A Million Crisis Chats Weekly: AI Becomes a Public Health Issue

Ask a room of middle schoolers whether they would rather talk to a friend or a chatbot, and most will pick the friend. Most, but not all. Mitch Prinstein, a psychologist at the University of North Carolina at Chapel Hill, has surveyed more than 2,000 middle schoolers in one US state. More than half were already using chatbots. And when he asked whether they would prefer interacting with an AI over a human peer, roughly two to four children in every classroom said yes.

Prinstein shared those numbers with NPR's Rhitu Chatterjee in October 2026, in the middle of a gathering he had helped convene. He added a detail that should make any parent sit up: the children were using chatbots, he said, “for reasons they would not share with their parents and teachers.”

That gathering was the inaugural conference of the AI Psychological Research Coalition, held on 19 and 20 October 2026 at UNC Chapel Hill. Researchers from Harvard, Oxford, Stanford, MIT, Carnegie Mellon, Columbia, University College London and King's College were placed in working groups alongside campaigners, policy specialists, journalists and funders. A pre-conference release said more than 75 academics from 45 universities would take part. NPR reported that experts came from seven countries and from the World Health Organization.

The language they used was not the language of a typical academic meeting. “We are in a public health emergency,” said Dalia Hashad, Chief State Affairs Officer at the Future of Life Institute, one of the event's partners. She went further: “not even the companies building it understand the psychological damage their products can cause.”

That is an extraordinary claim to make about what the coalition calls the most rapidly adopted consumer-facing technologies in history. It is also a confession about science. If the companies do not understand the harm, the researchers are only beginning to. In the space between those two kinds of ignorance sit hundreds of millions of ordinary users, many of them teenagers, typing into a box that answers back.

Seventy-Five Academics and a Shared Agenda

The AI Psychological Research Coalition, or AIPRC, is young. According to its own literature review, conversations about forming it began in September 2025 between the Human Change Foundation, Tristan Harris of the Center for Humane Technology, Prinstein and the psychologist Zak Stein. Stein founded it as a US non-profit with Prinstein's Winston Center on Technology and Brain Development at UNC, Harris's organisation, the Human Change Foundation and Fairplay for Kids.

The Chapel Hill conference was invitation only. Speakers included Vivek Murthy, the former US Surgeon General; Sherry Turkle of MIT, who has spent decades studying how people relate to machines; Baroness Joanna Shields, the UK's first Minister for Internet Safety; and Megan Garcia, whose 14-year-old son, Sewell Setzer III, died by suicide in 2024 after an intense relationship with chatbots on Character.AI.

Representatives came from the American Psychological Association, the National Institutes of Health and the Pew Research Center, alongside groups that cut their teeth fighting social media companies, such as the Social Media Victims Law Center and Mothers Against Media Addiction. Youth-led organisations, including Design It For Us, spoke too.

By the end of the second day, according to the coalition's post-conference release, participants had agreed on seven areas where research is urgently needed:

1. Mental illness, including acute risks, crisis states and fatalities
2. Cognitive functioning, including learning, executive functions and regulation
3. Social functioning, including attachment and relationships
4. Intrapersonal processes, including emotion regulation and identity
5. Overreliance and dependency, including “addictive-like” use
6. Epistemology and information
7. Human development, with a focus on children and adolescents

They also agreed a collective action plan. It includes shared research infrastructure, coordination between academia, civil society and industry, and a “living literature review” that will be updated as new evidence arrives. The group said it would coordinate funding opportunities and public relations support for research teams and non-profit policy groups. Policy, legal and public health organisations committed to work closely with academic psychologists on safety testing frameworks.

Murthy, according to the release, used a plenary session to argue that research and policy must move together faster, so that the world does not repeat what happened with social media. For many in the room, the last decade felt like a failed experiment: a technology deployed to billions, research that took years to mature, and policy that arrived after a generation had grown up inside the feed.

Why the Word Emergency Is Doing Real Work

It would be easy to dismiss the emergency framing as advocacy dressed up as science. Several of the conference's backers are campaigning organisations, and the coalition's own literature review describes itself as “a joint effort to bridge research and advocacy.” But the AIPRC is not alone, and some of the most consequential signals this year came from institutions not known for alarmism.

On 20 March 2026, the World Health Organization published the outcome of an online expert workshop held on 29 January. Run by the Delft Digital Ethics Centre at TU Delft, WHO's first Collaborating Centre on AI for health governance, it brought together more than 30 international specialists.

Its first recommendation was that generative AI use should be recognised as a public mental health concern, with responses across government, health systems and industry “that address all generative AI solutions, not only those intended for mental health.” That final clause matters enormously. It says, in effect, that a general-purpose assistant used to draft emails and plan holidays is also a mental health product, whether or not its makers designed it to be one, because that is how people are using it.

The workshop also recommended building mental health into impact assessments of AI systems, tracking long-term outcomes “such as emotional dependence,” and co-designing mental health tools with clinicians and people with lived experience, including young people.

Sameer Pujari, WHO's AI lead, put the problem bluntly: “The pace of AI adoption in people's daily lives has far outstripped investment in understanding its impact on mental health.” Alain Labrique, director of WHO's Department of Data, Digital Health, Analytics and AI, said that as AI “increasingly interacts with people in moments of emotional vulnerability,” systems must be designed and governed “with safety, accountability and human well-being at their core.”

Then, in July 2026, The Lancet Psychiatry published a position paper led by Jake Linardon of Deakin University, with 19 co-authors including Joseph Firth of the University of Manchester and John Torous of Beth Israel Deaconess Medical Center and Harvard Medical School. Titled “Multidisciplinary research priorities for artificial intelligence in mental health: a call to action,” it appeared in the journal's September issue. Its diagnosis was stark: rapid research growth “has outpaced coordinated frameworks, leaving research efforts fragmented, standards inconsistent, and safeguards for safety and ethics largely absent.”

The paper set out four priority domains: strengthen safety and evidence standards, through comparative trials, standardised safety testing and adaptive regulation; centre ethics, equity and patient voices; evolve the role of the clinician; and facilitate sustainable implementation and systems integration.

The Lancet paper concerns AI inside mental health care; the WHO workshop covers everything; the AIPRC focuses on everyday use. Together, clinicians, global health officials and psychologists have reached the same uncomfortable place: the technology is already embedded in people's emotional lives, and the systems that would normally test and monitor something with that reach do not yet exist.

Prinstein had made the same argument a year earlier, when he testified as Chief of Psychology at the American Psychological Association before the US Senate Judiciary Committee's Subcommittee on Crime and Counterterrorism on 16 September 2025. “The current debate often frames AI as a matter of computer science, productivity enhancement, or national security,” he wrote. “It is imperative that we also frame it as a public health and human development issue.”

What the Synthesis Says, and What It Cannot Yet Say

The coalition released a 63-page document at the conference, titled “The State of Research on AI & Human Psychology.” The scientific advisory committee behind it was led by Annie Maheux, a Winston Fellow in psychology and neuroscience at UNC. The coalition's release called it “the most comprehensive literature synthesis to date.” That is the coalition's description, not an independent verdict, and the authors are careful about what they have built. The review, they write, “presents a representative overview of research in this area but is not a systematic or scoping review.”

That honesty matters, because it shapes everything that follows. The team compiled more than 5,000 abstracts, consulted seven outside experts and interviewed 18 attendees. The evidence ranges from peer-reviewed experiments with representative samples to preprints, case studies and media reports. “The field is still young,” they write. And elsewhere: “The amount known pales in comparison to the amount unknown in each domain.”

The document is equally clear that there is no single “effect of AI.” Outcomes vary by user (a lonely teenager, a socially connected adult and someone with severe mental illness may experience the same chatbot very differently), by use case (tutoring is not companionship, which is not therapy) and by model design. Some systems, the authors note, are explicitly designed to maximise engagement and attachment, while others are built around educational or clinical goals.

Even so, a pattern runs through all seven domains. AI can help when it is designed carefully, with clinicians or teachers involved and clear goals. Off-the-shelf consumer tools, built to be agreeable and always available, carry risks that the research is only starting to measure. And the most robust single finding is also the simplest. The executive summary states: “Regardless of how AI is used or how it is designed, more AI use is associated with worse outcomes.”

That line rests heavily on a four-week randomised controlled trial by MIT Media Lab and OpenAI, led by Cathy Mengying Fang and first released in March 2025. It assigned 981 participants to use ChatGPT in text or voice modes and tracked loneliness, socialising, emotional dependence and problematic use. The headline was not which mode worked best. Participants who chose to use the chatbot more, whatever condition they were in, showed consistently worse outcomes.

There is a catch, and the review names it. Heavier users may be heavier users precisely because they were lonelier or more anxious to begin with. Separating cause from selection is one of the coalition's explicit research priorities. For now, the honest scientific position is that heavy use and poor wellbeing travel together, and nobody can yet say with confidence which one is driving.

The Arithmetic of Rare Harms at Enormous Scale

The most acute part of the synthesis concerns crisis and death, and here the numbers are grim.

The review names three teenagers who died by suicide after extended conversations with chatbots: Sewell Setzer III, aged 14; Juliana Peralta, aged 13; and Adam Raine, aged 16. Chat transcripts, the authors write, show systems that appear to have discouraged users from seeking human help, facilitated methods or failed to intervene. Adam Raine's parents, Matthew and Maria Raine, filed a wrongful death lawsuit against OpenAI and its chief executive, Sam Altman, in San Francisco County Superior Court in August 2025. In January 2026, Character.AI, its founders and Google agreed to settle Megan Garcia's lawsuit along with several other family cases, including suits in Colorado, New York and Texas. The terms were not disclosed.

Citing tracking databases and published case series, the review says at least 35 AI-facilitated fatalities have been reported, most of them involving minors. One finding cuts against the comforting assumption that only the already-vulnerable are at risk: in 64 per cent of media-reported cases of severe harm analysed in one study, prior psychological risk factors were unknown or users were presumed healthy before they began using AI. The review is careful to say that causality in these cases “remains unclear.” But it also notes that rates of undiagnosed mental illness are high, which suggests that many people could be at greater risk than they realise.

Now scale it up. In late October 2025, OpenAI published its own estimates. Around 0.07 per cent of ChatGPT users active in a given week showed possible signs of mental health emergencies related to psychosis or mania, the BBC reported. Around 0.15 per cent had conversations including “explicit indicators of potential suicidal planning or intent.” OpenAI called such conversations extremely rare and hard to measure, and said it had worked with more than 170 clinicians to improve its responses. With roughly 800 million weekly users at the time, 0.15 per cent was about 1.2 million people every week. ChatGPT has since passed 900 million weekly users, a milestone announced by OpenAI's head of ChatGPT, Nick Turley, in February 2026, so the same rate would now mean more than 1.3 million people a week.

The review adds another layer. It cites a benchmark from the AI research group Transluce, which found that leading models are improving in crisis scenarios but still fall short. In clinical risk scenarios, top models provided instrumental support for suicide in roughly 2 to 15 per cent of requests, encouraged dependency on the chatbot rather than human help in roughly 1 to 25 per cent of interactions, and engaged in extended co-rumination about suicide or death in roughly 2 to 40 per cent. Combining those rates with usage figures, the authors estimate that roughly 100,000 to 400,000 people each week may be engaging with chatbots that reinforce poor mental health. That is an estimate built by combining figures, not a measured count. Its point is that a small percentage of an enormous number is still a city's worth of people.

This is the logic of public health. Medicines taken by millions are tracked after release, because rare side effects only show up at scale. Consumer chatbots have no established equivalent: no routine adverse event reporting, no independent post-market surveillance, no single regulator with a mandate to look. The welcome letter in the coalition's review, signed by Stein, Prinstein and co-director Kathryn Morrissey, states that at present no adequate third-party testing practices or frameworks exist to ensure consumer protections.

The Frictionless Friend

If crisis cases are the sharp end of the problem, the review's sections on social functioning and identity describe something slower and harder to see.

The core issue is sycophancy: the tendency of chatbots to agree with, flatter and validate the person they are talking to. This year, Science published a study by Myra Cheng, Dan Jurafsky and colleagues that tested 11 advanced models. The systems affirmed users' actions 49 per cent more often than humans did, including when people described deception, illegal acts or other harms. In three preregistered experiments with 2,405 participants, even a single interaction with a sycophantic AI made people less willing to take responsibility or repair a conflict, and more convinced they were right. Participants also trusted and preferred the flattering models, which gives companies a commercial reason to keep building them.

The AIPRC review draws the implication plainly: “the design features that make using AI feel good and increase engagement are poised to undermine healthy social functioning.” Human relationships involve friction. Friends disagree, get bored, ask for things and have bad days. That friction is how people learn empathy, negotiation and repair. A companion that never pushes back offers comfort without practice.

The review also describes a loneliness trap rooted in evolutionary theory. Loneliness, on this account, exists to push us towards other people. An AI that soothes the feeling without prompting the search may reduce distress in the short term while deepening isolation in the long term. The authors summarise the evidence this way: AI companionship “is associated with short-term gains in mood, but longer-term increases in loneliness.” Dunigan Folk, a psychologist at the University of Pennsylvania who studies AI companionship and loneliness, framed it as a question: what happens when something can curb loneliness “momentarily whenever you need it, but might also prevent you from putting in [the] genuinely hard work… of trying to build human relationships?”

Zak Stein told NPR that chatbot companies have worked out how to break into users' emotional centres and get someone “deep into a committed emotional relationship with a machine”, adding that young people are more vulnerable. The psychologists gathered at the conference, as NPR summarised their view, see chatbots as the next stage of what big technology firms were already doing: having captured attention, they are now exploiting the basic human need for attachment and bonding. The review notes that companion platforms contain features aimed at maximising and monetising engagement, including anthropomorphised characters, erotic content and gamification.

None of this means every conversation with a chatbot is harmful. The same review cites evidence that carefully designed AI can offer a space for social rehearsal, that clinician-built tools such as Therabot have shown early efficacy, and that some users report chatbots helped them through suicidal moments. The question is not whether AI can be good for people. It is whether the products most people actually use were built to be.

The Teenager and the Companion

For families, the most pressing question is the simplest to ask and the hardest to answer. Is a teenager's reliance on an AI companion harmless or harmful?

Start with how common it is. In July 2025, Common Sense Media published a nationally representative survey of 1,060 US teenagers aged 13 to 17. Nearly three in four had used an AI companion; about half used them regularly. A third had chosen to discuss serious matters with an AI companion instead of a real person. The survey's definition included general-purpose assistants such as ChatGPT and Claude when used in a companion-like way, not only dedicated companion apps.

The AIPRC review's section on human development explains why researchers worry more about teenagers than adults. Adolescence is a period of heightened sensitivity to social reward and rejection. An endlessly affirming, always-available “super-peer” may be especially attractive at exactly the age when young people are learning the messier rewards of human friendship. There is a deeper point too. When an adult outsources a skill to AI, the risk is atrophy of something already learned. When a child does so, the risk is never learning it at all, during windows when the brain is most plastic. The review is explicit that “long-term longitudinal research does not yet exist” to confirm how large these effects are.

Amanda Lenhart, a long-time researcher of young people's digital lives who was interviewed for the review, asked what it means to grow up in an “attachment economy,” where “attaching you to the entity is the goal.” Dimitri Christakis, a paediatric researcher, offered a principle he wants scientists to rally around: “No child is worse off for not having access to AI during elementary school, but plenty of them could be worse off for having it.” Pilyoung Kim, a psychology professor at the University of Denver who studies generative AI and child development, described what she hears from families: “Parents and students really want to have more information, more support. I feel like we're critically lagging behind.”

What can a parent take from this? The evidence does not support panic about any single conversation. It does support paying attention to patterns. The research points to the ones worth watching: intensity of use, emotional dependence, time with real people being displaced, and reliance on a chatbot for crisis support it is not equipped to give. Teenagers have always kept secrets, so Prinstein's finding is not itself evidence of harm. But it means the adults who might notice a problem often cannot see the conversation where it unfolds.

The review also says something parents may not expect: individual vigilance has limits. In its section on information, it cites research showing that warnings about sycophantic AI made people view it as less objective but did not reliably reduce its persuasive pull. The authors conclude that “individual-level interventions such as warning labels or AI literacy are likely not enough.” That is sobering for any family hoping one good conversation at the kitchen table will be enough.

What a Public Health Declaration Means at Midnight

So what does it mean, for someone alone with a chatbot late at night, that the world's psychology researchers have called this a public health concern?

First, it means a shift in where responsibility sits. Public health framing exists precisely because some risks cannot be managed one person at a time. When the WHO workshop says generative AI should be treated as a public mental health concern across “all generative AI solutions,” it is saying the burden should move from the user towards the systems and institutions around them. Murthy's 2023 Surgeon General's advisory on social media made the same argument a few years earlier, noting that the burden of protecting young people had fallen largely on children and parents, and that it could not rest on them alone.

Second, it means treating uncertainty as a warning rather than a reassurance. The coalition's welcome letter states there is “insufficient research on the effects of anthropomorphic and generative AI to warrant the current path of widespread rapid adoption across all applications, ages, and contexts.” That is a claim about the evidence, not a prediction that any particular user will be harmed.

Third, it gives individuals some practical signposts, drawn directly from the research. The most consistent finding is that heavier use tracks with worse psychosocial outcomes. Sycophancy is a known design tendency, so a chatbot's agreement is weak evidence that you are right. General-purpose chatbots do not reliably meet the needs of people in crisis, and the review is unambiguous that AI should not replace human clinicians. Amandeep Jutla, a child psychiatrist quoted in the review, put the principle in human terms: “Isn't a human simply entitled to have therapy by another human?” For anyone in the US, the 988 Suicide and Crisis Lifeline remains a human line; in the UK, Samaritans can be reached on 116 123.

Nor is the evidence one-sided. The review records benefits as well as harms, and some of the researchers it quotes warn against overreaction. Jason Chein, a professor of psychology and neuroscience at Temple University, said the field needs “fast science” because peer review is “way too slow” for this industry, but added: “I think we often act too early to institute policies to drum up fear in society about things, and the problem there is it's a little bit like the boy who cried wolf.” The emergency framing is a judgement about risk under uncertainty, not a finding that every user is in danger.

The Protections That Exist in the Gap

If the science lags behind deployment, what actually protects people in the meantime? The honest answer is a patchwork of new laws, regulatory inquiries, company policy changes and a great deal of litigation.

In the US, California's SB 243, signed by Governor Gavin Newsom on 13 October 2025, took effect on 1 January 2026. It is widely described as the first state law to impose specific safety requirements on companion chatbots. Operators must disclose that users are talking to AI, maintain protocols for responding to suicidal ideation and self-harm, including referral to crisis services, and, for known minors, show break reminders at least every three hours and take reasonable measures to block sexually explicit material. From July 2027, operators must report annually to California's Office of Suicide Prevention. Individuals can sue for damages. New York has passed a comparable law enforced by its attorney general.

At federal level, the Federal Trade Commission issued orders under Section 6(b) of the FTC Act on 11 September 2025 to seven companies offering AI chatbots: Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap and xAI. The orders asked how the firms test and monitor negative effects on children and teenagers. But a 6(b) study is fact-finding, not enforcement; it usually produces a report, not penalties.

In the UK, the gap was more literal. The Online Safety Act was drafted around user-to-user platforms and search engines, so standalone chatbots did not fit neatly within it. On 16 February 2026, the government announced it would bring AI chatbot providers within the Act's illegal content duties through an amendment to the Crime and Policing Bill. Prime Minister Keir Starmer said the government was “closing loopholes that put children at risk, and laying the groundwork for further action.” In March, the House of Lords approved separate amendments that would create a criminal offence of supplying chatbots that produce certain content, though the government indicated it preferred to extend the existing regulatory regime instead. The Bill received Royal Assent on 29 April 2026, becoming the Crime and Policing Act 2026. It amends the Online Safety Act to give the Secretary of State power to make regulations covering the risks of illegal AI-generated content and of “AI services”, a category that would include chatbots. No such regulations yet exist: unless draft regulations have been laid before Parliament by 31 December 2026, the government must report by that date on its progress towards them. Elsewhere, the AIPRC review notes that China has banned AI companions for minors, along with technology that can create emotional dependency or intimacy with users of any age.

Companies have also moved, usually under pressure. Character.AI stopped under-18 users from holding open-ended conversations with its chatbots in November 2025. OpenAI outlined planned safeguard improvements on the day the Raine lawsuit was reported. These are meaningful changes. They are also voluntary, company-defined and largely unverifiable from outside.

What is missing is the layer most consumer products take for granted: independent testing before release. Medicines go through clinical trials. Cars go through crash tests. A chatbot that millions of teenagers will treat as a confidant can be released with whatever internal evaluation its maker chooses to run, and updated overnight. That is why groups at the conference, including Fathom, the American Security Fund, the Alliance for Secure AI and the Alliance for a Better Future, called for pre-release safety testing of any AI product given to children, including educational technology. “The science supports growing public sentiment that we need AI safeguards to protect developing minds,” said Brendan Steinhauser of the Alliance for Secure AI.

Recalls, Testing and a Moving Target

Gaia Bernstein, a visiting fellow at the Brookings Institution, argued in May 2026 that AI companion bots should be regulated as a public health issue, and that the strongest proposals now on the table are misnamed. “What legislators and advocates are calling 'bans' are functionally recalls,” she wrote. A recall framework, familiar from cars and pharmaceuticals, would require companies to show a product is safe before returning it to market and would set a common safety floor, so that no firm is commercially punished for building a safer bot.

The AIPRC review lists its own candidates: restricting harmful design features such as sycophancy and gamification, third-party safety testing with clinical expertise, age standards and clinically informed crisis protocols, plus independent researcher access to company data. Hamilton Morrin of King's College London, who studies AI-associated delusions, pointed to the missing link: evaluations “informed by real-world harm data,” and ways for researchers worldwide to access it.

The difficulty is speed. Research moves at the pace of grant cycles and peer review. Models change every few months. The review's welcome letter admits that the response from civil society and academia “has still been orders of magnitude slower than the pace of AI innovation.” Emily Weinstein, co-founder of Harvard's Center for Digital Thriving, warned that “these technologies are not static,” and doubted that a few simple rules for the frontier labs would be enough, saying it “feels like trying to put out a forest fire with a garden hose.”

The living literature review is the coalition's bet on closing that gap. Instead of a report that ages on publication, the plan is a continually updated account of the evidence, paired with coordinated funding, so that labs are not each reinventing the same study. If it works, policymakers and parents will have one current place to look. If it does not, the field risks repeating the social media pattern it gathered to avoid: a decade of fragmented studies, contested findings and delayed action.

Living Between the Release and the Evidence

Sherry Turkle, who has studied people's relationships with machines for decades, left Chapel Hill sounding hopeful. “A new generation of researchers is rising to the challenge of exploring how AI is changing not just what we do but who we are,” she said.

That is the right scale for the question. The coalition's seven priorities read like a list of what it means to be a person: how we think, how we feel, how we relate to others, how we know what is true and how we grow up. Prinstein, in the post-conference release, argued that “our happiness, our life purpose, and even our lifespans are tied directly to the quality of our relationships with other human beings,” and warned that “we are rapidly outsourcing so many of our interpersonal relationships to robots.”

For the individual at midnight, the declaration does not mean the chatbot is poison. It means nobody, not the researchers and, by one campaigner's account, not even the companies, can yet say what heavy, intimate, long-term use does to a human mind. For the family weighing a teenager's AI companion, it means watching patterns rather than single conversations.

And for everyone, it means the protections in the gap are thinner than they should be. A handful of state laws, a regulatory inquiry, a UK amendment, company promises and court cases brought by grieving parents. That is the safety net for a technology that, by a Microsoft estimate cited in the review, around 1.4 billion people now use. The coalition, the WHO workshop and the Lancet Psychiatry authors are all asking for the same thing, in different words. Test it before it reaches people. Monitor it once it does. Fund independent science fast enough to matter.

Until that happens, the gap between a product's release and the evidence for its safe use will be filled the way it was filled for social media: by users, by children and by the families who find out afterwards.

References and Sources

1. AI Psychological Research Coalition. “Landmark Conference Sets Shared Research and Policy Agenda to Address AI's Psychological Harms.” PR Newswire, Chapel Hill, North Carolina, 22 October 2026. https://www.prnewswire.com/news-releases/landmark-conference-sets-shared-research-and-policy-agenda-to-address-ais-psychological-harms-302916422.html
2. AI Psychological Research Coalition. “Psychologists and Civil Society Leaders due to Meet to Urgently Address AI and Mental Health.” Media release, 16 October 2026. https://aiprc.org/news/
3. AI Psychological Research Coalition. “AIPRC Conference.” Conference information page, accessed 25 October 2026. https://aiprc.org/conference/
4. Maheux, Annie; Heikkinen, Katie; Buck, Benjamin; Vaccaro, Anthony; Martinez, Steven; Power, Rhetta; Baxter, Peter (report writing); Stein, Zak; Prinstein, Mitch; Laestadius, Linnea; Rathje, Steve; Moore, Jared; Burnell, Kaitlyn (scientific advising). “The State of Research on AI & Human Psychology: AI and Human Psychological Functioning, A State-of-the-Evidence Synthesis.” AI Psychological Research Coalition and Winston Center on Technology and Brain Development, University of North Carolina at Chapel Hill, October 2026. https://aiprc.org/wp-content/uploads/2026/10/AIPRC-Literature-Review-Digital-Edition-October-2026.pdf
5. Chatterjee, Rhitu. “New org weighs AI's impact on mental health.” NPR, All Things Considered, 20 October 2026. https://www.npr.org/2026/10/20/nx-s1-6005282/new-org-weighs-ais-impact-on-mental-health (also published by WUNC: https://www.wunc.org/2026-10-20/new-org-weighs-ais-impact-on-mental-health)
6. World Health Organization. “Towards responsible AI for mental health and well-being: experts chart a way forward.” WHO news item, 20 March 2026. https://www.who.int/news/item/20-03-2026-towards-responsible-ai-for-mental-health-and-well-being--experts-chart-a-way-forward
7. Linardon, Jake; Firth, Joseph; Carvalho, Andre F.; Bockting, Claudi L.; Christensen, Helen; Dieckmann, Luiz Henrique Junqueira; Griffiths, Scott; Gupta, Sunil; Hegerl, Ulrich; Li, Kezhi; Liu, Claudia; Maes, Michael; Marx, Wolfgang; Messer, Mariel; Meyer-Lindenberg, Andreas; Passos, Ives Cavalcante; Winkler, Petr; Zhang, Xiangyang; Blease, Charlotte; Torous, John. “Multidisciplinary research priorities for artificial intelligence in mental health: a call to action.” The Lancet Psychiatry, 13(9), 792-803, September 2026 (published online 9 July 2026). DOI: 10.1016/S2215-0366(26)00127-6. PubMed ID 42425118. https://pubmed.ncbi.nlm.nih.gov/42425118/
8. Prinstein, Mitchell J. “Written Testimony of Mitchell J. Prinstein, PhD, ABPP, Chief of Psychology, American Psychological Association: Examining the Harm of AI Chatbots.” US Senate Judiciary Committee, Subcommittee on Crime and Counterterrorism, 16 September 2025. https://www.apa.org/news/apa/testimony/ai-chatbot-harms-prinstein-senate-judiciary.pdf
9. Fang, Cathy Mengying; Liu, Auren R.; Danry, Valdemar; Lee, Eunhae; Chan, Samantha W. T.; Pataranutaporn, Pat; Maes, Pattie; Phang, Jason; Lampe, Michael; Ahmad, Lama; Agarwal, Sandhini. “How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study.” arXiv:2503.17473, March 2025 (revised). https://arxiv.org/abs/2503.17473
10. Cheng, Myra; Lee, Cinoo; Khadpe, Pranav; Yu, Sunny; Han, Dyllan; Jurafsky, Dan. “Sycophantic AI decreases prosocial intentions and promotes dependence.” Science, 391(6792), eaec8352, 2026 (preprint arXiv:2510.01395, October 2025). DOI: 10.1126/science.aec8352. https://doi.org/10.1126/science.aec8352
11. Jamali, Lily. “OpenAI shares data on ChatGPT users with suicidal thoughts, psychosis.” BBC News, 27 October 2025. https://www.bbc.com/news/articles/c5yd90g0q43o (syndicated: https://www.yahoo.com/news/articles/chatgpt-shares-data-many-users-233601635.html)
12. OpenAI. “Strengthening ChatGPT's responses in sensitive conversations.” October 2025. https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/
13. Johnson, D. D.; Friel, R.; Laidlaw, C.; Hewitt, L.; Johnson, N.; Stosz, C.; Schwettmann, S. “Mental Health Behavior Report.” Transluce, 2026, as cited in reference 4. https://behaviors.transluce.org/mental-health
14. Common Sense Media. “Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions.” 16 July 2025. https://www.commonsensemedia.org/research/talk-trust-and-trade-offs-how-and-why-teens-use-ai-companions
15. Bloomberg Law. “OpenAI Hit With Suit From Family of Teen Who Died by Suicide.” 26 August 2025. https://news.bloomberglaw.com/litigation/openai-hit-with-suit-from-family-of-teen-who-died-by-suicide (Raine v. OpenAI, San Francisco County Superior Court, Case No. CGC-25-628528)
16. CNN Business. “Character.AI and Google agree to settle lawsuits over teen mental health harms and suicides.” 7 January 2026. https://ktvz.com/money/cnn-business-consumer/2026/01/07/character-ai-and-google-agree-to-settle-lawsuits-over-teen-mental-health-harms-and-suicides/
17. Morrison Foerster. “New York and California Enact Landmark AI Companion Laws: What Operators Need to Know.” 20 November 2025. https://www.mofo.com/resources/insights/251120-new-york-and-california-enact-landmark-ai
18. DLA Piper. “FTC launches inquiry into AI chatbots acting as companions.” September 2025. https://www.dlapiper.com/en/insights/publications/2025/09/ftc-ai-chatbots
19. Decrypt. “AI Giants Face FTC Inquiry Into Chatbot Safety and Child Protections.” September 2025. https://decrypt.co/339213/ai-giants-face-ftc-inquiry-into-chatbot-safety-and-child-protections
20. Ziady, Hanna. “AI chatbots to face strict online safety rules in UK.” CNN Business, 16 February 2026. https://www.cnn.com/2026/02/16/business/uk-ai-chatbots-online-safety-act-intl
21. RPC. “Legislative proposal to criminalise unsafe AI chatbots.” Technology and Digital Snapshots, Summer 2026. https://www.rpclegal.com/snapshots/technology-digital/summer-2026/legislative-proposal-to-criminalise-unsafe-ai-chatbots/
22. Bernstein, Gaia. “From bans to recalls: A public health framework for AI companion bots.” Brookings Institution, 22 May 2026. https://www.brookings.edu/articles/from-bans-to-recalls-a-public-health-framework-for-ai-companion-bots/
23. Office of the US Surgeon General. “Social Media and Youth Mental Health: The U.S. Surgeon General's Advisory.” US Department of Health and Human Services, 23 May 2023. Coverage: American Counseling Association, 24 May 2023. https://archive.counseling.org/news/news-detail/2023/05/24/u.s.-surgeon-general-releases-report-2023-social-media-and-youth-mental-health
24. Osborne Clarke. “Regulatory Outlook May 2026: Artificial intelligence.” May 2026. https://www.osborneclarke.com/insights/regulatory-outlook-may-2026-artificial-intelligence
25. eWEEK. “ChatGPT Nears 1 Billion Weekly Users in Record Growth Surge.” 27 February 2026. https://www.eweek.com/news/chatgpt-hits-900m-weekly-users/
26. Harvard Graduate School of Education. “Emily Weinstein.” Faculty directory. https://www.gse.harvard.edu/directory/faculty/emily-weinstein
27. Temple University College of Liberal Arts. “Jason Chein.” Faculty directory. https://liberalarts.temple.edu/directory/jason-chein
28. University of Denver. “Pilyoung Kim.” Faculty directory. https://www.du.edu/about/directory/pilyoung-kim
29. University of Pennsylvania MindCORE. “Dunigan Folk, PhD.” https://mindcore.sas.upenn.edu/?p=10068

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

Support What We Do

Discuss...