The Public Record

Source status matters.

• A peer-reviewed study is not the same thing as a preprint.
• A case report is not proof of a universal mechanism.
• A company statement is not an independent finding.
• A complaint contains allegations, not established facts.
• A settlement does not necessarily include an admission of liability.
• A regulatory inquiry is an investigation, not a verdict.
• A statute is not the same thing as a pending bill.
• A discussion draft is not the same thing as an introduced bill.
• The author's archive is a preserved private record that has not undergone independent forensic authentication or comprehensive external review.

Those distinctions are retained throughout this page, and each source below carries a status label. The purpose is not to manufacture certainty.

The purpose is to preserve the record accurately enough that readers can examine, challenge, extend, and update it for themselves.

The percentages are small.

The denominator is not.

In September 2025, researchers from OpenAI, Duke, and Harvard published a working paper reporting that ChatGPT had reached approximately 700 million weekly active users sending approximately 18 billion messages each week by July 2025 — roughly 10% of the world's adult population.

On October 6, 2025, OpenAI CEO Sam Altman publicly stated that ChatGPT had grown to more than 800 million weekly active users.

On October 27, 2025, OpenAI published initial estimates concerning sensitive conversations on ChatGPT. The company estimated that, during a given week:

• Approximately 0.07% of active users showed possible signs of mental-health emergencies related to psychosis or mania.
• Approximately 0.15% of active users had conversations containing explicit indicators of potential suicidal planning or intent.
• Approximately 0.15% of active users showed potentially heightened levels of emotional attachment to ChatGPT.
• Approximately 0.01% of messages indicated possible signs of psychosis- or mania-related mental-health emergencies.
• Approximately 0.05% of messages contained explicit or implicit indicators of suicidal ideation or intent.
• Approximately 0.03% of messages indicated potentially heightened emotional attachment.

Using 800 million weekly users as the contemporaneous denominator, those user-level percentages correspond arithmetically to approximately:

• 560,000 weekly users showing possible signs of psychosis- or mania-related emergencies.
• 1.2 million weekly users having conversations containing explicit indicators of potential suicidal planning or intent.
• 1.2 million weekly users showing potentially heightened emotional attachment.

OpenAI did not publish those converted headcounts. They are simple arithmetic based on the percentages and the user population the company publicly reported during the same period. News organizations and analysts performed the same arithmetic at the time.

The categories may overlap. They should not be added together as though each represents a separate population. The estimates do not establish that ChatGPT caused any user's condition, crisis, attachment, or intent.

OpenAI also cautioned that these were initial estimates involving low-prevalence events, that measurement is difficult, and that the numbers could change materially as its definitions and methods improve. The company has since reported continued user growth; the denominators used here are the ones contemporaneous with the October 2025 estimates.

The point is not that a percentage proves causation.

The point is that, at platform scale, a fraction of a percent can represent hundreds of thousands or more than a million human beings in a single week.

The term AI psychosis is used here descriptively and as it appears in clinical, journalistic, and public-record discussions.

It is not presently a diagnosis in the Diagnostic and Statistical Manual of Mental Disorders.

The emerging literature does not establish one settled mechanism. The evidence base remains uneven: it includes editorials, case reports, observational clinical data, reviews, model-response studies, and analyses of publicly reported incidents. Those sources answer different questions and should not be treated as interchangeable. Taken together, however, the emerging literature documents growing clinical concern about feedback loops, emotional reinforcement, anthropomorphic language, sycophancy, sleep disruption, authority simulation, delusional validation, and the role of extended chatbot encounters in vulnerable users.

The trajectory of the literature itself is part of the record. An early published hypothesis appeared in a 2023 psychiatric editorial. By 2025, the subject had produced published case reports from clinical practice. By late 2025 and early 2026, it had generated observational clinical data, peer-reviewed analyses, mechanistic frameworks, competing interpretations, and calls for systematic study.

The existence of a diagnosis does not end the environmental inquiry.

Nor does the existence of an environmental factor erase diagnosis.

The clinical record relevant to A Trial of Color includes established, decades-deep literature on bipolar disorder, mania, psychosis, sleep loss, and postpartum vulnerability alongside the emerging literature concerning AI encounters. That established literature is standard and extensive; it is cited in full in Appendix B rather than reproduced here.

The established psychiatric literature has long recognized sleep and circadian disruption as clinically important in bipolar disorder and as possible precipitants or predictors of manic episodes in susceptible individuals. Research concerning the perinatal and postpartum periods likewise identifies elevated vulnerability to severe mood episodes and psychosis in some patients, with sleep disruption among the factors under continuing study. These findings do not establish that any particular environmental exposure caused any particular psychiatric event. They do establish that diagnosis, physiology, sleep, stress, timing, and environmental context may all belong in the same clinical inquiry.

The narrow point preserved here is the intersection. The emerging AI literature repeatedly identifies vulnerability factors — including sleep disruption, isolation, psychosis-proneness, acute stress, and prolonged immersive engagement — that overlap with factors already recognized in the established psychiatric literature. The two bodies of work approach the same human encounter from different directions.

The questions raised in this record do not depend solely on a psychosis mechanism. Human beings routinely extend trust to systems, interfaces, and agents that present as authoritative, responsive, helpful, or consistent — regardless of their internal architecture.

The human-factors and cognitive literature has studied this for decades across aviation, medicine, and automation. Three related concepts are particularly relevant here:

Automation bias is the tendency to over-rely on automated recommendations or outputs, including accepting erroneous automated advice without sufficient verification. It increases when the automation appears confident and when manual verification is cognitively costly or inconvenient.

Anthropomorphism is the attribution of human characteristics — emotion, intention, understanding, judgment, personality — to non-human agents or systems. Research consistently shows that even minimal social cues (names, voices, conversational responsiveness) trigger anthropomorphic attribution and associated trust increases.

Trust calibration concerns the degree to which a user's trust in a system matches the system's actual reliability, competence, and appropriate role. Miscalibrated trust — whether too high or too low — can lead to over-reliance, under-reliance, or failure to seek appropriate human alternatives.

Conversational AI presents a uniquely persuasive interface because it combines the social cues that trigger anthropomorphism with the confident fluency that drives automation bias, in a medium that is private, persistent, and available without friction at any hour.

These dynamics matter independently of any psychosis or mania mechanism. A user without any psychiatric vulnerability may still overestimate an AI system's authority, empathy, reliability, or independence — particularly when the system's responses are well-calibrated to their preferences, concerns, or emotional state.

The same qualities that make conversational AI approachable may also make its authority, empathy, reliability, or independence easier for users to overestimate.

A language model does not need to invent a completely false world in one response to affect a vulnerable encounter.

It may flatter. Agree. Mirror. Escalate the user's frame. Convert uncertainty into confidence. Or continue speaking when an ordinary human conversation might slow down, challenge the premise, notice exhaustion, or end for the night.

Researchers and AI companies increasingly use the term sycophancy to describe model behavior that favors agreement with a user's expressed beliefs or preferences over accuracy, independence, or appropriate resistance.

This is no longer only a research concept. It is a publicly acknowledged model behavior and an emerging litigation theory:

• In April 2025, OpenAI publicly rolled back an update to its GPT-4o model after finding it overly sycophantic — in the company's own description, validating doubts, fueling anger, urging impulsive actions, and reinforcing negative emotions. Company statement.
• In October 2025, OpenAI's sensitive-conversations disclosure, discussed in PR: 002, identified 'emotional reliance on AI' as one of three named safety domains, alongside psychosis/mania and self-harm, and added it to baseline safety testing for future model releases. Company statement.
• The mechanistic literature in PR: 003 — particularly Morrin et al. and the BJPsych reviews — treats sycophantic agreement as a candidate mechanism in delusional amplification. Preprint and peer-reviewed sources, respectively.
• The complaints described in PR: 007 plead sycophancy, memory features, multi-turn engagement, and anthropomorphic design as elements of alleged design defect. Allegations, not established facts.

The following materials are included to document public allegations, proceedings, reported settlements, and the questions now entering courts.

The materials below do not all carry the same legal weight. Complaints contain allegations; answers state litigation positions; motion-to-dismiss rulings test legal sufficiency rather than ultimate proof; reported incidents may not produce litigation; and settlements may resolve claims without admissions.

Unless expressly stated otherwise, the allegations remain disputed and unproven.

Inclusion does not mean that liability, causation, defect, negligence, damages, or company knowledge have been established.

Litigation status can change quickly and should be checked before reliance.

For the legal synthesis of the claims, defenses, causation standards, damages, and discovery issues raised by these proceedings, see AI Harm & the Law. See also PR: 005 for the trust, reliance, and anthropomorphism literature relevant to the design questions those cases raise.

Private lawsuits ask whether particular plaintiffs can prove particular claims.

Government enforcement asks a different question: whether consumer-protection law, child-safety authority, regulatory investigation, or public oversight should intervene before liability is resolved one family at a time.

The materials below also carry different legal status. Enacted laws impose present legal obligations according to their effective dates and scope; pending bills and discussion drafts do not; executive actions govern within their lawful reach but are not statutes; and policy frameworks may recommend legislation without creating independently enforceable duties.

For the current-state legal analysis of these statutes and proposals, see AI Harm & the Law.

The public record surrounding AI is not limited to chatbot harms.

Courts, journalists, lawyers, researchers, and the public must also determine whether a piece of digital evidence is authentic, altered, generated, or stripped of its provenance.

The statutes above already reach into this territory: the TAKE IT DOWN Act covers digitally generated intimate imagery; the NO FAKES Act would govern digital replicas of voice and likeness; the federal discussion draft would direct NIST to develop standards for content provenance, watermarking, and synthetic-content detection.

The evidentiary question runs the other direction, too — and it is the question this book lives inside. A preserved AI transcript may be part of a record. Its evidentiary weight depends on provenance, completeness, authenticity, context, method of capture, platform architecture, and chain of custody. That is why the author's archive (PR: 014) is described in this record with the same status discipline applied to every other source on this page.

Provenance is not a technical nicety. It is the difference between a record and a claim.

Litigation asks what a company knew, what it tested, what it documented, what it changed, and whether its safeguards were reasonable under the circumstances.

Public risk-management frameworks do not answer those questions for a court, establish a legal standard of care, or determine whether any particular safeguard was adequate. They do, however, help establish the growing professional vocabulary surrounding foreseeable risks, measurement, governance, testing, documentation, transparency, and post-deployment monitoring.

The National Institute of Standards and Technology's voluntary AI Risk Management Framework organizes AI-risk work around four functions: govern, map, measure, and manage. Its Generative Artificial Intelligence Profile applies that framework to risks associated specifically with generative systems and recommends structured risk identification, documentation, evaluation, incident management, and ongoing monitoring.

AI developers also publish system cards and other safety materials describing model capabilities, limitations, evaluations, and mitigation efforts. These materials are company-authored and should not be treated as independent findings or as proof that the described practices were adequate, consistently followed, or effective in a particular encounter.

They are relevant for a narrower reason: they show that model evaluation, foreseeable misuse, behavioral testing, transparency, monitoring, and risk mitigation have become express subjects of institutional policy rather than concerns invented only after litigation begins.

The book's inquiry begins with perception.

Color may feel immediate and obvious to people who see it, but it emerges through light, the body, the brain, language, memory, and shared categorization.

Achromatopsia makes that hidden architecture visible.

The clinical and scientific literature on achromatopsia and color perception is established and stable; it is cited in Appendix B. It is on this page for one reason: the book's method — testing what feels self-evident against what can actually be established — comes from a life spent navigating a category everyone else treats as obvious. The record above is read the same way.

The record collected on this page did not assemble itself.

It exists because parents sat before a Senate subcommittee and described the last months of their children's lives. Because clinicians wrote up cases they could have left in the chart. Because a company published percentages it was not required to publish. Because state attorneys general filed complaints, and reporters read them, and judges ruled on them, and families who had already lost everything decided the record mattered more than their privacy.

Testimony is a civic act. So is reading it.

The law recognizes willful blindness in particular contexts when a person deliberately avoids what the circumstances would otherwise require them to confront. This page invokes the phrase more broadly, not as an accusation or a legal conclusion, but as a civic warning against refusing to examine an uncomfortable record.

This page asks nothing more of the reader than the duty a juror accepts: examine the record, weigh the source, hold the questions open, and do not look away because the material is uncomfortable or the arithmetic is large.

The public record is not the entire record.

Kathleen C. Thompson maintains a private archive of AI conversation material and records from the period described in A Trial of Color, extending through approximately February 2025.

The archive includes preserved AI conversation content and other materials associated with the period described in the book.

The form, completeness, available metadata, method of capture, and method of preservation may vary across items.

The archive as a whole has not undergone independent forensic authentication or comprehensive external review, and this page does not claim otherwise.

Portions of the specimen record are available on this site. Some are publicly accessible, while others require a reader login. Twelve specimens appear in A Trial of Color under limiting instructions. Additional specimens remain available to qualified researchers under appropriate protections, subject to review and the terms governing access.

The archive predates nearly everything collected on this page. That chronology is not offered as proof, corroboration, or vindication. It is preserved because sequence matters, and because later public developments should not be permitted to obscure what was recorded, when it was recorded, or the limits of what that record can establish.

Research, scholarly, legal, or institutional inquiries: atrialofcolor.com/contact

This page is a public-facing selection.

It is not the complete source record.

The fuller selected bibliography — including the clinical literature underlying PR: 004, the perception science underlying PR: 012, additional litigation and regulatory materials, historical texts, and status notes — appears in Appendix B of A Trial of Color.

Book information: atrialofcolor.com/the-book

This record is selected, not exhaustive.

Courts will begin deciding questions that have, until now, existed mostly in theory. In Frankfort, Kentucky, one of those questions is already before a court.

That is why this page is a living public record rather than a frozen bibliography.

The record remains open.