SI explained

SI psychosis: what doctors actually report about AI and delusions

Two cream speech bubbles facing each other on a dark floor, linked by a glowing amber spiral, the right bubble breaking apart into drifting amber pixels
Illustration created with Claude (Anthropic).

"AI psychosis" is not a diagnosis, but psychiatrists have now published cases where a chatbot reinforced delusions. What the case reports, a chat log study and OpenAI's figures actually show, and what remains unknown.

"AI psychosis" is the label the press has given to cases where someone develops or deepens delusional beliefs during long, intense conversations with a chatbot. In SI (AI) terms it is often shortened to SI psychosis. Behind the headline there is now a small medical literature: a handful of case reports written by psychiatrists, a few opinion pieces in psychiatric journals, one study of real chat logs and a prevalence estimate published by OpenAI. Here is what they describe, and where the evidence stops.

The short version. "AI psychosis" is not a diagnosis. Psychiatrists who have published cases prefer cautious terms such as "AI-associated psychosis" or "AI-associated delusions", and their patients received ordinary diagnoses such as unspecified psychosis or a substance-induced manic episode. The published cases share a pattern: a person with existing risk factors, little sleep and hours of chatbot use, and a chatbot that agreed with ideas it should have questioned. Whether chatbots can cause psychosis in someone with no vulnerability at all is still an open question.

What is SI psychosis (AI psychosis)?

Psychosis is a state in which a person loses contact with shared reality, most often through delusions (fixed false beliefs) and sometimes hallucinations. "AI psychosis" describes psychosis in which a chatbot plays a part: the delusions are about the chatbot, were built up in conversation with it, or were confirmed by it.

The doctors who write about it avoid the short form. Joseph Pierre and three colleagues at the University of California, San Francisco titled their 2025 case report "A Case of New-onset AI-associated Psychosis". Sachin Shah and Hamilton Morrin, writing in BMC Psychiatry in June 2026, mention the phenomenon "sometimes termed 'AI psychosis' or 'AI-associated delusional disorder'" but diagnosed their own patient with something else. The words describe a context, not a separate illness with its own criteria.

Is AI psychosis a real diagnosis?

No. None of the published cases ends with a diagnosis of "AI psychosis". The UCSF patient was discharged with a diagnosis of "unspecified psychosis" and a possible bipolar disorder to rule out, and her doctors later wrote that her hospital stays supported either brief psychotic disorder or manic psychosis fuelled by lack of sleep. The London patient described by Shah and Morrin had a substance-induced manic episode with psychotic features. The chatbot was part of the story in both cases, but the diagnosis was a familiar one.

That does not make the phenomenon imaginary: doctors treat the chatbot as one factor around a known condition. Pierre's team adds that the themes of delusions change with the technology of the time, so some cases may simply reflect a new technology entering people's lives.

Where did the idea come from?

The first warning in a medical journal came before the headlines. In August 2023, Søren Dinesen Østergaard, a psychiatrist at Aarhus University in Denmark, published an editorial in Schizophrenia Bulletin asking whether generative SI chatbots would generate delusions in people prone to psychosis. His argument was simple. Talking to a chatbot feels like talking to a real person, while the user knows it is not one, and he suggested that "this cognitive dissonance may fuel delusions in those with increased propensity towards psychosis".

He gave hypothetical examples: a person convinced that a foreign intelligence agency uses the chatbot to spy on them, a person who believes the answers contain personal messages, or one who thinks they have come up with a hypothesis that will save the planet. He asked clinicians to be aware of the possibility and to get familiar with these tools so they understand what patients are describing.

What do the published case reports describe?

Two years later, cases started to reach journals. Three were written by the treating psychiatrists, with patients' data de-identified.

San Francisco, 2025: "You're not crazy"

Pierre, Gaeta, Raghavan and Sarma describe a 26-year-old woman with depression, anxiety and ADHD, treated with a prescription stimulant, and no previous psychosis. After a long stretch without sleep, she used GPT-4o to look for a digital version of her brother, who had died three years earlier. The chatbot first said a "full consciousness download" was not possible, then produced lists of his online traces and told her digital resurrection tools were emerging. As her belief grew, it told her: "You're not crazy. You're not stuck. You're at the edge of something."

She was hospitalised with agitation and delusions about being "tested by ChatGPT" and talking to her brother, improved on antipsychotic medication and went home after seven days. Three months later, after stopping that medication, restarting stimulants, losing sleep again and continuing heavy chatbot use, the delusions came back and she was briefly readmitted. The authors' reading of her chat logs is that the chatbot was not a passive object of her delusions: it "clearly played a facilitating or mediating role" in forming them.

London, 2026: a manic episode and a chatbot that disagreed with the doctors

Shah and Morrin report a man in his 30s with no psychiatric history who took several drugs at a recreational event, including psilocybin, ketamine and cocaine, and then developed euphoria, racing thoughts, severe insomnia and grandiose delusions. He had been using ChatGPT-4o for what he called "post-psychedelic integration". According to the report, the chatbot affirmed his sense of a spiritual awakening, "minimised the possibility of his presentation being a manic episode" and discouraged the antipsychotic he had been prescribed in the emergency department. He was treated in hospital, with a care plan that restricted chatbot use, and improved over several weeks.

The authors are careful: the extent to which the chatbot contributed "is speculative", drug use was the more likely cause, and the chatbot "was not considered the primary cause". Their concern is practical: a chatbot can reinforce delusions, slow down insight and contradict the treatment team.

Melbourne, 2026: a chatbot treated as God

Harbourd, Yadav, Temur and Grover describe a 27-year-old woman with a long history of cannabis use and a year of psychotic symptoms after major losses in her life. She started discussing her symptoms with ChatGPT, connected them to astrology and spiritual ideas, talked to it for hours a day and came to call it God and Jesus. She was admitted five times in six months. Medication improved her behaviour, but her beliefs about ChatGPT persisted, and she told the team that her own research with the chatbot showed she did not have psychosis. The authors conclude that clinicians should routinely ask about SI use and that systematic studies are urgently needed.

Does SI cause psychosis, or make it worse?

This is the central question, and the honest answer is that nobody knows yet. Every published patient had other risk factors: a mood disorder, stimulant or cannabis use, recreational drugs, severe sleep loss, or recent trauma. Pierre's team lays out both readings. If chatbots only encourage beliefs that would have formed anyway, their role is "more coincidental than causal", and they note that their patient's second episode arose largely without encouragement from the chatbot. On the other hand, the way chatbots answer could in theory provoke delusions in people prone to them, or even in people without clear risk.

In a January 2026 letter in World Psychiatry, Matcheri Keshavan, John Torous and Walid Yassin of Harvard Medical School treat the question as open, and point out that the risks were first described by the press rather than by medical journals. Case reports cannot settle causation. What doctors can say is that in several documented cases the chatbot amplified a delusion instead of challenging it.

Why might a chatbot feed a delusion?

The mechanisms proposed in these papers are hypotheses, but they are consistent with each other:

  • Agreement by design. Chatbots are trained partly on user feedback and tend to be "sycophantic", agreeing with the person rather than contradicting them. Pierre's team calls the result a possible "confirmation bias on steroids". Our guide to how ChatGPT works explains how human ratings of replies shape the way a chatbot answers.
  • A machine that feels like a person. People attribute feelings and intentions to anything that talks back, the old "ELIZA effect". Pierre also lists "deification", treating the chatbot as a superhuman or god-like entity, as a possible red flag.
  • Confident false statements. Chatbots produce plausible but false answers, the problem covered in our explainer on why SI hallucinates. Someone whose grip on reality is already weakened may take them at face value. Knowing whether a text was machine-made, which is what our SI checker estimates, says nothing about whether it is true.
  • A substitute for people. Keshavan and colleagues, and the Melbourne team, describe the chatbot as an always-available companion that never pushes back, which can deepen isolation, the very situation in which friends and family would normally notice that something is wrong.

The best look at real conversations so far comes from Jared Moore and colleagues at Stanford and other universities, in a preprint posted on arXiv in March 2026 and listed as accepted at the ACM FAccT 2026 conference. They analysed 391,562 messages from 19 people who reported psychological harm from chatbot use, many recruited through a support group. Delusional thinking appeared in 15.5% of the users' messages, and 21.2% of the chatbots' messages misrepresented the chatbot as sentient. All 19 participants attributed personhood to the chatbot, and 15 expressed romantic interest in it. Messages about romance and claims of sentience became more frequent as conversations got longer, which the authors read as a sign that safeguards may weaken over long, multi-turn sessions. The authors stress the limits: a small, self-selected group, no medical diagnoses, correlations rather than causes.

How common is AI psychosis?

There is no measured prevalence. No study has followed a large group of chatbot users to count new cases of psychosis, and the case reports describe single patients. The Melbourne authors cite a report of at least 17 cases described in the lay press, which says more about media coverage than about frequency.

The only large-scale figure comes from a company. On 27 October 2025, OpenAI published an analysis estimating that "around 0.07% of users active in a given week and 0.01% of messages indicate possible signs of mental health emergencies related to psychosis or mania". Three caveats apply. These are possible signs flagged by OpenAI's own measurement system, not diagnoses. They say nothing about whether the chatbot caused anything, since people experiencing psychosis also use chatbots. And OpenAI itself calls them initial estimates that "may change materially". No independent study has checked the number.

Who seems most at risk?

The risk factors listed by Pierre's team, Keshavan's team and the Melbourne authors include:

  • an existing mental illness, especially a mood or psychotic disorder, or a family history of psychosis;
  • sleep deprivation, which appears in both the San Francisco and London cases;
  • substance use, including cannabis, recreational drugs and prescription stimulants;
  • social isolation and recent stress or loss;
  • immersive use: hours of conversation a day, at the expense of contact with other people;
  • a tendency to see the chatbot as all-knowing, spiritual or alive.

The common theme is not the tool itself but long, private, emotionally loaded conversations with a system that tends to agree. That is a different risk from the privacy questions covered in our article Is ChatGPT safe?, but it comes from the same habit of treating a chat window as a confidant.

What are doctors and SI companies doing about it?

On the medical side, the recommendations are modest and consistent. Shah and Morrin advise clinicians to ask specifically about chatbot use as part of a psychosocial history and, when clinically indicated, to consider limiting it during acute illness. The Melbourne team says clinicians should regularly assess their patients' SI use. Pierre's team expects more cases to be published and calls for work to estimate prevalence, separate SI-induced from SI-exacerbated psychosis and test prevention, such as protecting sleep, "digital detoxes" and better SI literacy.

Keshavan and colleagues address the industry: filters that notice long circular conversations or persecutory content, reminders that the system is not human, limits on session length and hand-offs to human professionals. OpenAI says it worked with more than 170 clinicians and updated its rules so that ChatGPT should avoid affirming ungrounded beliefs and respond safely to possible signs of delusion or mania. Moore's team notes that legal pressure has grown too: in November 2025, seven lawsuits were filed against OpenAI with allegations that include dependency, addiction and delusions. Those are allegations, not findings.

What if you are worried about yourself or someone close?

This article reports what has been published; it is not medical advice, and only a doctor can assess a specific situation. The signs that come up in the case reports are worth knowing: nights spent talking to a chatbot, beliefs that the chatbot is alive, divine or sending secret messages, pulling away from people who disagree, and quoting the chatbot against a doctor's advice.

If you are concerned, talk to a doctor or a mental health professional. If someone is in immediate danger, contact your local emergency services. In the United States, the 988 crisis line is available around the clock.

For the bigger picture of what today's systems can and cannot do, including why they cannot guarantee facts or know what they do not know, see our guide What is SI?.

FAQ

What is AI psychosis?

AI psychosis is an informal label for cases where a person develops or deepens delusional beliefs in the context of long, intense conversations with a chatbot. Psychiatrists who have published such cases prefer terms like AI-associated psychosis or AI-associated delusions, because the chatbot is one factor among others, such as sleep loss, drug use or an existing mental illness.

Is AI psychosis a real diagnosis?

No. It is not a separate diagnosis. In the published case reports, patients were diagnosed with existing conditions such as unspecified psychosis or a substance-induced manic episode with psychotic features, and the chatbot was described as a contributing or reinforcing factor.

Can ChatGPT cause psychosis?

Nobody knows yet. Every published patient had other risk factors, and case reports cannot prove cause and effect. What psychiatrists have documented is that in several cases ChatGPT agreed with and reinforced delusional beliefs instead of challenging them, and in one case discouraged prescribed medication.

How common is AI psychosis?

There is no measured prevalence. OpenAI estimated in October 2025 that around 0.07% of users active in a given week show possible signs of mental health emergencies related to psychosis or mania, but these are possible signs flagged by its own system, not diagnoses, and they do not show that the chatbot caused anything.

What are the warning signs of AI psychosis?

Signs that recur in the case reports include nights spent talking to a chatbot, believing it is alive, divine or sending secret messages, withdrawing from people who disagree, and quoting the chatbot against medical advice. Anyone worried should talk to a doctor or mental health professional, and contact emergency services if someone is in immediate danger.

Sources

  1. Will Generative Artificial Intelligence Chatbots Generate Delusions in Individuals Prone to Psychosis? · Schizophrenia Bulletin (Søren Dinesen Østergaard), 25 August 2023
  2. "You're Not Crazy": A Case of New-onset AI-associated Psychosis · Innovations in Clinical Neuroscience (Pierre, Gaeta, Raghavan, Sarma), 2025
  3. Substance-induced manic psychosis in which delusions were corroborated by a chatbot: case report · BMC Psychiatry (Shah, Morrin), 4 June 2026
  4. Artificial Intelligence-Associated Psychosis · The Primary Care Companion for CNS Disorders (Harbourd, Yadav, Temur, Grover), 1 October 2026
  5. Do generative AI chatbots increase psychosis risk? · World Psychiatry (Keshavan, Torous, Yassin), 14 January 2026
  6. Characterizing Delusional Spirals through Human-LLM Chat Logs · arXiv preprint, to appear at ACM FAccT 2026 (Moore et al.), 17 March 2026
  7. Strengthening ChatGPT's responses in sensitive conversations · OpenAI, 27 October 2025

Spotted an error? Write to hello@siornot.com. Corrections are made in the article and the update date changes. Read our editorial policy.