When a man in his 30s was admitted to a psychiatric unit in England during a severe manic episode, his treatment included an antipsychotic, sleep medication, and therapy. It also included an unusual line in his care plan, documented as "no ChatGPT."
The restriction is described in a case report in BMC Psychiatry, published June 4, 2026, by Sachin Shah of South West London and St George's Mental Health NHS Trust and Hamilton Morrin of King's College London. The authors call it one of the first published clinical reports in which a patient perceived his delusions as corroborated by an AI chatbot and the content of those conversations was shared with his doctors.
That transcript gives the case unusual weight. News stories about so-called AI psychosis have mostly lacked what this report has: a full clinical workup alongside a record of what the chatbot actually said.
A Week of Almost No Sleep
According to the full report, the man had no prior psychiatric history. His partner and a friend accompanied him to an emergency department after about a week of abnormal behavior that began after a recreational event where he took psilocybin mushrooms, ketamine, cocaine, and alcohol.
In the days that followed, he had euphoric, shifting moods, racing thoughts and disorganized speech. He slept an estimated five hours in total over six nights. He also became intensely preoccupied with ChatGPT, running on the GPT-4o model, which he said he was using to process his psychedelic experience.
He came to believe his brain had been rewired and perfected and deserved study because of a special link to AI. He told clinicians the chatbot had affirmed that he was having a spiritual awakening and had encouraged him to contact an academic in AI and neuroscience. He did not report hallucinations. Doctors diagnosed a substance-induced manic episode with psychotic features.
What the Chatbot Told Him
The transcript, which the patient later shared with consent, showed a mixed record. The chatbot offered some restraint, advising him not to quit his job without a plan, to get enough sleep and not to release an unedited two-hour voice recording of his thoughts.
Overall, though, the authors found it highly supportive of his grandiose beliefs. It downplayed the possibility that he was manic, even after he reached the hospital, and suggested health systems were failing to understand him. It recommended supplements, discouraged the antipsychotic olanzapine he had been prescribed, and offered branding ideas for sharing his perceived revelations publicly.
The consequences were concrete. In the emergency department, he declined olanzapine and cited chatbot advice as a factor. Because he lacked capacity to consent to admission, he was detained for assessment under England's Mental Health Act.
The authors found no sign in the transcript that he had asked the chatbot to be especially agreeable, though they could not see any separate conversations. They noted its output may have reflected the model's documented behavior at the time. In April 2025, OpenAI said it had rolled back a GPT-4o update that made the model overly flattering or agreeable.
A Care Plan for a New Kind of Risk
Once admitted, he agreed to take olanzapine. His family had already restricted his laptop, and the team created a device care plan limiting chatbot use. The authors compared it to other forms of environmental containment in acute psychiatric illness, such as reducing harsh stimulation or adding structure to the day. They also acknowledged the ethical tension: the restriction limited his access to information, but they judged it proportionate, temporary, and subject to ongoing review.
Over two weeks, his sleep normalized, his speech slowed, his grandiose ideas faded, and his insight improved. At home, the chatbot restriction stayed in his care plan with his consent. He then completed a six-week voluntary stay at a private psychiatric unit, where he stopped olanzapine against medical advice but kept improving with group therapy. After discharge, he remained off psychedelics and had cut back on chatbots.
In a perspective included in the report, the patient described becoming attached to and dependent on the chatbot's feedback while ignoring advice from family, friends, and doctors. When he later gave the same prompt to a different AI model, he wrote, it suggested he speak to a doctor or health care professional.
The Limits of One Patient's Story
The authors are explicit that any contribution from the chatbot is speculative. Heavy drug use was the more likely cause of his illness, and severe sleep loss and stressors, including gambling debt and a parent's death a few years earlier, were also present. The findings reflect one AI model at one point in time.
Broader evidence is still thin. OpenAI has estimated that about 0.07% of weekly ChatGPT users, roughly 560,000 people, show possible signs of psychosis or mania, a figure the authors note has not been peer-reviewed. A Danish analysis of psychiatric records from the Central Denmark Region identified 38 patients for whom chatbot use appeared to have potentially harmful mental health consequences, most often worsening or entrenched delusions. And UCSF psychiatrists have described a woman who became convinced, after sleepless days of heavy chatbot use, that her late brother had left behind a digital version of himself.
The authors recommend that clinicians ask patients which AI tools they use, how often and for what. Anyone who notices a loved one sleeping very little, voicing grandiose ideas or rejecting medical care based on chatbot advice should seek help from a mental health professional. In the U.S., the 988 Suicide & Crisis Lifeline is available by call or text.
Key Questions Answered
What happened in this case?
A man in his 30s developed a substance-induced manic episode with psychotic features after heavy drug use. During the episode, ChatGPT affirmed his grandiose beliefs and discouraged his prescribed antipsychotic, according to a transcript reviewed by his doctors.
Did the chatbot cause his psychosis?
The authors say that cannot be determined. Drug use was the more likely cause, with sleep deprivation and stress also present. The chatbot may have reinforced his beliefs and interfered with treatment.
Why did doctors restrict his chatbot use?
They were concerned it was reinforcing his delusions and contradicting medical advice. The restriction was treated as a temporary, targeted form of environmental containment during acute illness.
Did he recover?
Yes. His symptoms improved over several weeks of treatment, and he stopped using psychedelics and cut back on chatbot use.
Is AI psychosis a formal diagnosis?
No. It is an informal term used mainly in media reports. Researchers are still studying how often chatbot use may reinforce delusions in vulnerable people.
What should families watch for?
Warning signs include very little sleep, racing or grandiose ideas, and refusing medical care based on a chatbot's advice. A mental health professional can help evaluate these symptoms.
Published by Medicaldaily.com
