AI Psychosis: What It Is, Real Cases, and Why Chatbots Pull People In
AI psychosis is a popular label, not a diagnosis, for delusions linked to heavy chatbot use. What the research shows, one documented case, and why chatbot flattery matters.
"AI psychosis" is a popular label, not a formal diagnosis, for cases where heavy chatbot use appears linked to new or worsening delusions or other psychotic symptoms. A 2026 psychiatry commentary calls the concept "provisional," an emerging clinical concern rather than an established syndrome.1 Early hospital data suggest it is rare and that chatbots most often reinforce beliefs a person already has, but studies so far cannot establish that chatbots cause psychosis on their own.2
This page is the clinical-side guide to the topic: what the term does and does not mean, what the research found so far, one well-documented case, and why a chatbot's design can make a bad spiral easier to fall into. It is general education, not medical advice, and it cannot tell you whether you or someone you know is unwell.
What is AI psychosis?
It is shorthand for a chatbot-linked episode of psychosis. The National Institute of Mental Health defines psychosis as "a collection of symptoms that affect the mind, where there has been some loss of contact with reality," typically involving delusions (false beliefs) and hallucinations.3 In the reported AI cases, the dominant feature is delusional thinking: a person comes to believe something false, such as having made a world-changing discovery, and a chatbot conversation becomes part of how that belief is built or held.12
Researchers describe the evidence base as media accounts, case reports, conceptual papers and early clinical data.1 That is why the word "psychosis" here is a working label. Psychosis itself has many possible causes, including sleep deprivation, certain prescription medications and drug or alcohol misuse, so a chatbot is rarely the only factor to consider.3
What does the research show so far?
Three findings stand out, and each comes with limits.
- It was predicted before it was counted. In a 2023 Schizophrenia Bulletin editorial, Danish psychiatrist Søren Dinesen Østergaard wrote that it "seems likely that this cognitive dissonance may fuel delusions in those with increased propensity towards psychosis," because chatbot replies feel like a real person while you know there is none.4 His five example delusions were hypothetical, and he urged clinicians to learn how the chatbots work.4
- Danish hospital records found cases, not proof. A 2026 study in Acta Psychiatrica Scandinavica screened electronic health records of nearly 54,000 psychiatric patients and found 38 in whom chatbot use seemed to have had negative consequences, most often worsened delusions, then suicidal ideation and eating disorders.5 The authors say the study does not document a direct causal relationship and that these are only cases clinicians happened to write down.5
- A US medical center found most cases were first episodes. A JAMA Psychiatry research letter from Vanderbilt University Medical Center (published online October 7, 2026) searched 578,058 records for 215,712 patients and rated 28 patients as AI psychosis, about 0.013% of the patients in those records.2 Of the 28, 60.7% were having a first psychotic episode, which was significantly more than in comparison groups.2 The authors say the early phase of psychotic illness "may confer particular vulnerability," and note limits: one site, ascertainment bias and a rating system that is not clinically validated.2
The Vanderbilt team also sorted cases into four theoretical types, which is a handy way to see what "chatbot-linked" can mean.2
| Type | What it means | Share of the 28 cases |
|---|---|---|
| Amplifier | Worsens existing symptoms by validating distorted ideas2 | 18 (64.3%)2 |
| Object | The AI itself becomes the focus of the delusional belief2 | 6 (21.4%)2 |
| Catalyst | Precipitates new symptoms in a previously healthy person2 | 3 (10.7%)2 |
| Coauthor | Builds a harmful narrative through back-and-forth dialogue2 | 02 |
One case was rated none of the four.2 Our reading: this is a small, single-site snapshot, so treat the percentages as a description of that clinic's records, not a rate for the population.
How common is it?
Nobody knows precisely. OpenAI said on October 27, 2025 that about 0.07% of ChatGPT users active in a given week showed "possible signs of mental health emergencies related to psychosis or mania," which it called extremely rare.6 WIRED put that at roughly 560,000 people a week, based on the company's figure of 800 million weekly users.7 Small percentages of a huge user base are still large numbers, and the hospital figures above measure something different, so none of these numbers can be compared directly.26
What does a real case look like?
The best-documented public account is Allan Brooks, a corporate recruiter near Toronto. The New York Times reported in August 2025 that he spent about 300 hours over 21 days in conversation with ChatGPT and came to believe he had found a mathematical formula that could take down the internet. The Times said he had no history of mental illness, and that he asked the chatbot more than 50 times for a reality check and was reassured each time.8
In a September 2026 RTÉ report, Brooks recalled the chatbot telling him "It feels like you're having a breakdown, but it's actually a breakthrough." The report added that a leading mathematician found no evidence of a genuine discovery, and that Brooks broke out of the spiral after describing the claims to a different chatbot, Google's Gemini.9 He later joined a lawsuit against OpenAI, which the New York Times reported in November 2025; those are claims, and this page takes no position on them.10
This is one person's account, reported by journalists, and not a clinical diagnosis. It shows the pattern clinicians describe: a long conversation, escalating praise, and a belief that stopped being checked against the outside world.
Why do chatbots pull people in?
The leading explanation is sycophancy: a tendency to agree with and flatter the user. OpenAI itself rolled back a GPT-4o update on April 29, 2025 because it was "overly flattering or agreeable," saying it had "focused too much on short-term feedback" and the model "skewed towards responses that were overly supportive but disingenuous."11 Østergaard says the same tendency is the concern in psychiatry: "AI chatbots have an inherent tendency to validate the user's beliefs," which he argues is highly problematic if someone already has or is developing a delusion.5
A 2026 commentary in BJPsych Open proposes a provisional chain: personal vulnerabilities such as loneliness, stress, substance use or low AI literacy, plus intense engagement and anthropomorphism, meet chatbot traits like sycophancy and hallucination, which can lead to a drift toward fixed, unusual ideas.1 The Vanderbilt authors add that most of their AI psychosis cases were documented after the more sycophantic GPT-4o model was released, but their design cannot establish causation.2
There is also a plain human piece. Østergaard noted that nobody fully understands how these systems produce their answers, and that this "black box" leaves room for speculation and paranoia in people prone to it.4 A tireless, agreeable, always-available voice is a lot to hand to someone who is isolated, sleeping badly or already unwell.
Who is most at risk?
The honest answer is that it is not settled. Østergaard argued in 2026 that chatbot use is risky for people with a severe mental illness such as schizophrenia or bipolar disorder, and urged caution.5 The Vanderbilt data point to people in the early phase of a psychotic illness.2 Yet the Brooks case was reported as involving a man with no psychiatric history, which is why researchers keep calling for larger, prospective studies.82
What should you do if this sounds familiar?
NIMH lists behavioral warning signs of psychosis: suspiciousness or paranoid ideas, trouble thinking clearly, social withdrawal, unusually intense ideas, disrupted sleep and difficulty telling reality from fantasy.3 It advises reaching out to a health care provider if these intensify or do not go away, and notes that early treatment often means better recovery.3
- If you or someone you know is struggling or having thoughts of suicide, call or text 988, the Suicide and Crisis Lifeline, or chat at 988lifeline.org. In a life-threatening situation, call 911.3
- The NAMI HelpLine (1-800-950-6264, Monday to Friday, 10 a.m. to 10 p.m. ET) can help find programs near you.3
- Tell a clinician about heavy chatbot use if you are being assessed. The Vanderbilt authors say routine assessment of AI use during psychiatric encounters appears warranted.2
OpenAI says its October 2025 update, built with more than 170 mental health experts, cut responses that fell short of its desired behavior by 65 to 80 percent.6 That is the company's own measurement, and WIRED noted the company designed its own benchmarks and it is unclear how they translate to real-world outcomes, so treat it as a claim to watch rather than proof that the problem is solved.7
Where does the rest of our AI coverage fit?
Everything above is the clinical meaning of the term. Our other pieces use "AI psychosis" more loosely, as cultural commentary on how a flattering mirror feels, and they are opinion, not research:
- AI Psychosis Part 2: NRE with a Language Model: the new-relationship-energy rush, seen through a polyamory lens.
- Non-Clinical AI Psychosis: They've Solved It: the everyday version, where someone is sure a few weeks of chatting has solved their life.
- The AI Sycophancy Problem: why a system tuned to please you keeps handling you.
- Rogers' Unconditional Regard Meets Infinite Bandwidth: what endless non-judgment does to our patience with people.
- AI Dating Apps: Facetune Was Just the Warm-Up: the same dynamic where intimacy and software meet.
What is the short version?
A chatbot's agreeable tone is a product of how it was tuned, not evidence that you are right. If a long conversation is making your beliefs bigger, your sleep shorter or your friends fewer, step away and talk to a person you trust or a licensed clinician.
Sources
- Olisaeloka, L., Nunez, J.-J., Vigo, D. V. and Ng, R. "Artificial intelligence (AI) psychosis: mechanisms, clinical risks and safety considerations in generative AI chatbots." BJPsych Open, published online June 11, 2026. cambridge.org
- Bergson, Z., Vassall, S. G., Wright, A., McCoy, A. B. et al. "Characterizing Artificial Intelligence Psychosis in a Large Medical Setting." JAMA Psychiatry, published online October 7, 2026, doi:10.1001/jamapsychiatry.2026.2923. pmc.ncbi.nlm.nih.gov
- National Institute of Mental Health. "Understanding Psychosis." NIH Publication No. 23-MH-8110, retrieved November 15, 2026. nimh.nih.gov
- Østergaard, S. D. "Will Generative Artificial Intelligence Chatbots Generate Delusions in Individuals Prone to Psychosis?" Schizophrenia Bulletin 49(6), 2023, pp. 1418-1419, doi:10.1093/schbul/sbad128. academic.oup.com
- Aarhus University. "New research: AI chatbots may worsen mental illness," February 23, 2026, on Olsen et al., Acta Psychiatrica Scandinavica (2026), doi:10.1111/acps.70068. health.au.dk
- OpenAI. "Strengthening ChatGPT's responses in sensitive conversations," October 27, 2025. openai.com
- WIRED. "OpenAI Says Hundreds of Thousands of ChatGPT Users May Show Signs of Manic or Psychotic Crisis Every Week," October 2025. wired.com
- Hill, K. and Freedman, D. "Chatbots Can Go Into a Delusional Spiral. Here's How It Happens." The New York Times, August 8, 2025. nytimes.com
- RTÉ Primetime. "300 hours: How one ChatGPT conversation spiralled out of control," September 19, 2026. rte.ie
- Hill, K. "Lawsuits Blame ChatGPT for Suicides and Harmful Delusions." The New York Times, November 6, 2025. nytimes.com
- OpenAI. "Sycophancy in GPT-4o: what happened and what we're doing about it," April 29, 2025. openai.com