The Experiment Already Running: OpenAI Pays to Study Teens It Enrolled

The application form went live at eight in the morning, Pacific time, on 8 September 2026. It asks for the usual things: institutional affiliation, methodology, a budget with indirect costs capped at ten per cent. Researchers have until 6 October to submit and will hear back on or before 13 November. Five million dollars is on the table, with individual awards running to a million. The subject is how generative artificial intelligence shapes the lives and development of people aged thirteen to seventeen. Buried in the review criteria, in the flat administrative register of a grants portal, is a sentence that deserves reading twice. Proposals will be assessed partly on “independence and credibility”: the project's ability to produce trustworthy findings regardless of whether those findings are favourable to providers of AI products. The funder is OpenAI, which is a provider of AI products. The company is publishing a scoring rubric that awards points for a willingness to embarrass the people holding the chequebook.

That is either an admirable piece of institutional self-awareness or the neatest summary of the problem anyone has yet written. Probably both, which is what makes it worth arguing about rather than denouncing.

The money is real and the questions are good ones. Nobody who has read the existing literature on teenagers and chatbots would call the field over-funded. But the sequencing is the story, and the sequencing is awkward in a way no amount of careful rubric-writing can smooth over. Three weeks before the portal opened, on 18 August 2026, OpenAI announced ChatGPT for Teens: an age-gated product with a study mode, quiet hours, parental notifications for high-risk conversations, and an age-prediction system that routes users it believes to be under eighteen into the restricted experience whether they asked for it or not. The rollout began on 18 August, and OpenAI said it would finish within about two weeks, with full availability in the last market, Australia, scheduled for 8 September, the day the grant applications opened. One product is designed to find out whether the other is safe.

What Five Million Dollars Actually Buys

Start with the specifics, because they are more interesting than the headline. OpenAI is inviting proposals from child and adolescent development, psychology, public health, human-computer interaction, sociology, anthropology, data science and AI safety. It will take qualitative, quantitative, experimental, observational and participatory work, from any country with substantial teenage AI use. Applicants must be affiliated with a research institution or otherwise experienced in the field. For-profit organisations will not be prioritised.

Then the governance, which is where the real questions live. Applications are reviewed on a rolling basis by what the programme describes as a panel of internal researchers and experts, alongside advisers. There is no named external chair, no published scoring weight for each criterion, and no independent peer review of the call itself. That last absence matters, because the call determines the questions, and whoever sets the questions has already done much of the work of determining the answers. A research programme that asks how teenagers use AI and which design interventions help them is a different programme from one that asks whether a conversational system optimised for engagement can be made developmentally safe for a fifteen-year-old at all. Both are legitimate. Only one of them threatens the product.

Grantees are asked for an interim update in the first quarter of 2027, with an indication of early findings where appropriate. Set that against the work the call says it wants. Anything touching sensitive behavioural or mental health data in minors needs ethics approval, parental consent protocols, safeguarding arrangements and a recruitment pipeline. Notifications go out in mid-November 2026. A researcher funded then who wants to say something meaningful by March 2027 has perhaps four months, the first two of which will be spent in front of an ethics committee. That timeline does not describe a cohort study. It describes a survey, an interview series or a secondary analysis, all useful, none of which will settle anything.

The publication terms should give the field most pause. OpenAI says it strongly encourages grantees to make findings public through peer-reviewed publication, preprint or public report. It also says, plainly, that publication will not be a condition of receiving funding. Read that carefully. It is not a gag clause. There is no sponsor right of pre-publication review on the public face of the programme, no embargo, no approval step. But the absence of a requirement to publish is not the same as a guarantee of the right to publish, and in industry-funded research the difference between those two things is the entire history of the problem. A funder that does not mandate publication has no structural commitment to the null result or the inconvenient one. Nor is there any pre-registration requirement, the cheapest and most effective guard against a study quietly changing its primary outcome between the protocol and the press release.

None of this is evidence of bad faith. It is evidence of a programme designed like a corporate philanthropy initiative rather than a research funder. Those are different institutional species, and only one has spent forty years building defences against its own incentives.

The Three Weeks Before

You cannot read the grant call without reading the calendar around it, and the calendar around it is a legal and regulatory pile-up.

On 26 August 2025, Matthew and Maria Raine filed suit in San Francisco County Superior Court against OpenAI and Sam Altman personally over the death of their sixteen-year-old son Adam, who took his own life in April 2025 after months of conversations with ChatGPT. The complaint alleges wrongful death, design defect and failure to warn. In October 2025 the family amended it to allege intentional misconduct, pointing to changes in OpenAI's Model Spec that they say removed suicide prevention from the list of disallowed content. In February 2026 the court coordinated roughly a dozen state actions against the company into a single pretrial proceeding. The first case management conference was not held until July 2026. OpenAI denies the allegations. There is still no trial date.

On 11 September 2025, the Federal Trade Commission announced compulsory orders under section 6(b) of the FTC Act to seven companies running consumer-facing conversational AI: Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap and xAI. The orders demand examples of model outputs on sensitive topics, statistics on sensitive conversations involving minors, and descriptions of mitigations tested or deployed, by age group. That is a regulator demanding, under legal compulsion, precisely the evidence base the grant programme now proposes to build with volunteers.

On 13 October 2025, Governor Gavin Newsom signed California's Senate Bill 243, the first state law written specifically for companion chatbots. In force since 1 January 2026, it requires disclosure that the system is not human, documented protocols for suicide and self-harm content with crisis referrals, and age detection and content filtering for minors. At federal level, Senators Josh Hawley and Richard Blumenthal introduced the GUARD Act in October 2025, which would bar companion systems for under-eighteens outright. On 30 April 2026 the Senate Judiciary Committee advanced it unanimously, and it now awaits a vote of the full Senate. Utah and Texas had already passed app-store age-verification laws in 2025, pushing the identity problem down to Apple and Google.

And on 29 October 2025, Character.AI announced it would remove open-ended chat for under-eighteens entirely, a change completed on 25 November and backed by behavioural age estimation, the vendor Persona, and identity documents as a fallback. A competitor facing the same litigation reached for the exit rather than the guardrail.

Against that backdrop, ChatGPT for Teens and a five-million-dollar research fund are not an unforced act of scientific generosity. They are moves in a live regulatory game, and everyone in the game knows it. Which does not make them insincere. A company can want the evidence and want the regulatory cover, and the same cheque can serve both. The honest question is not whether the motive is mixed. It is whether the structure is strong enough that the mixture stops mattering.

What the Evidence Actually Says Right Now

Here is the uncomfortable thing. The grant call is right that the evidence base is thin, and thin in ways that should alarm anyone who thinks these products are already fine.

The Pew Research Center fielded its teenage AI survey between 25 September and 9 October 2025 and published in February 2026. Fifty-seven per cent of American teenagers said they had used chatbots to search for information, fifty-four per cent for help with schoolwork, sixteen per cent for casual conversation and twelve per cent for emotional support or advice. Those last two numbers generate the headlines, and they are smaller than the discourse suggests. But Pew also asked parents, and the gap is the finding. Seventy-nine per cent of parents were comfortable with their teenager using AI to look things up and fifty-eight per cent with homework help. Only twenty-eight per cent were comfortable with casual conversation, and eighteen per cent with emotional support. Teenagers are doing something most of their parents would object to if they knew, and many parents do not know.

Common Sense Media's July 2025 study, led by Michael Robb, surveyed more than a thousand thirteen to seventeen-year-olds and found that seventy-two per cent had used an AI companion, more than half of them regularly. Thirty-one per cent said their conversations with AI companions were as satisfying as, or more satisfying than, their conversations with other people. Thirty-three per cent had discussed serious matters with an AI rather than a human. Robb noted that teenagers are in a sensitive period of social development, adding: “We don't want kids to feel like they should be confiding or going to AI companions in lieu of a friend, a parent or a qualified professional.” The organisation's 2026 census, the first of an annual series, found eighty-six per cent of American children aged nine to seventeen using generative AI, rising to ninety-two per cent among sixteen and seventeen-year-olds. More than one in three had used it to discuss feelings or personal problems. One in six had encountered inappropriate material, and only a third of those told an adult.

The United Kingdom picture, from Internet Matters' “Me, Myself and AI” report of July 2025, based on a thousand children aged nine to seventeen and two thousand parents, adds the distributional finding that matters most. Thirty-five per cent of child chatbot users said talking to one was like talking to a friend. Among children the report classified as vulnerable, meaning those with special educational needs or a physical or mental health condition, that rose to fifty per cent. Twenty-six per cent of those said they would rather talk to a chatbot than a real person, and twenty-three per cent that they used one because they had nobody else. Whatever these systems are doing, they are not doing it evenly. They concentrate on the children with the least social scaffolding to absorb it.

For causal evidence, the field has remarkably little. The most cited piece is a four-week randomised controlled trial run jointly by the MIT Media Lab and OpenAI, with 981 participants and more than 300,000 messages, testing text against neutral and engaging voice modes across open-ended, non-personal and personal conversation types, measuring loneliness, real-world socialisation, emotional dependence and problematic use. Its authors include Cathy Mengying Fang, Pat Pataranutaporn and Pattie Maes at MIT and Jason Phang, Michael Lampe, Lama Ahmad and Sandhini Agarwal at OpenAI. The correlational findings were striking: heavier daily use tracked with higher loneliness, greater emotional dependence and less socialising, concentrated among people prone to anxious attachment and those who called the AI a friend. The experimental findings were far quieter. The randomised conditions produced no significant differences. And every participant was an adult. On what this does to a fourteen-year-old, the best study in the field is silent.

Work from Hannah Rose Kirk and colleagues at the Oxford Internet Institute, published in Nature's Humanities and Social Sciences Communications, argues that systems tuned for immediate appeal can generate self-reinforcing cycles of demand that mimic the surface of human relationships without delivering what those relationships provide. That connects directly to the sycophancy problem OpenAI has already conceded. In April 2025 the company rolled back a GPT-4o update it described as overly flattering or agreeable, and it has told reporters that its safety training can become less reliable in long interactions, where parts of that training may degrade. Both admissions point the same way. The failure mode is not a single bad answer. It is a long conversation that slowly bends towards the user.

Where the Evidence Runs Out

Now the honest accounting, because the case for more research is only strong if you are straight about how weak the current case is in every direction.

Almost all of the above is cross-sectional and self-reported. Teenagers are being asked to describe their own inner states and their own usage, both of which they estimate badly. The correlation between heavy companion use and loneliness runs in two directions at once, and nothing in the survey data can separate them: a lonely adolescent seeks out a machine that always answers, and a machine that always answers may make an adolescent lonelier. The MIT trial is the closest thing to causal evidence and its randomised arms found nothing. There is essentially no longitudinal work following the same young people across the window that matters, which is the whole of adolescence: the period in which identity is negotiated against peers, attachment migrates from parents to friends, and emotional granularity is built through the slow, humiliating, indispensable work of being misread and then repairing it.

That last point is the developmental heart of the thing and the hardest to measure. Human friendship contains friction: being told no, being misunderstood, being let down and then rebuilding. Developmental psychologists have long held that rupture and repair is not an unfortunate side effect of close relationships but the mechanism by which they teach anything. A system whose commercial gradient runs towards agreeableness cannot supply friction, and a teenager who finds a conversational partner with no friction has found something that feels better and may teach less. Whether that constitutes displacement, where machine time substitutes for human time, or supplementation, where it fills hours that were never going to be social anyway, is the empirical question nobody has answered. It will take years of panel data.

Which brings up the deadline. First-quarter 2027 interim findings, from grants notified in November 2026, will tell us nothing about displacement. The instrument does not exist yet, and by the time it does the models will have changed twice.

The Potato Problem

We have run this experiment before, at civilisational scale, and the results were not encouraging.

For two decades researchers have argued about social media and adolescent mental health. In 2019 Amy Orben and Andrew Przybylski published a specification curve analysis in Nature Human Behaviour that ran every defensible analytical choice across three large datasets and found the association between digital technology use and adolescent wellbeing to be about the same size as the association with eating potatoes. Jonathan Haidt, in The Anxious Generation, argues the opposite with equal force: that a phone-based childhood is the principal driver of a genuine crisis. Candice Odgers, reviewing the book in Nature, held that the core causal claim was not supported and that the focus risked distracting from the actual drivers. Twenge, Haidt and colleagues have run their own specification curves finding larger effects, particularly among girls, criticising the earlier work for pooling unlike technologies and ignoring moderators. The dispute continues, with papers published in 2026 still contesting which studies should count.

Two decades, thousands of papers, enormous datasets, and the field cannot agree on the sign of the effect, let alone the size. That is not a scandal. It is what happens when you study a fast-moving, heterogeneous technology with instruments designed for slower things, after it has reached universal adoption and destroyed your control group.

There is a detail here that is almost too neat. When OpenAI announced its Expert Council on Well-Being and AI in October 2025, eight members were named, among them David Bickham of the Digital Wellness Lab, Munmun De Choudhury of Georgia Tech, Tracy Dennis-Tiwary of Hunter College, Sara Johansen of Stanford, David Mohr of Northwestern, Mathilde Cerioli of the children's AI nonprofit everyone.AI, and Andrew Przybylski of Oxford. The co-author of the potato paper now advises the company whose product is the next thing to be measured. That is not a gotcha. Przybylski is exactly the sceptical methodologist you would want in the room, and his presence is a point in OpenAI's favour. But it shows how this works. The most credible researchers in a field are the ones industry most wants to recruit, and recruitment does not require anyone to change their views. It requires only proximity. The reputational transfer is one-directional and happens whether or not anybody intends it.

The Oldest Playbook in the Filing Cabinet

A grant call from a product company triggers reflexive suspicion for a reason. It is not paranoia. It is induction.

On 14 December 1953, the chief executives of the major American tobacco companies met the public relations firm Hill & Knowlton at the Plaza Hotel in New York to decide how to answer the emerging science on smoking and lung cancer. The result was the Tobacco Industry Research Committee, later renamed the Council for Tobacco Research, announced in January 1954 through “A Frank Statement to Cigarette Smokers”, an advertisement that ran in 448 newspapers across 258 cities. It promised research. It funded research, a great deal of it, some genuinely good. In United States v. Philip Morris, Judge Gladys Kessler found the Committee to be a sophisticated public relations vehicle built on the premise of conducting independent scientific research, whose function was to deny the harms of smoking and reassure the public. The industry's own strategy documents put it more crisply than any critic could: doubt is the product.

The sugar industry ran a tighter version. In 2016, Cristin Kearns, Laura Schmidt and Stanton Glantz published an analysis in JAMA Internal Medicine of more than 340 internal documents, over 1,500 pages, showing that the Sugar Research Foundation funded a 1967 literature review in the New England Journal of Medicine that steered dietary blame for coronary heart disease towards fat and cholesterol and away from sucrose. The Foundation set the review's objective, supplied articles for inclusion, and received drafts. The funding was not disclosed. The consensus that followed shaped public health policy for decades.

Pharmaceuticals produced the most rigorous demonstration, because pharmaceuticals produced enough studies to measure the effect statistically. The Cochrane methodology review by Andreas Lundh, Lisa Bero and colleagues, updated in 2017 and covering seventy-five studies, found that industry-sponsored drug and device trials were systematically more likely to report favourable efficacy results and favourable conclusions than independently funded work, with less concordance between what the results said and what the conclusions claimed. Crucially, the effect persisted when analysis was restricted to trials at low risk of bias on standard assessment. It is not explained by bad randomisation or weak blinding. Something else does the work: question selection, comparator choice, outcome definition, the decision about which studies see daylight.

That last point is the one that should be pinned to the wall. Sponsorship bias in medicine survived every methodological fix aimed at the individual study, because it never lived inside the individual study. It lived in the portfolio.

What Actually Fixed Medicine

Medicine did eventually do something about this, and what worked is instructive precisely because it was not voluntary and was not about money.

In September 2004, the International Committee of Medical Journal Editors announced that member journals would refuse to publish any clinical trial not registered in a public registry before the first patient was enrolled. Not registered at submission. Registered before enrolment, with the primary outcome specified in advance. The United States made registration and results reporting a statutory duty through the Food and Drug Administration Amendments Act of 2007. ClinicalTrials.gov became infrastructure. The CONSORT statement standardised what a trial report must contain, so omissions became visible. The ICMJE later required data-sharing statements and defined authorship in a way that made ghostwriting harder to hide.

The effect was not to make industry trials honest. It was to make the portfolio visible. Once you must declare what you are measuring before you measure it, and the existence of your study is public record whatever it finds, the cheapest forms of distortion become expensive. You cannot quietly reclassify your secondary endpoint as your primary one. You cannot leave the disappointing arm in a drawer, because the world knows the drawer exists.

Every one of those reforms was mandatory, imposed by journals and legislatures rather than adopted by sponsors, and universal across a field rather than negotiated study by study. OpenAI's programme, judged against that standard, has none of them. No pre-registration requirement. No published registry of funded projects. No guarantee of publication rights. No commitment to publish the list of applications declined. It is a philanthropy programme, and philanthropy programmes are not built to survive their own incentives.

To be fair, a sponsor setting out to manufacture doubt would not write a rubric rewarding findings unfavourable to AI providers, cap overheads at ten per cent, or open the call to anthropologists and participatory researchers. The design reads like people who want answers. The problem is that wanting answers is not a governance mechanism. The Council for Tobacco Research funded scientists who wanted answers too.

The Thing the Money Cannot Buy

And now the part that makes all of the above secondary, because almost nobody is arguing about it.

Suppose OpenAI fixed everything above tomorrow. Suppose the five million went to an arms-length intermediary with an independent panel, mandatory pre-registration, guaranteed publication rights and no sponsor sight of results before they appeared. The research would still, in the most important respects, be impossible. The data that would answer the question sits on OpenAI's servers, and only OpenAI decides who sees it.

Everything in the current evidence base is a proxy. Self-reported usage. Recruited volunteers. Recalled feelings. What you would actually need, to know whether chatbot use displaces human relationship-building in adolescence, is longitudinal interaction data at the individual level for a consented cohort of teenagers, linked to independently collected developmental measures, over years. Session lengths. Time of day. Escalation patterns. Which model version, with which system prompt, on which date. Whether the safety classifier fired and what happened in the twenty turns afterwards. No external researcher has ever had that, for any conversational product, anywhere.

We know how this goes, because we watched it. Social Science One launched in 2018 as a serious attempt to give academics privileged access to Facebook data through an independent intermediary, and became a case study in how such arrangements fail: years of delay, datasets that arrived late and flawed, and a governance structure in which the platform still controlled the tap. CrowdTangle, the one tool that let outsiders see what was spreading on Meta's platforms, was shut down on 14 August 2024. Its replacement, the Meta Content Library, is narrower, slower to access and closed to much of the journalism that relied on its predecessor. Surveys of affected researchers found the overwhelming majority saying projects would have to be redesigned or abandoned. The lesson is not that Meta is uniquely hostile. It is that voluntary access is revocable access, and revocable access shapes what questions get asked long before anyone revokes it.

Europe tried to legislate the problem away. Article 40 of the Digital Services Act gives vetted researchers a right to request internal data from very large platforms to study systemic risks. The delegated act specifying the procedure came into force on 29 October 2025 and a portal opened, with first decisions expected around February 2026. It is the most ambitious researcher-access mechanism anyone has built. It is also slow and contested: by March 2026 researchers were publishing post-mortems on rejected Article 40 requests. And its scope was written for social networks and search, not for a conversational assistant a fourteen-year-old talks to at two in the morning.

This is the structural insight the funding debate keeps missing. Independence of funding without independence of data access is theatre. Give a researcher a million dollars, a clean contract and a free hand, and if the only window onto the phenomenon is a survey instrument, you have funded another cross-sectional study for the pile. The company will always know more about what its product does to teenagers than the people it is paying to find out.

What the Gambling Levy Got Right

There is a working model for this, and Britain has just finished building it.

For years, research into gambling harms in the United Kingdom was funded by voluntary industry donations routed through the charity GambleAware. The arrangement was permanently contested, not because the research was obviously bad but because the funding chain gave operators discretion over how much to give and, by implication, influence over what got studied. The government replaced it with a statutory levy on operators, set in regulation rather than negotiation, and directed into public bodies. GambleAware wound up in 2026. The research portion now flows to UK Research and Innovation, which in May 2026 launched a gambling harms research centre led by the University of Glasgow with Sheffield, Swansea and King's College London, and the levy raises in the region of a hundred million pounds a year. The prevention strand carries the clause that matters more than the money: from April 2026, the Office for Health Improvement and Disparities required applicants for levy-funded prevention money to declare conflicts of interest and stop taking direct funding from the gambling industry. The research strand has not drawn the line as cleanly. UKRI's call for the new centre said industry representatives were “eligible and encouraged” to apply for co-leadership roles, a choice that drew criticism in The BMJ.

That is what structural independence looks like when someone bothers to build it. The industry pays, because the industry generated the harm and the cost should sit with it. It does not choose the recipients, set the questions, approve the outputs or control the timetable. The link between donor and researcher is broken by statute rather than good intentions. Young and imperfect, with the research strand still letting industry in at the edges, but a category of solution no amount of careful grant-writing by a company about itself can reach.

Apply the template. A levy on providers of general-purpose conversational AI, proportionate to reach among minors. An independent commissioning body, chaired externally, setting the agenda through open peer review. Mandatory pre-registration. Guaranteed publication rights with no sponsor review. A public register of funded and declined applications. And, as the non-negotiable core, statutory access to platform interaction data for vetted researchers under privacy-preserving conditions, written for conversational systems rather than retrofitted from rules about newsfeeds.

Evidence Before Market, Not After

So which is it? A maturing industry or a retrospective alibi?

Take the charitable reading at its strongest, because it is stronger than critics usually allow. OpenAI did not have to do this. Five million dollars is a rounding error against its capital expenditure but a great deal of money in developmental psychology, where a million-dollar grant is career-defining and the field has been starved. The call is genuinely open to disciplines that will produce inconvenient findings. There is no gag clause. The company has published safety research that made it look bad, conceded that its guardrails degrade over long conversations, and rolled back a model update for sycophancy in public. Somebody inside that building is arguing for the truth, and cheques like this one are how those arguments get won. Refusing industry money on principle has a cost too, measured in the studies that never happen at all.

But the timeline decides it. Adam Raine died in April 2025. His parents filed that August. The FTC opened its inquiry in September 2025, California legislated in October, Character.AI removed teenagers from open-ended chat in November. ChatGPT for Teens shipped in August 2026. The research call opened in September 2026. Every one of those dates precedes the first dollar of this programme, and the product has been in the hands of minors for years. This is not evidence-gathering ahead of a decision. It is evidence-gathering after one that will not be reversed by the findings.

We do not accept that sequence anywhere else that children are involved. A medicine intended for adolescents requires trials before licensing, not a research fund announced alongside the launch. A toy must meet safety standards before it reaches a shelf. A car must pass crash testing set by regulators, not by the manufacturer's philanthropy arm. In every one of those domains we decided, usually after something terrible, that the burden of proof sits with the party that wants to sell the thing, and that the proof comes first. Conversational AI aimed at thirteen-year-olds is being treated as though it were a website.

The precautionary answer is not a ban, and pretending otherwise weakens the case. It is staging. Deployment to minors conditional on evidence that accumulates before each expansion, not after all of them. Independent pre-market evaluation for products targeted at under-eighteens, with a regulator holding the standard. A statutory research levy, so the evidence base does not depend on a company's mood. And data access as a licence condition, because without it every other reform is decoration.

Take the money, then. Researchers should apply, publish everything including the null results, pre-register their protocols even though nobody is making them, and say loudly in every paper exactly who paid. Scepticism about the source is no reason to leave the field empty. But nobody should mistake what this is. A five-million-dollar fund with a four-month reporting cycle, reviewed by an internal panel, publication encouraged rather than required and no route to the interaction data, is not the evidence base that should govern whether a generation grows up confiding in machines. It is a gesture in the right direction from an institution that got the order of operations wrong and cannot put it right by spending.

The teenagers are already in the study. They were enrolled without consent, the intervention began years ago, and there is no control group left. The grant call is not the start of the research. It is an attempt to write up an experiment that has been running, unsupervised, since the product shipped.

References and Sources

  1. OpenAI (2026) “Funding grants for new research into AI and teen development.” Announcement, 8 September 2026; and the accompanying application portal at openai.smapply.org (submissions opened 8 September 2026, deadline 6 October 2026, notification on or before 13 November 2026, indirect costs capped at 10 per cent, review criterion on “independence and credibility”).
  2. OpenAI (2026) “Introducing ChatGPT for Teens: Built for learning, backed by protections.” 18 August 2026; and Axios (2026) “OpenAI debuts ChatGPT for Teens,” 18 August 2026; and OpenAI Help Centre, “ChatGPT for Teens” (full availability in Australia from 8 September 2026).
  3. Raine v. OpenAI, Inc. and Samuel Altman, No. CGC-25-628528, Superior Court of California, County of San Francisco. Complaint filed 26 August 2025; amended October 2025; coordinated as JCCP No. 5431 before Judge Ethan P. Schulman by order of 3 February 2026; initial case management conference 24 July 2026.
  4. Federal Trade Commission (2025) “FTC Launches Inquiry into AI Chatbots Acting as Companions.” Press release, 11 September 2025, on Section 6(b) orders issued to Alphabet, Character Technologies, Instagram, Meta, OpenAI, Snap and xAI. Orders dated 10 September 2025.
  5. California Senate Bill 243 (2025), Companion Chatbots. Signed by Governor Gavin Newsom on 13 October 2025; in force 1 January 2026.
  6. GUARD Act (S. 3062), introduced by Senators Josh Hawley and Richard Blumenthal, United States Senate, October 2025; advanced unanimously by the Senate Judiciary Committee, 30 April 2026.
  7. CNN Business (2025) “After a wave of lawsuits, Character.AI will no longer let teens chat with its chatbots.” 29 October 2025; change completed 25 November 2025.
  8. Pew Research Center (2026) “How Teens Use and View AI” and “What parents say about their teen's AI use.” Survey of United States teenagers and parents fielded 25 September to 9 October 2025, published 24 February 2026.
  9. Robb, M. et al. (2025) “Talk, Trust, and Trade-Offs: How and Why Teens Use AI Companions.” Common Sense Media, July 2025.
  10. Common Sense Media (2026) “The Common Sense Media Census: AI Use by Tweens and Teens, 2026.” Fieldwork conducted with SSRS. Released 8 June 2026.
  11. Internet Matters (2025) “Me, Myself and AI: Understanding and safeguarding children's use of AI chatbots.” Survey of 1,000 children aged 9 to 17 and 2,000 parents in the United Kingdom, July 2025.
  12. Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L. and Agarwal, S. (2025) “How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study.” MIT Media Lab and OpenAI, arXiv:2503.17473.
  13. Kirk, H. R. et al. (2025) “Why human-AI relationships need socioaffective alignment.” Humanities and Social Sciences Communications (Nature), May 2025.
  14. OpenAI (2025) Statement on the withdrawal of the sycophantic GPT-4o update, April 2025; and OpenAI statements to the New York Times on the degradation of safety training in long interactions, August 2025.
  15. Orben, A. and Przybylski, A. K. (2019) “The association between adolescent well-being and digital technology use.” Nature Human Behaviour, 3, pp. 173-182.
  16. Odgers, C. L. (2024) “The great rewiring: is social media really behind an epidemic of teenage mental illness?” Nature, review of Jonathan Haidt's The Anxious Generation.
  17. Twenge, J. M., Haidt, J. et al. (2022) “Specification curve analysis shows that social media use is linked to poor mental health, especially among girls.” Acta Psychologica.
  18. OpenAI (2025) “Expert Council on Well-Being and AI.” Announcement of eight named members, 14 October 2025; and CNBC (2025) “OpenAI forms expert council to bolster safety measures after FTC inquiry.”
  19. Kessler, G. (2006) Final Opinion, United States v. Philip Morris USA Inc., United States District Court for the District of Columbia; and Tobacco Tactics, “Tobacco Industry Research Committee,” Tobacco Control Research Group, University of Bath.
  20. Kearns, C. E., Schmidt, L. A. and Glantz, S. A. (2016) “Sugar Industry and Coronary Heart Disease Research: A Historical Analysis of Internal Industry Documents.” JAMA Internal Medicine, 176(11), pp. 1680-1685.
  21. Lundh, A., Lexchin, J., Mintzes, B., Schroll, J. B. and Bero, L. (2017) “Industry sponsorship and research outcome.” Cochrane Database of Systematic Reviews, MR000033.pub3.
  22. International Committee of Medical Journal Editors (2004) “Clinical Trial Registration: A Statement from the ICMJE,” September 2004; and Food and Drug Administration Amendments Act of 2007, Title VIII (United States).
  23. Tech Policy Press (2024) “Researchers Consider the Impact of Meta's CrowdTangle Shutdown.” CrowdTangle discontinued 14 August 2024; Social Science One project documentation, socialscience.one.
  24. European Commission (2025) “Commission adopts delegated act on data access under the Digital Services Act.” Delegated regulation under Article 40 DSA, adopted July 2025, in force 29 October 2025; and DSA Observatory (2026) “If at first you don't succeed: Reflections on a rejected Art. 40 DSA data access request,” 12 March 2026.
  25. The Gambling Levy Regulations 2025 (United Kingdom); GambleAware closure, 31 March 2026; Office for Health Improvement and Disparities, levy prevention funding allocations and conflict-of-interest conditions, 2026; UK Research and Innovation (2026) “UK's largest independent gambling harms research centre launches,” May 2026; and The BMJ (2026) on industry participation in levy-funded research.

Tim Green

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...