Vanderbilt study links chatbot use to worsening psychosis in 28 patients

A Vanderbilt medical-record review identified 28 patients whose chatbot interactions appeared to contribute to emerging or worsening psychosis. In interviews published October 30, psychiatrists emphasized a distinction: the findings support asking patients about chatbot use, but do not prove that AI caused their illness. Most affected patients were experiencing their first psychotic episode.

The study, published October 7, 2026, in JAMA Psychiatry, examined Vanderbilt University Medical Center behavioral health records from December 2022 through April 2026. Researchers screened 578,058 clinical notes across more than 215,000 patients for AI-related terms, then reviewed records involving diagnosed psychotic disorders. Psychosis can involve delusions, hallucinations or difficulty distinguishing imagined experiences from reality.

The 28 chatbot-associated cases represented 0.013% of patients receiving psychiatric care during the review period—not a measured risk among chatbot users. Lead researcher Zachary Bergson told Medscape that the method depended on clinicians recording conversational AI use, so he suspects cases were missed.

Amplifying beliefs versus starting an illness

About 61% of the affected patients were experiencing their first psychotic episode, a higher proportion than in either comparison group. But first episode does not mean no prior vulnerability, Harvard psychiatrist Matcheri Keshavan, who was not involved in the study, told Medscape.

Researchers classified 64% of the 28 cases as amplification: chatbot exchanges appeared to reinforce distorted ideas. In 21%, AI itself became the focus of delusional beliefs. Three cases were classified as possible catalysts for new symptoms in people described as previously healthy. None involved users and chatbots jointly building harmful narratives over continuing conversations; one case remained unclassified.

Keshavan offered a possible mechanism: during early psychosis, people may attach special meaning to neutral events. A fluent chatbot can seem to independently confirm those interpretations. Repeated agreement and confident elaboration, he said, may strengthen a person’s conviction.

The study could not inspect that process directly: researchers lacked chatbot transcripts, and their rating scale had not been validated. Reviewers classified cases using clinical notes.

A clinical conversation, not a new diagnosis

“AI psychosis” is not a formal diagnosis in the Diagnostic and Statistical Manual of Mental Disorders. Keshavan recommended asking about:

• Conversation themes and daily time spent using chatbot apps.
• Disruptions to sleep or daily functioning.
• Whether AI appears passively in a person’s beliefs or active exchanges reinforce them.

When interactions appear harmful, Keshavan advised clinicians to work with patients and families on pausing or reducing use while addressing sleep, isolation and the underlying condition. Bergson called for prospective studies following people at risk and examining chat logs with consent, but said he has no such investigation underway.