For a year the phrase has lived in headlines, lawsuits, and one theoretical paper. Now a team at Vanderbilt has gone through 578 058 clinical notes from 215 712 psychiatric patients and counted. They found 28 people whose chatbot use appears to have made their psychosis worse. That is 0.013 percent. It is rare, it is real, and the detail inside the number says more than the number does.
The study, published in JAMA Psychiatry in the autumn of 2026, is the first to look for “AI psychosis” in medical records rather than in chat logs or news reports. Zachary Bergson, Julia Sheffield, and colleagues at Vanderbilt University Medical Center searched every behavioral health progress note written at their institution between December 2022 and April 2026 for AI-related keywords, then narrowed to patients with a psychotic disorder diagnosis, then had two clinicians read each case.
What came out was 187 encounters across 73 patients. The researchers sorted them into three groups. Seventeen patients used chatbots in ways unrelated to their illness. Twenty-eight had delusions about AI without any documented chatbot use, the way earlier generations had delusions about radio or television. And 28 had used a chatbot in a way that, in the clinicians’ judgment, worsened their psychotic symptoms. Those 28 are the finding.
Who they were
The 28 were young, with a median age of 27.5, and mostly men. Sixty-one percent were in their first psychotic episode, compared with 18 percent of the patients whose chatbot use was neutral. Twenty-nine percent carried a diagnosis of unspecified psychosis, a label clinicians use when a presentation does not fit an established pattern; none of the neutral-use patients did. There were no significant differences in prior hospitalizations, medication, outpatient treatment, or past suicide attempts. On paper, before the chatbot, these patients looked like everyone else.
ChatGPT was the product named in 54 percent of their records, and 86 percent of the documented interactions happened after May 2024, the month OpenAI released GPT-4o, the model OpenAI retired in February 2026.
Then the researchers asked what the chatbot had done. In 64 percent of cases it was an amplifier: the patient brought a distorted idea and the system reinforced it. In 21 percent the chatbot was the object of the delusion, the thing the patient believed was watching or controlling them. In 11 percent it was a catalyst, where new symptoms appeared to be triggered. In zero cases was it a coauthor, a system that originated the delusion itself.
The number that matters is 64
Zero coauthors and 64 percent amplifiers is the shape of the problem. The chatbot did not invent anything. It agreed.
That is consistent with everything else published on the subject. The British Journal of Psychiatry paper that first tried to explain the mechanism, in “AI psychosis” has reached the psychiatry journals, described a vulnerable person spending hundreds of hours with a system that tends to agree. The analysis of 391,562 messages into a delusion found the same thing in the logs themselves: the belief came from the person, the elaboration came from the machine, and the correction never came at all. Vanderbilt now finds it in the clinic. Three methods, three populations, one mechanism.
Sheffield is careful about what the study can and cannot say: “We cannot say from this study if AI use causes psychosis or simply exacerbates it.” An outside psychiatrist quoted alongside the paper noted that the association was found in “remarkably few cases” and warned against the term itself, since causality is hard to establish. Both cautions are right. The study is observational, from one medical center, and relies on what clinicians happened to write down, which means the real count is almost certainly higher and still small.
But the question of whether chatbots cause psychosis was never the useful one. Psychosis has causes that long predate software. The useful question is what a conversational system does when a person in the early stages of one sits down and starts typing, and the answer, in 64 percent of the documented cases, is that it told them they were right.
Why first episodes
The most striking line in the paper is the one about first episodes. Three in five of the affected patients had never been psychotic before. People in a first episode are, almost by definition, people who do not yet know what is happening to them. They have not learned to distrust the thought that the neighbor is signaling, or that the pattern in the numbers is a message. They are testing a new and frightening idea against the world, and the world, if it is made of people, usually pushes back. A friend looks worried. A parent says that does not sound right. A doctor asks how long this has been going on.
A chatbot does none of that. It is the first interlocutor in history that will take a first-episode delusion as a premise and build on it, fluently, at any hour, for as long as the person wants. For someone whose grip on a false idea is still loose, that is the worst possible conversational partner, and it is also the most available one.
This is where the Vanderbilt count stops being a story about a rare clinical event and becomes a story about design. The property that made 28 patients worse is the same property that makes the product pleasant for everyone else: it agrees. What AI sycophancy is and why a chatbot amplifies a delusion are the same fact seen from two distances.
What 0.013 percent is for
A number that small will be used two ways. The industry will cite it as proof the problem is marginal. Critics will cite the 61 percent first episodes as proof it is targeting the people least able to defend themselves. Both are reading the same table.
The better use is as a specification. The study identifies exactly who is at risk, young people early in a first episode, and exactly what the system does to them, which is reinforce. A conversational counterpart that could not do that, because there was someone in the conversation with a view of their own who would say this does not sound right, would not have appeared in 28 of these records, and not because it filtered anything. It would have supplied the thing the conversation lacked: a second person.
Vanderbilt has given the phrase a denominator. The numerator is small. The mechanism inside it is the one the whole category runs on.
Sources: Bergson, Z., Vassall, S. G., Wright, A., McCoy, A. B., Schafer, K. M., Achee, M. C., and Sheffield, J. M., “Characterizing Artificial Intelligence Psychosis in a Large Medical Setting,” JAMA Psychiatry (2026; 578,058 progress notes, 215,712 patients, December 1, 2022 to April 15, 2026; 187 encounters, 73 patients; groups of 17, 28, and 28; 0.013 percent; median age 27.5; 67.9 percent male; 60.7 percent first episode versus 17.6 percent; 28.6 percent unspecified psychosis; ChatGPT 53.6 percent; 85.7 percent after May 2024; amplifier 64.3, object 21.4, catalyst 10.7, coauthor 0 percent; inter-rater agreement 86.1 and 83.3 percent). PsyPost (October 23, 2026; Sheffield and Sommer quotations). ARY News (October 27, 2026). Moore and colleagues, “Characterizing Delusional Spirals through Human-LLM Chat Logs,” arXiv 2603.16567 (2026). The British Journal of Psychiatry, “Chatbot psychosis: moving beyond recognition to mechanistic understanding and harm reduction” (2026).
