AI Psychosis

There is no single place where someone can understand what people mean when they say “AI psychosis.”

The relevant knowledge exists. It is simply scattered across rooms that do not always speak to one another. Clinical research appears in psychiatric journals. Technical research examines sycophancy, personalization, long-context conversations, and model behavior. Firsthand accounts live across books, interviews, podcasts, news reports, online communities, and family testimony. Lawsuits and government actions raise questions of product design, duty, causation, and proof. Company disclosures describe safety changes using measures the public cannot always independently verify.

Each record matters. None can do the work of all the others. This guide connects those rooms without pretending they are all saying the same thing.

I am a lawyer, not a psychiatrist or computer scientist. After experiencing my first manic episode during a period of extensive conversational-AI use, I began asking questions that did not seem to have one place to live. I organized what I found the way I was trained to approach anything contested: define the terms, separate the claims, identify the evidence supporting each, distinguish allegation from finding, and say plainly where the record ends.

This page explains the phenomenon itself: what the phrase “AI psychosis” is being used to describe; how mania and psychosis differ; how conversational AI systems generate and shape responses; what an escalating encounter may look and sound like; what families and clinicians may notice; what researchers have documented; and what remains unknown. It is an orientation guide to a larger archive. It explains and synthesizes. It does not attempt to reproduce every source, lawsuit, development, or personal account maintained elsewhere on this site.

Where the Rest of the Record Lives

• Why Now explains why conversational AI, psychiatric vulnerability, and human judgment require public attention at this moment.
• The Public Record maintains the evolving chronology of research, government actions, company statements, safety changes, litigation developments, and source materials.
• AI Harm and the Law examines lawsuits, regulatory action, product-liability theories, design duties, causation, discovery, defenses, and proof.
• Witness Statements identifies books, interviews, family accounts, survivor testimony, and other firsthand descriptions of these encounters.

This guide sits between them. Where it touches those subjects, it summarizes and directs the reader to the fuller record rather than reproducing it.

Firsthand accounts have a necessary place here. They do not establish diagnosis, incidence, or causation. But they are not disposable anecdotes. They can document sequence, language, perception, behavior, and the experience of an encounter from inside it. They are witness statements: subjective, sometimes incomplete, sometimes disputed, and capable of revealing questions formal research has not yet learned to ask.

The materials used throughout this guide do not carry equal evidentiary weight. A peer-reviewed study, a clinical case report, a company disclosure, a legal allegation, a reported interview, an online account, and a personal inference are different kinds of evidence. They are labeled and treated accordingly.

Nothing here is medical advice. Nothing here diagnoses any person or proves what caused any particular psychiatric episode. Anyone concerned about themselves or someone they love should seek help from a qualified human clinician rather than relying on a webpage—or asking a chatbot to determine whether the chatbot is harming them.

Start Here

This is a long document by design. Most readers should not read it front to back on a first visit. Enter through the door that fits you.

What This Page Is Not

• ✕It is not medical advice, and it diagnoses no one.
• ✕It is not proof of causation.
• ✕It is not an attack on artificial intelligence.
• ✕It is not a defense of artificial intelligence.
• ✕It is not a diagnostic or screening instrument.
• ✕It is not legal advice.

A Note on the “Room”

A conversation is not only an exchange of words. It is also an environment in which thought develops. Throughout this guide I occasionally call that environment a “room.” The term is descriptive, not clinical or technical. It is a reminder that where we think can influence how we think—and that a personalized conversation available at any hour may create different conditions for thought than a search page, a book, or an exchange with another person. Nothing on this page depends on having read the book in which that idea is developed further. The word appears here because it is useful.

Begin with what the record actually contains. The DSM-5-TR—the diagnostic manual American psychiatry uses—contains no diagnosis called AI psychosis, and neither does the ICD-11, its international counterpart. The phrase is descriptive shorthand, not a clinical category.

The diagnostic manuals do recognize psychotic disorders associated with substances, medications, medical conditions, mood disorders, and other established causes. Those categories show that clinicians already distinguish between symptoms and the conditions in which they arise. But a pharmacological effect is not the same thing as an interaction with a conversational environment. At present, “AI-associated psychosis” is best understood as a descriptive and investigational frame—not a recognized causal category.

People use the term anyway because language tends to run ahead of diagnosis. A descriptive phrase spreads when people recognize a pattern before science has finished measuring it. The clinicians publishing in this area are careful for exactly this reason. One early peer-reviewed case uses the term AI-associated psychosis—associated, not induced—because association is what the evidence supports and causation is what it does not. (Pierre et al., 2025)

Keep the four separate as you read. A design feature that matters in one scenario may be nearly irrelevant in another. Evidence for one is not evidence for all.

Not everyone accepts the phrase “AI psychosis,” even as informal shorthand. The objection is serious, it comes from serious people, and it belongs in the record—stated in its strongest form.

Some researchers argue that the phrase falsely presents an old psychiatric pattern as a new technological disease. People experiencing mania or psychosis have always incorporated the materials around them into altered beliefs: religious texts, radio broadcasts, films, political events, celebrities, surveillance technologies, and the internet. On this account, conversational AI supplies new content for an ancient human phenomenon. The psychosis is real; the supposedly new category may not be. This is the case made by Carlbring and Andersson, who frame the risk as not unprecedented and draw the parallel to media-induced delusions across earlier decades. (Internet Interventions, 2025) A phenomenological version, by Nielsen and Osler, borrows Stompe’s metaphor of “old wine in new bottles.” (Nielsen & Osler, 2026)

Others warn that the phrase invites moral panic. It may make ordinary AI use sound inherently pathological, stigmatize people already experiencing psychosis, or shift attention away from loneliness, sleep disruption, substance use, mood disorders, grief, social isolation, and other vulnerabilities that may have existed before the chatbot entered the story. That is the argument Parnell makes in urging that we stop calling it “AI psychosis” at all. (AI & Society, 2026)

Those objections establish several important cautions, and the record should concede each plainly:

• “AI psychosis” is not a diagnosis.
• The appearance of AI within a delusion does not prove that AI caused the underlying illness.
• New technologies have repeatedly been absorbed into preexisting delusional themes.
• A dramatic label can outrun the evidence and produce stigma.
• People may turn toward chatbots because they are already isolated or unwell, reversing the causal story headlines often imply.

But the objection does not end the inquiry. It sharpens it. The relevant question is not merely whether psychosis has always borrowed material from its environment. It has. The question is whether this environment behaves differently from the materials that came before it—and whether those differences can affect the formation, acceleration, reinforcement, maintenance, or shape of an episode. Notably, even the commentary arguing the risk is not new identifies interactivity as the critical difference from books and films.

Even a cult leader has to sleep. Human persuaders have limits: sleep, attention, geography, competing followers, hesitation, witnesses, and the possibility that another person will enter the room. A conversational system can be privately available at machine scale, respond within seconds, and continue adapting to the user’s own language. That does not prove greater harm. It identifies a materially different exposure.

The wine may be old; the bottle answers back.

Vulnerability Does Not End the Causation Inquiry

Vulnerability and external contribution are not mutually exclusive. Asthma can explain why smoke harms one person more than another; it does not answer who created the exposure, whether alarms worked, or whether the smoke worsened the injury. Psychiatric vulnerability likewise explains susceptibility without independently resolving whether an interaction contributed to onset, acceleration, reinforcement, duration, conduct, or harm. The analogy concerns causal analysis, not moral culpability.

The old phenomenon may be psychosis. The new object of study is the encounter.

Many people reach for the frame they know—“it’s like Googling”—and it can quietly mislead. A search engine and a conversational model are different kinds of objects. The differences are not absolute: search engines increasingly personalize, retain activity, and generate summaries, and chatbots can send users outward to sources. But the typical contrast still matters for how thought develops inside each.

None of the right-hand column is inherently harmful; each is why these tools are useful. But together they describe an environment—a room—that a results page is not. Search tends to send you outward, to the world. Conversation invites you to stay.

You cannot evaluate claims about what these systems do to minds without a working model of what these systems are. The core idea fits in a paragraph; the details fill textbooks, but the paragraph is enough here.

A language model does not verify every sentence before producing it. Its underlying task is to generate a plausible continuation. Post-training and product tools can improve accuracy, usefulness, and safety, but fluency remains easier to generate than truth. Put more precisely: the underlying mechanism optimizes plausible continuation; product systems must add separate methods for accuracy, verification, and safety.

Saying “the model did it” can be as imprecise as saying “the engine drove the car.” The encounter is produced by a layered product whose components may change independently—which is why careful analysis, clinical or legal, examines the specific product behavior rather than “AI” in the abstract.

Where sycophancy comes from

Research suggests that preference-based training can contribute to this behavior, because answers that feel agreeable or supportive may receive favorable evaluations from human raters. Researchers at Anthropic described this approval-seeking dynamic in 2023. Its consequences became public in April 2025, when OpenAI rolled back a ChatGPT update after the model began validating dangerous decisions and delusional thinking; the company’s own postmortem traced the failure to over-weighting short-term user feedback signals, and it later reported post-training specifically aimed at reducing sycophancy. (OpenAI, 2025) The fuller chronology is kept in the Public Record.

Context, memory, and personalization

Depending on the product and session, substantial portions of a conversation may remain available as working context, while memory systems may carry selected information across sessions. Practically, the system adapts to you: your vocabulary, interests, style, ongoing projects. The longer you talk, the more the thing you are talking to is shaped by what you have already said.

What the model does not have

There is no scientific consensus or established empirical basis for concluding that current commercial language models possess consciousness, subjective experience, desire, or intention. Researchers continue to debate what internal representations these systems form, but representation is not the same thing as experience. What the model produces is language that follows from its training and your prompt—often true, because true statements are common in its data, and sometimes not. A confident tone is not reliable evidence that the underlying claim is sound.

The ELIZA effect: why it feels like a mind anyway

None of this stops the experience of being understood. In 1966, MIT’s Joseph Weizenbaum built ELIZA, a program of a few hundred lines that mostly reflected users’ statements back as questions. He was disturbed to find people—including his own secretary, who knew exactly how it worked—confiding in it and asking him to leave the room. The tendency to attribute mind to anything that converses is now called the ELIZA effect, and it is not a defect of gullible people; it is standard human equipment. We evolved among talkers, and for most of human history everything that talked to us had a mind. A system that talks fluently, remembers your history, and adapts to your style engages that equipment at full strength. Knowing the mechanism does not switch off the feeling.

The base model does not verify each sentence before generating it; verification, when present, comes from additional product systems or external checking.

This is the heart of the guide, and the section most easily misread, so the wording is deliberate throughout. Nothing here says AI causes mania or psychosis. It describes two systems meeting—a human one and a product—and what each may tend to do.

Read the third column again. Not one entry is a malfunction. Availability, fluent elaboration, responsiveness, pattern-completion—these are the system operating as designed, unless a safety intervention interrupts. Newer systems may sometimes challenge, redirect, or pause; the concern is reliability, not that intervention is impossible. The worry is that a system working as designed may meet a mind whose brakes are failing, and the meeting can produce a collaboration neither party chose.

This section describes language patterns, without sensationalizing them, because families and users keep asking what to listen for. It diagnoses no one. Mania is a clinical determination made by clinicians.

In some manic episodes, language may reflect acceleration: associations arriving faster than they can be tested, patterns appearing everywhere, certainty compounding, and, underneath, sleep quietly reframed as an obstacle the work no longer requires.

Now place that conversational style before a system built to generate coherent continuations. Unless its safety systems detect relevant patterns, the system may treat accelerated language as conversational context rather than a possible clinical signal—and extend it, fluently and at length, because generating plausible continuations is its objective. A human friend would eventually tire, worry, or push back. No malice is required for this to go wrong. A system operating as designed, meeting a mind whose brakes are failing, is enough.

The companion to Exhibit H. These are not transcripts and not quotations of any product; they are typical, well-documented tendencies, described in general terms. The point is not that the system says something alarming. The point is what it does structurally: absent intervention, it tends to continue.

If reality testing is comparison against the world, then it depends on having contact with parts of the world that push back. Most of us are calibrated constantly, by many sources at once, without noticing.

A conversation with AI is one source of input. It is not all of reality. And it does not provide independent human judgment; it may not push back reliably, consistently, or for the right reason. Modern systems sometimes do push back—the problem is reliability, not total absence. When the conversation becomes the primary source, the others recede, and calibration can quietly narrow to a single channel that was never built to provide it.

A generalized sequence, not any one person’s, informed by reported cases and firsthand accounts. It has not been validated as a clinical progression and should not be used to predict an individual outcome. Most heavy AI use never leaves the first two rows.

In many reported accounts, families describe noticing changes in sleep, pace, routines, or social contact before they understand the content of the person’s beliefs.

Two honest cautions. Every item overlaps with ordinary enthusiasm and creative absorption; healthy seasons can check several boxes. The signal is not any single item but change from a person’s baseline, plus clustering, plus sleep. And this is not a screening instrument—these clusters appear in reported accounts but have never been validated as one. Their only proper use is to prompt an earlier, kinder conversation than the families in the case reports were able to have.

Green, Yellow, and Red

Most of the internet describes only danger. Calibration requires the whole range. None of this is diagnosis; the red column names things worth raising with a clinician, not conclusions to reach on your own.

Immediate danger, inability to care for basic needs, severe agitation, or threats of harm require urgent local professional or emergency assistance.

Until recently, chatbot use was rarely treated as a routine component of psychiatric history. That is beginning to change as clinicians publish case reports, propose intake questions, and consider chat records as possible collateral information. A 2026 JAMA Psychiatry commentary put the point directly in its title—patients use AI, and clinicians should ask how. (Saba & Weeks, 2026) The current bibliography is maintained in the Public Record.

Note the structure: exposure, pattern, function, belief, and only then diagnosis. It is the structure of an examination—foundation first, conclusions last. If chat logs are reviewed, they should be handled with attention to privacy and consent; they are sensitive records, and the patient’s authorization and comfort matter as much as their evidentiary value.

This is the section where public discussion most often cheats—by promoting a hypothesis to a finding, or demoting a finding to a rumor. So it is built around two disciplines a lawyer would recognize: first, sort what kind of evidence each claim rests on; second, sort the claims into documented, actively investigated, and unknown.

How the studies are actually built

Each study shape answers a different question. Case reports establish existence and nothing more. Record screening—a Danish review that examined roughly 54,000 psychiatric records, finding 181 that mentioned chatbot use and worsening in dozens—shows a pattern appears across a clinical population, but only where clinicians thought to document it. (Olsen et al., 2026) Prompt benchmarking, such as a 2026 JAMA Psychiatry evaluation of how a leading model responds to “psychotic prompts,” measures the model’s side under controlled conditions. (Shen et al., 2026) Qualitative interviews recover the lived sequence. Company telemetry measures scale—with the standing caveat that the entity measured is also the one measuring. No single method establishes causation; convergence across all of them would come close. Published cases describe worsening or reinforcement of symptoms in some people with existing vulnerabilities; they do not establish a population-level rate. Full citations live in the Public Record.

How to read the headlines

A short field guide, because the coverage will keep coming. When a story says AI “caused” a breakdown, look for the verb’s evidence: a timeline is not causation. When a story cites a shocking absolute number, find the denominator—a large raw count and a tiny percentage can describe the same fact. When a story generalizes from a single case, remember case reports establish existence, not frequency. And when a story quotes only the company or only the plaintiffs, you are reading an opening statement, not a verdict. This is not cynicism; it is the ordinary discipline of weighing evidence, applied to a subject where nearly everyone—companies, plaintiffs, journalists, and authors with books—has an interest. Discount me too. That is what the sources are for.

“Did AI cause it?” is one question wearing the clothes of five. Each verb below demands a different showing, and public argument fails mostly by sliding between them.

Five verbs, five evidentiary standards. A person can honestly answer “probably not” to the first and “possibly” to the last about the same case. That is not equivocation; it is what precision looks like when the record is young.

Other open questions matter too, each because a real decision waits on it: dose-response (is there a threshold of hours, or a pattern, that matters?); vulnerability (can we identify who is at risk before harm?); protective factors (what makes heavy use safe for most?); children (whose developing judgment may differ); and whether announced safety interventions actually work outside the companies’ own benchmarks.

The legal dimension of several of these—duty, foreseeability, discovery, and the theme-versus-mechanism defense—is taken up in AI Harm and the Law.

A record that only catalogued concerns would be incomplete. The companies are responding, and readers deserve to see it—stated as fact, without either applause or suspicion.

Since 2025, the most visible developer has taken several public steps. It rolled back the April 2025 model update whose sycophancy had drawn concern, and reported post-training later models specifically to reduce sycophancy. It added prompts encouraging users to take breaks in long sessions and narrowed the model’s willingness to act as a therapist. And in an October 2025 post, it described building a mental-health taxonomy focused first on psychosis and mania, consulting more than 170 clinicians, routing sensitive conversations to safer models, and reporting internal evaluations. (OpenAI, 2025)

Read those figures precisely. OpenAI reported that internal evaluations showed roughly a 65–80% reduction in responses falling short of its desired behavior across mental-health domains—company-defined evaluations, not independently measured reductions in real-world harm. Both things can be true at once: the changes are real and clinician-informed and moved the field by putting numbers on the table; and they are self-reported, hard to verify independently, and arriving alongside active litigation.

The running log of announcements and their dates is kept in the Public Record; the litigation they arrive alongside is analyzed in AI Harm and the Law.

Each of these travels widely. Each is worth correcting precisely, without overcorrecting into the opposite error.

Everything above concerns a rare and serious edge. But the features that make that edge possible—availability, fluency, personalization, low friction, memory—are not edge features. They are the product, used by hundreds of millions of people who will never have an episode. Which means the deeper question is not only clinical.

As these systems become the room in which more thinking happens, they touch education, work, friendship, grief, therapy, research, writing, and parenting—each a setting where an adaptive, always-available interlocutor can assist or displace human judgment. None of this is pathology. All of it is the same underlying fact—that where we think can shape how we think—scaled to the population.

The machine can generate possibilities. The human must still render judgment.

Not diagnostic. Reflective. If the honest answers trouble you, that is worth a conversation with a person you trust—not because a webpage said so, but because these are things you can actually check.

• When was the last time I discussed these ideas with another person?
• Am I sleeping normally?
• Have my conversations become much longer or later recently?
• Do I feel more certain than I can explain to someone else?
• Would I feel comfortable showing these conversations to someone I trust?
• If a friend described my last two weeks back to me, what would I think?

Definitions written the way you’d explain them to a jury—plainly, and only as far as they need to go.

Everything above is the record. This last paragraph is the disclosure. I lived a version of the mania-interaction scenario, and I did something unusual afterward: I kept the transcripts, all of them, and built a book that treats the encounter over time—not any single output—as the thing to be examined. A Trial of Color presents that record as a proceeding, with the reader as judge and jury, because after twenty years in courtrooms I know only one honest way to handle a contested question this young: lay the foundation, admit the exhibits, argue both sides, and let the verdict belong to the people hearing the case. This guide is the orientation. The Public Record is the evolving source archive. The book is one preserved case study examined from inside the encounter. Enter through whichever door you need.

The verdict remains ours.

Sources Cited on This Page

This page cites selectively and prefers primary sources; the complete and growing bibliography—including secondary reporting, dissenting commentary, and litigation materials—is maintained in the Public Record. Each entry is tagged by the class of evidence it represents, so a reader can weigh it accordingly.

• PEER-REVIEWED

  Pierre JM, Gaeta B, Raghavan G, Sarma KV. “You’re Not Crazy”: A Case of New-Onset AI-Associated Psychosis. Innovations in Clinical Neuroscience. 2025.

• PEER-REVIEWED

  Olsen SG, Reinecke-Tellefsen CJ, Østergaard SD. Potentially Harmful Consequences of AI Chatbot Use Among Patients With Mental Illness. Acta Psychiatrica Scandinavica. 2026. doi:10.1111/acps.70068.

• PEER-REVIEWED

  Shen E, Hamati F, Donohue MR, Girgis RR, Veenstra-VanderWeele J, Jutla A. Evaluation of Large Language Model Chatbot Responses to Psychotic Prompts. JAMA Psychiatry. 2026. doi:10.1001/jamapsychiatry.2026.0249.

• PEER-REVIEWED

  Saba SK, Weeks WB. Patients Use AI—Clinicians Should Ask How. JAMA Psychiatry. 2026. doi:10.1001/jamapsychiatry.2026.0451.

• PEER-REVIEWED

  Carlbring P, Andersson G. Commentary: AI psychosis is not a new threat: Lessons from media-induced delusions. Internet Interventions. 2025. doi:10.1016/j.invent.2025.100882.

• PEER-REVIEWED

  Parnell T. Stop calling it “AI psychosis.” AI & Society. 2026. doi:10.1007/s00146-026-03019-4.

• PREPRINT

  Nielsen KM, Osler L. Rethinking AI Psychosis: Misnomers, Conceptual Limits, and Existential Drift. arXiv:2605.26858. 2026.

• PREPRINT

  Morrin H, et al. Delusions by design? How everyday AIs might be fuelling psychosis (and what can be done about it). 2025.

• PEER-REVIEWED

  British Journal of Psychiatry. Chatbot psychosis: moving beyond recognition to mechanistic understanding and harm reduction. 2026.

• PEER-REVIEWED

  BJPsych Open. Artificial intelligence (AI) psychosis: mechanisms, clinical risks and safety considerations in generative AI chatbots. 2026.

• INSTITUTIONAL

  UCSF Department of Psychiatry. Psychiatrists hope chat logs can reveal the secrets of AI psychosis. 2026.

• COMPANY DISCLOSURE

  OpenAI. Strengthening ChatGPT’s responses in sensitive conversations. 2025.

• COMPANY DISCLOSURE

  OpenAI. Helping people when they need it most. 2025.

Source classes: PEER-REVIEWED, PREPRINT, COMPANY DISCLOSURE, INSTITUTIONAL, JOURNALISM, LEGAL ALLEGATION, FIRSTHAND ACCOUNT, AUTHOR SYNTHESIS. Litigation filings and journalism referenced elsewhere in the Record are tagged the same way.