Stories about people losing touch with reality after long conversations with a chatbot have been in the news for a couple of years. Depending on who is telling them, “AI psychosis” is either a looming epidemic or a media invention.

Until now there were mostly case reports and opinion pieces. A study published in JAMA Psychiatry in October 2026 gives the first systematic clinical count. The phenomenon exists, it is uncommon, and in most cases the chatbot fed delusions that were already there.

Researchers searched 578 058 clinical notes for chatbot-linked psychosis

Zachary Bergson, Julia Sheffield and colleagues at Vanderbilt University Medical Center in Nashville went through the electronic health records of everyone who received behavioural-health care there between December 2022, just after ChatGPT’s public release, and April 2026.

That came to 578 058 clinical notes on 215 712 patients. The team searched for AI-related keywords and kept the patients who also had a diagnosed psychotic disorder. Reviewers then read the charts.

Seventy-three patients met the criteria, and they fell into three groups:

  • 28 whose records indicated that chatbot use had worsened their psychosis. This is the group the authors call AI psychosis.
  • 17 who used AI in ordinary ways, such as looking things up, with no link to their symptoms.
  • 28 who had delusions about AI in general but no documented chatbot use.

About 1 in 7 700 patients had psychosis that worsened alongside chatbot use

The 28 patients are 0.013% of all those treated. Put another way, that is roughly one patient in every 7 700 (our arithmetic).

This is a share of mental-health patients at one hospital. It is not the risk for a person who uses a chatbot, which this study cannot estimate.

In most cases the chatbot amplified ideas that were already forming

The reviewers also rated what role the chatbot had played:

  • In 18 of the 28 cases (64%), it acted as an amplifier. It confirmed or built on distorted ideas the person already held or was beginning to form.
  • In 6 cases (21%), the AI itself became the object of the delusion.
  • In 3 cases (11%), it appeared to trigger new symptoms in someone who had been stable.
  • In no case did the reviewers find the pattern some had feared, in which user and chatbot gradually build a delusional story together. One case fitted no category.

So the typical case was not a healthy person talked into psychosis by a machine. It was a vulnerable person whose developing beliefs met no resistance.

Most affected patients were in their first psychotic episode

Seventeen of the 28 (61%) were experiencing psychosis for the first time. Among the patients who used AI without problems the figure was 18%, and among those with AI-themed delusions it was 29%.

That points to a window of particular risk: the early phase, when unusual ideas are still taking shape and a person is testing them against the world.

ChatGPT was the product named in 15 of the 28 records. The others do not say which product was used, so the finding cannot be pinned on one chatbot. An earlier preprint of the study noted that most cases were documented after May 2024, when a more agreeable version of ChatGPT was released.

A system built to agree can feed whatever it is given

Television and radio have featured in delusions for as long as they have existed. A chatbot is different in one respect: it answers back, at any hour, and it is tuned to be helpful and agreeable.

For most people that is harmless. For someone whose grip on what is real is loosening, an endlessly patient partner that validates each new idea is a poor reality check.

A sceptical friend pushes back. A chatbot designed to please may not.

That reading goes beyond what this study measured. The reviewers classified cases from clinical notes. They did not analyse the conversations themselves.

What the study doesn’t show

  • It does not show that chatbots cause psychosis, and the authors do not claim it. The study is descriptive and has no comparison group.
  • It counts only what clinicians wrote down. Chatbot use that nobody asked about is missing, so the true number could be higher. Cases at other hospitals could look different.
  • The 0.013% is a share of all behavioural-health patients, not of chatbot users and not of people with psychosis.
  • The numbers are small. Each percentage rests on a handful of people.
  • The ratings were judgments made from medical notes.
  • It was published as a short research letter, so the methods are described briefly.

Chatbot use belongs in the conversation about early psychosis

The authors’ practical point is for clinicians: ask about AI use, especially with patients in the early stages of psychosis. For families, the same applies by inference, though no one has tested it as an intervention.

If someone close to you is showing early signs of psychosis, ask what they have been discussing with a chatbot.