A Chart Count Is Not an AI Psychosis Rate

October 24, 2026 | The Psychiatric Record | Assessment

A careful history for a first psychotic episode covers sleep, cannabis, stimulants, recent stress, and family history. Whether it also asks about conversational AI is less settled. If no one asks, chatbot use reaches the chart mainly when someone volunteers it. That gap shapes how we should read a new US chart review of what has come to be called "AI psychosis."

In a research letter published online in JAMA Psychiatry on October 7, Bergson and colleagues searched progress notes at Vanderbilt University Medical Center from December 2022 to April 2026. They began with patients receiving behavioral health care whose notes contained AI-related keywords, kept only those with a documented psychotic-disorder diagnosis, and had raters read the notes by hand.

Among 578,058 records from 215,712 unique patients, 73 patients met criteria after screening. Raters placed 28 in an "AI psychosis" group, which the study defines as psychosis-worsening chatbot use. Another 28 had AI-related psychotic content, meaning AI was incorporated into a delusion without documented chatbot use, and 17 had neutral interactions, such as using ChatGPT to look up recipes.

Raters also sorted the 28 cases into an a priori typology. The most common was "amplifier," in which chatbot use was judged to worsen existing symptoms by validating distorted ideas (18 of 28). The chatbot was the focus of delusional belief in 6 cases, and a "catalyst" for new symptoms in a previously healthy person in only 3. These are readings of clinical notes, made without chat logs, using a rating system the authors describe as not clinically validated.

The authors report that the 28 patients represent 0.013% of those receiving mental health care. That is a documented proportion in one health system, not a reliable estimate of how often chatbot use worsens psychosis, and not a measure of causal risk. The paper itself describes the figure two ways: as a share of all behavioral health patients in its Results, and as conditional on a psychosis diagnosis in its Discussion.

A keyword search of notes can only find what a clinician wrote down, in words the search recognized. The authors list ascertainment bias, a single site, and the unvalidated rating system among their limitations. A Danish study of psychiatric records, which searched for "chatbot" and "ChatGPT," was more explicit. Because its patients were not systematically questioned about chatbot use and its search terms were narrow, the authors wrote, its results do not speak to incidence.

The Vanderbilt count therefore reflects cases that were noticed, recorded, and judged to fit. Unrecorded cases would push the true figure up; cases in which chatbot use was incidental to the illness would push it down. The authors state that their design cannot establish causation.

Seventeen of the 28 patients in the AI psychosis group (60.7%) were in their first psychotic episode, compared with 3 of 17 in the neutral group and 8 of 28 in the AI-related content group. The authors suggest the early phase of psychotic illness may confer particular vulnerability. A first episode is not the same as a chatbot-initiated one; raters judged only three cases to be catalysts.

The vulnerability reading is plausible, but it is not the only one. One untested explanation is that clinicians record chatbot use more readily at a first presentation, when they are looking for a precipitant. The direction of any association is also open. In a cross-sectional US survey of 1,003 young adults, Buck and Maheux found that those with elevated psychosis-risk scores were no more likely to have used generative AI but were more likely to report intensive use. Such a survey cannot show whether symptoms influence use, use influences symptoms, or other factors influence both, and a risk-screen score is not a first psychotic episode.

That makes the first-episode finding a hypothesis for systematic questioning to test, not an established vulnerability.

The Danish notes also show that delusions are not the only concern. Among 53,974 patients with psychiatric service contact, notes for 38 were judged compatible with potentially harmful chatbot use. Delusions were the most common category, followed by suicidality or self-harm and eating disorders. The same notes described 32 patients whose use seemed constructive, such as psychoeducation or companionship, though the study did not measure clinical benefit and the authors note these tools were generally not developed or validated for such purposes. They concluded that some patients would likely benefit from reduced or no use in their current form, a judgment for the individual case rather than a blanket rule.

The case for asking rests on neither the 0.013% figure nor the first-episode finding. It rests on taking a complete history of something many patients now do. The American Psychological Association's practitioner resource, published in June, recommends proactively asking patients about AI use, exploring what they have found helpful or unhelpful, and considering a question about AI use on intake forms. The Vanderbilt authors close on a similar note: routine assessment of AI use during psychiatric encounters, they write, appears warranted.

Our September piece on family work covered online life in psychosis care generally. A chatbot raises a narrower question, because it answers back. For a person presenting with psychosis or mania, these chatbot-specific details may be worth recording, as an editorial proposal rather than a validated screen:

The third item adapts a viewpoint by Morrin and colleagues, who suggest tracking whether a person still entertains counterevidence and caution that a transcript alone cannot reveal a person's mental state. The last item follows the APA resource's suggestion that patients might bring outputs to sessions.

Recording the answer informs the plan for the person in front of us, and it gives future studies a denominator of patients who were asked, rather than a count of incidental mentions.

We found no published trial of whether reducing chatbot use changes the course of psychotic illness. Until asking is routine, chart-based estimates will remain sensitive to how clinicians ask and document.

At the next first-episode assessment, is chatbot use part of the history, or does it appear only if the patient brings it up?

Source note: This commentary interprets two retrospective record reviews, one cross-sectional survey, one viewpoint, and one professional practice resource for clinical assessment. It does not estimate how often chatbot use causes or worsens psychosis, evaluate any AI product, or propose a validated screening instrument; the history items above are offered for clinical consideration and have not been validated as a screening tool. Søren Dinesen Østergaard is an author of both sources 2 and 4.

Educational Disclaimer: The Psychiatric Record provides general educational information for psychiatric and mental-health professionals. Content does not constitute medical, legal, regulatory, compliance, billing, or other professional advice; does not establish a standard of care; and is not a substitute for independent professional judgment. Appropriate assessment and treatment depend on the individual, setting, and current authoritative guidance. Patients should coordinate treatment decisions with their treating clinicians.