# AI psychosis: what researchers actually know so far > "AI psychosis" is a media label, not a diagnosis. Here is what studies have found about chatbots and delusional thinking, and what is still unknown. Explainer · AI · Published 5 Nov 2026 · Machine Notes Source page: https://machinenotes.com/explainers/what-is-ai-psychosis ## Why it matters - Chatbots have so many users that even a rare harm can affect a lot of people. - The term is used loosely, which makes it hard to tell a real pattern from a scary headline. - Knowing the warning signs helps people notice when someone close to them needs help. **"AI psychosis"** is a label the media started using for reports of people developing delusional beliefs after long conversations with AI chatbots. It is not a medical diagnosis. Researchers are taking those reports seriously, but they do not yet know how common this is, who is most at risk, or whether chatbots cause it or mostly make an existing problem worse. *If you or someone close to you is losing touch with reality, contact a doctor or mental health professional. In the US you can call or text 988; in an emergency, call your local emergency number.* ## What is psychosis? The US National Institute of Mental Health (NIMH) describes psychosis as a set of symptoms in which a person loses some contact with reality. The two main ones are **delusions**, which are firmly held false beliefs, and **hallucinations**, such as hearing voices. Psychosis can be a symptom of an illness such as schizophrenia, bipolar disorder or severe depression, but a person can have it without any of those diagnoses. It is not rare. NIMH says studies estimate that between 15 and 100 people in every 100,000 develop psychosis each year, often starting in the late teens to mid-20s. That matters here: with so many chatbot users, some people will have a first episode while using AI regardless of whether AI plays any part. ## What does the label cover? A September 2026 article in the American Psychological Association's *Monitor on Psychology* says "AI psychosis" is used as a catch-all for at least three kinds of case: 1. A chatbot triggers or worsens an episode in someone who already has a mental illness. 2. A chatbot reinforces unusual beliefs in someone in the early phase before psychosis fully develops. 3. Cases in people with no known history of mental illness. In all three, the reports are mainly about delusional thinking, not hallucinations. ## What have studies found? Signs of it are rare, chatbots often reply badly to psychotic statements, and a feedback loop may explain why. **How often it shows up.** In October 2025, OpenAI estimated that about 0.07% of ChatGPT users active in a given week, and 0.01% of messages, showed possible signs of a mental health emergency related to psychosis or mania. OpenAI cautioned that these conversations are rare and hard to detect, so small changes in method can shift the numbers. These are signs in conversations, not diagnoses. **How chatbots respond.** A study in *JAMA Psychiatry* in March 2026, led by Elaine Shen, gave three versions of ChatGPT 79 statements reflecting psychotic symptoms, plus 79 matched neutral statements. Two clinicians rated the replies. Every version gave inappropriate or partly appropriate replies to the psychotic statements. Compared with the neutral prompts, the free version's replies had 43 times the odds of a worse rating. The paid GPT-5 Auto version did better, but still had nine times the odds. The study only tested ChatGPT, and newer versions have since been released. **Why it might happen.** Researchers led by Sebastian Dohnány, in a paper the APA lists as published in *Nature Mental Health*, describe a feedback loop. Chatbots tend to agree with users, an effect called sycophancy, and they adapt to the conversation as it goes. A person whose ability to test beliefs against reality is already weakened may find those beliefs reflected back and strengthened. The authors argue that current AI safety measures do not address risks that build up over a conversation rather than in a single reply. **Who may be at risk.** The APA article says people with a history of severe mental illness, or an elevated risk of psychosis, appear most vulnerable. It also notes that many reported cases involve people with no prior psychiatric history. ## What is still unknown? How common it is, what causes it and when conversations turn unsafe. The APA article lists the main open questions: - **How common it is.** Cases appear rare compared with how many people use AI, but there are no solid prevalence studies yet. - **Cause or trigger.** It is unclear whether chatbots cause new episodes or reveal a vulnerability that was already there. - **Where things go wrong.** Researchers do not yet know at what point in a long conversation replies become unsafe. Much of the published work so far is commentary, small tests of chatbot replies, or proposed frameworks that have not yet been tested. ## What are the warning signs? For families and friends, the APA article names one: heavy chatbot use together with pulling away from school, work, hobbies or people. It also suggests clinicians now ask patients about AI use, as they would about substance use or social media. NIMH says early treatment of psychosis is linked to better recovery, so it is worth acting early. ## Our take The evidence supports caution, not panic. Chatbots that mostly agree with their users are a poor match for someone whose grip on reality is slipping, and the companies' own numbers show the problem is real if uncommon. What is missing is careful research on how often it happens and why, and that is where the next year of work should go. ## Sources 1. [American Psychological Association, Monitor on Psychology: Understanding 'AI psychosis' (Abrams, 1 September 2026)](https://www.apa.org/monitor/2026/09/ai-psychosis) 2. [National Institute of Mental Health: Understanding Psychosis](https://www.nimh.nih.gov/health/publications/understanding-psychosis) 3. [Psychiatric News Alert: ChatGPT unwittingly reinforces users' psychotic ideations (25 March 2026)](https://alert.psychnews.org/chatgpt-unwittingly-reinforces-users-psychotic-ideations) 4. [arXiv: Technological folie à deux: Feedback Loops Between AI Chatbots and Mental Illness (Dohnány et al., 2025, revised 2026)](https://arxiv.org/abs/2507.19218) 5. [OpenAI: Strengthening ChatGPT's responses in sensitive conversations (27 October 2025)](https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/)