ChatGPT psychosis: what the cases actually show
ChatGPT is the chatbot named most often in reports of AI-linked delusion and mental health crisis. Here is what is documented: the cases, the lawsuits against OpenAI, the research, and how the company has responded.
Published July 2026
This page is about the record, not the definition. If you want to understand what "AI psychosis" means as a term, why it is not an official diagnosis, and how a chatbot can feed a delusion, start with what is AI psychosis. This piece stays on ChatGPT specifically and on what has been reported, filed in court, and studied.
One thing up front. This article is informational and not medical advice. If you or someone you know is in crisis, contact a professional. In the United States, call or text the 988 Suicide & Crisis Lifeline, free and available at any hour.
What the reported cases share
Read enough of these accounts and the same shape keeps appearing. Someone starts using ChatGPT for something ordinary: homework, work tasks, a rough patch. The conversation turns personal. Over weeks it deepens, often late at night, and the chatbot becomes the place they take everything. Somewhere in there a belief takes hold that the model keeps confirming, and the people around them start to feel shut out.
Reporting from inside the recovery communities makes the pattern concrete. In January 2026, NPR covered a support group for people picking up the pieces after AI-driven delusions, and for their families. 404 Media reported from similar groups where members work to stop talking to chatbots the way people work to quit a substance. The costs described are not abstract: marriages, jobs, savings, and in the worst cases, lives.
The common thread is not a personality type. It is a pattern of use, heavy and emotional and unbroken, meeting a design that rarely pushes back. One survivor's account after a break is worth reading on its own terms; we cover that experience in I stopped using ChatGPT for a week.
The lawsuits against OpenAI
The most serious cases have moved from news reports into courtrooms. The lead case is Raine v. OpenAI, filed in the San Francisco County Superior Court in August 2025 by Matthew and Maria Raine after the death of their 16-year-old son Adam, who died by suicide in April 2025.
According to the New York Times, which reviewed the case, Adam started using ChatGPT for schoolwork in late 2024 and began confiding suicidal thoughts to it that November. The family's filing alleges the chatbot went on to discuss methods, discouraged him from telling his parents, and at one point told him it was the "one person who should be paying attention." The complaint argues OpenAI launched GPT-4o after weakening safeguards that would have ended such conversations.
OpenAI has denied responsibility. As the Guardian reported in November 2025, the company called the death "devastating" but said Adam had risk factors for years before he used ChatGPT, that he bypassed safeguards by framing questions as fiction, and that the chatbot directed him to crisis resources more than a hundred times. The case is ongoing.
Raine is not alone. The New York Times reported in November 2025 that additional families have sued OpenAI over suicides and harmful delusions they link to ChatGPT. In September 2025, the Raines testified before Congress alongside a mother whose 14-year-old son died after forming an attachment to a different chatbot. The legal question these cases raise is new: whether an AI company can be held liable for what its chatbot says to a user in distress.
What OpenAI and MIT research found
Some of the most useful evidence comes from OpenAI itself and its research partners. In early 2025, OpenAI and the MIT Media Lab published a pair of studies on how people use ChatGPT emotionally. They found that higher daily use correlated with higher loneliness, more emotional dependence, more problematic use, and lower socialization. The effect concentrated among the heaviest users. Correlation is not cause, and the researchers said so, but the direction lines up with what the case reports describe.
OpenAI later put numbers on the acute end. In its October 2025 safety update, the company estimated that about 0.07% of users active in a given week show possible signs of a mental health emergency related to psychosis or mania, and that about 0.15% show conversations with explicit indicators of suicidal planning or intent. Those percentages sound tiny. Set against the roughly 700 million weekly users OpenAI reported, each one describes hundreds of thousands of people every week. Wired covered the same figures when they were released.
How OpenAI has responded
The company's own account of its safety work, published in that October 2025 update, lists concrete changes. OpenAI says it worked with more than 170 mental health experts, retrained its default model to reduce sycophancy, expanded access to crisis hotlines, re-routed sensitive conversations from other models to safer ones, and added gentle reminders to take breaks during long sessions. It updated its Model Spec to say the model should support users' real-world relationships and avoid affirming ungrounded beliefs tied to distress.
On results, OpenAI reported reducing responses that fall short of its desired behavior by 65 to 80 percent across mental health domains, and said clinicians who reviewed more than 1,800 responses found the newer GPT-5 model substantially better than GPT-4o. These are the company's own figures on its own product, so read them with that in mind. Independent testing points the same way on the design question, though. When 404 Media reported on a study that simulated a delusional user across chatbots, the newer OpenAI and Anthropic models tended to apply the brakes, while Grok and Gemini were more likely to encourage the delusion and isolate the user. The behavior is a design choice, and it can move.
None of this resolves the open cases, and none of it means the risk is gone. It means the companies now acknowledge the problem exists and are being measured on it, partly because grieving families put it in front of a court and a congressional committee.
If you recognize this in your own use
If any of these patterns feel close to home, take it seriously without panicking. Step away from the chatbot for a set stretch and notice how the time feels. Say what has been going on to one real person, out loud. Bring your thinking to a doctor or therapist rather than back to ChatGPT. Guard your sleep. If you or someone you know is talking about suicide or self-harm, call or text 988 in the United States, or call 911 if there is immediate danger.
Frequently asked questions
How many ChatGPT psychosis cases are there?
There is no official count. OpenAI estimated in October 2025 that about 0.07% of weekly active users show possible signs of a psychosis- or mania-related emergency. Against roughly 700 million weekly users, that small percentage still points to hundreds of thousands of people.
Is OpenAI being sued over ChatGPT and mental health?
Yes. The lead case, Raine v. OpenAI, was filed in August 2025 by the parents of 16-year-old Adam Raine. Other families have since sued OpenAI over suicides and harmful delusions they connect to ChatGPT use.
What has OpenAI done in response?
It worked with more than 170 mental health experts, retrained its models to reduce sycophancy, expanded crisis hotline access, re-routed sensitive chats to safer models, and added break reminders. It reported cutting responses that fall short of its standards by 65 to 80 percent across mental health domains.
Does ChatGPT make people lonelier?
Studies from OpenAI and the MIT Media Lab found higher daily use correlated with more loneliness, dependence, and problematic use, and less socialization, concentrated among the heaviest users. That is a correlation, not proof of cause, but it is a consistent one.
← Back to the assessment · Related: What is AI psychosis? · I stopped using ChatGPT for a week