNot reviewed

Beyond Skills and Experience: Psychological Capital in the Modern Workplace

The report summarises a survey of 2,009 employed U.S. adults and finds that workers with high psychological capital more often reported job satisfaction (98% vs. 69%) and outstanding performance (61% vs. 15%), and less often planned to seek a new job within a year (19% vs. 51%), than those with low psychological capital. The cross-sectional, self-reported findings show associations rather than cause and effect.

American Psychological Association (2026). Beyond Skills and Experience: Psychological Capital in the Modern Workplace. American Psychological Association, Work in America™ 2026 report. https://www.apa.org/pubs/reports/work-in-america/2026/full-report.pdf

Open resource
Continuum
Also relevant to
READ
Evidence reviewPeer reviewed

A systematic review of AI induced technostress and human resource development interventions

The review searched Scopus for peer-reviewed journal articles published from 2020 to 2025, identifying 125 records and including 36 studies. It describes how AI-related technostress can affect employees’ psychological health, job satisfaction and work engagement, and highlights training, upskilling, coaching, mentoring and supportive workplace practices as ways to help employees adapt.

Wankhede, V., Khandelwal, K. (2026). A systematic review of AI induced technostress and human resource development interventions. Discover Artificial Intelligence. https://doi.org/10.1007/s44163-026-01855-3

Open resource
Continuum
Also relevant to

Part 5: Curator's preprints

Mark Anns's preprints, kept together. Each is also tagged to the debates it addresses.

Preprints

Preprints by the curator. Not peer reviewed; read them as position papers.

READ Preprint by Mark Anns
Position statementPreprint, not yet reviewed

The Debate That Arrived Too Late: Same Field, Different Realities in Mental Health AI

Mark's preprint argues that asking whether AI should be allowed in mental health is the wrong question: general-purpose assistants and companion platforms already handle mental health conversations at a scale no dedicated service matches, with no clinical governance applied. It identifies two reasons the debate keeps failing. Regulatory frameworks were built for objects with fixed identities, and clinicians at the high-acuity end and public health practitioners at the population end each describe their part of the spectrum as if it were the whole. It proposes a three-domain governance framework spanning population flourishing, clinical practice and the structural conditions that shape mental health. Version 2 (30 March 2026) expands the EU AI Act analysis and replaces the earlier governance leverage framework with two concrete asks: a domestic adverse event reporting obligation modelled on pharmacovigilance, and an internationally coordinated cross-cultural and neurodivergent test case library.

Why it matters now: The hub's central argument in full: the field is arguing about permission while deployment has already happened.

Anns, M. (2026, March 30). The debate that arrived too late: Same field, different realities in mental health AI (Version 2) [Preprint]. SocArXiv. https://doi.org/10.31235/osf.io/hxsnz_v2

Open resource
Continuum
Also relevant to
Related
READ Version 1 (5 March 2026), superseded
READ Preprint by Mark Anns
Position statementPreprint, not yet reviewed

The Same Field, Different Realities: Mental Health's Spectrum Problem and Why Its Central Debates Cannot Be Resolved From Within

Mark's companion preprint takes the spectrum argument beyond AI. Mental health runs from population flourishing and structural determinants through subclinical distress to severe and persistent conditions, and people working at different points use the same words for different things. The paper argues this conceptual failure is field-wide and shows up in funding frameworks that conflate wellbeing programs with clinical services, research that applies clinical outcome measures to population interventions, and governance built on a coherence that does not hold. AI mental health governance is the main case study, since general-purpose systems operate across the whole spectrum at once. It offers a governance-oriented taxonomy for telling spectrum positions apart.

Why it matters now: The underlying argument behind the hub's continuum labels: which part of the spectrum is a claim actually about?

Anns, M. (2026, March 30). The same field, different realities: Mental health's spectrum problem and why its central debates cannot be resolved from within [Preprint]. SSRN. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6498520

Open resource
Continuum
Also relevant to