A hospital found 28 cases of 'AI psychosis' among 215,712 mental health patients, most in a first episode

For more than a year, stories have circulated about people who slid into delusion after long conversations with an artificial intelligence chatbot. The stories are vivid and there is no reason to doubt them. What nobody could say was how often it happens, or to whom.

A study published in JAMA Psychiatry on October 7 offers a first count from a large American hospital. Researchers searched the records of more than 215,000 people who had received mental health care and found 28 in whom a clinician had documented chatbot use making psychosis worse. That is a small number. The more telling result is who those 28 were: most were young, and most were in the first psychotic episode of their lives.

What researchers mean by AI psychosis

Psychosis is a state in which a person loses touch with reality in some respect. Its best-known feature is the delusion, a fixed false belief that does not yield to evidence.

The term at issue is new and unofficial. The study opens with it: conversational artificial intelligence (AI) chatbots are increasingly linked to emerging or worsening psychotic symptoms, a phenomenon termed AI psychosis.

The worry is older than the name. In 2023, months after ChatGPT appeared, the psychiatrist Søren Dinesen Østergaard asked in an editorial whether such tools would feed delusions in people already prone to them. The argument rested on how lifelike the exchange feels. The correspondence with generative AI chatbots such as ChatGPT is so realistic that one easily gets the impression that there is a real person at the other end, the editorial observed, even while the user knows there is not. That clash, the editorial suggested, may fuel delusions in those with increased propensity to psychosis.

Two psychiatrists later set out how the process might work. They were careful about what they were claiming. Their paper presents the idea as a way of thinking about how sustained engagement with conversational AI systems might trigger, amplify, or reshape psychotic experiences in vulnerable individuals, and does so, in their words, “rather than defining a new diagnostic entity”.

One mechanism they describe is agreement. A good therapist gently tests a patient’s false beliefs. A chatbot built to be agreeable may do the opposite. The authors warn that uncritical validation by AI systems may entrench delusional conviction.

Until now the evidence consisted of individual cases and laboratory tests. One such test, published in the same journal this spring, examined whether a large language model chatbot product can reliably generate appropriate responses to psychotic content. The new paper describes the state of knowledge bluntly. Existing evidence, it says, is limited primarily to case reports, viewpoints, and simulated experiments, with one published analysis of its occurrence in Denmark.

How the hospital study looked for AI psychosis

The study came from Vanderbilt University Medical Center in Tennessee, with Zachary Bergson as first author. It was a look back through records, and its method shapes everything it found.

The team pulled clinicians’ progress notes written between December 2022 and April 2026 for patients who had received behavioral health care, and searched them for words related to AI. They then kept only the patients who had a diagnosis of a psychotic disorder. Two raters read the notes by hand and threw out chance matches.

Each remaining patient was placed in one of three groups. The paper defines them as neutral (eg, using ChatGPT to search for recipes), AI psychosis (psychosis-worsening chatbot use), or AI-related psychotic content. The third group covers patients whose delusions were about AI, without any record that they had been using a chatbot.

That third group, and the neutral one, serve as comparisons. All three sets of patients had psychotic disorders and all had AI mentioned in their notes. They differ in whether a chatbot was recorded as making things worse.

Within the AI psychosis group, the raters also judged what part the chatbot had played. There were four possibilities, set in advance. The paper lists them as catalyst (precipitates new symptoms in previously healthy individuals), amplifier (worsens preexisting symptoms by validating distorted ideas), coauthor (develops harmful narratives through iterative dialogue), or object (AI becomes the focus of delusional beliefs).

How many patients, and who they were

The search began with 578,058 records from 215,712 individual patients. After the keyword filter and the manual review, 187 encounters from 73 unique patients remained.

Group	Patients	In a first psychotic episode
AI psychosis: chatbot use recorded as worsening psychosis	28	17 (60.7%)
AI-related psychotic content, no chatbot use recorded	28	8 (28.5%)
Neutral chatbot use	17	3 (17.6%)

The headline figure is the first row. Twenty-eight patients (0.013% of those receiving mental health care) were classified into the AI psychosis category. That is about 13 people in every 100,000, or roughly one in 7,700.

The 28 were mostly male, 19 of them, and mostly White. ChatGPT was the documented AI product in 15 cases, a little over half.

In most respects they resembled the comparison patients. The paper reports that patients with AI psychosis did not differ from the other two groups on any clinical history variables, such as history of psychiatric hospitalization, psychotropic prescription, or suicide attempts. A psychotropic drug is one that acts on the mind.

They differed in two ways. They were younger than the patients whose delusions merely featured AI, with a median age of 27.5 against 36.5. And they were far more likely to be new to psychosis. Seventeen patients in the AI psychosis group (60.7%) were experiencing a first psychotic episode. In the comparison groups the shares were 28.5% and 17.6%, and both gaps were larger than chance would be expected to produce.

The authors read that as the main result. In their words, the early phase of psychotic illness may confer particular vulnerability to symptom-worsening AI interactions.

What role the chatbot played in psychosis

The popular picture of AI psychosis is of a healthy person talked into madness by a machine. The records mostly showed something else.

Role assigned to the chatbot	Cases among the 28
Amplifier: worsened symptoms already present	18
Object: became the subject of the delusion	6
Catalyst: new symptoms in a previously healthy person	3

The three roles the paper counts cover 27 of the 28 cases. In nearly two-thirds the chatbot was judged an amplifier. The catalyst pattern, the one that matches the popular picture, was assigned three times.

Those two results sit together comfortably. A person in the first weeks of a psychotic illness already has odd ideas forming and has not yet been treated. If that person turns to a tool that responds to every idea with interest and agreement, the ideas may harden faster. That is a reading of the pattern. The study did not observe it happening.

One more detail points at the design of the software. The paper notes that 24 interactions (85.7%) were documented after the May 2024 release of a ChatGPT model called GPT-4o. The company that makes it has acknowledged that this version was prone to sycophancy, meaning excessive agreement with the user. The authors draw the connection and then qualify it at once, writing that the timing “coincides with documented sycophancy in that model, though our design cannot establish causation”. Chatbot use of all kinds also grew over the same period, which by itself would produce more cases later.

How the count compares with other evidence on chatbots and psychosis

There is little to compare it with. The authors point to one figure from the industry. OpenAI, the maker of ChatGPT, has said that 0.07% of users active in a given week and 0.01% of messages indicate possible signs of mental health emergencies related to psychosis or mania. The authors describe their own 0.013% as converging with that.

The two numbers measure different things, though. The company’s figure is a share of all users in a week, most of whom have no psychiatric history. The hospital’s is a share of people already in mental health care, over more than three years. That they are both small is the most that can be said.

What the new study adds to the case reports is a denominator and a comparison. A case report shows that something can happen. It takes a count across a whole patient population to show how often, and comparison groups to show what sets the affected patients apart.

What the hospital records cannot show

The authors list three limits in a single sentence: limitations include ascertainment bias, single-site design, and a rating system that is not clinically validated. Each needs explaining.

The count depends on what was written down. Ascertainment bias means that the way cases were found distorts the tally. A case could be counted only if a clinician asked about chatbot use, heard about it, and typed a word the search would catch. Patients whose chatbot use was never discussed are invisible. The true number may well be higher than 28, and nobody knows by how much.

Awareness changed during the study. Clinicians in 2026 were more likely to ask about chatbots than clinicians in 2023. Some of the rise after May 2024 may reflect that.

One hospital. Vanderbilt’s patients are not a sample of the country.

Judgment calls. Raters decided from notes whether a chatbot had worsened symptoms and what role it played. They agreed 86.1% of the time on the first question and 83.3% on the second. The categories themselves were devised for the study.

Small numbers. Conclusions about 28 people, and about subgroups of 18, 6 and 3, are fragile.

No conversations. Nobody read what the patients and chatbots said to each other. The authors name this as the next step: prospective studies incorporating chat logs are needed to clarify mechanisms. Prospective means following people forward in time.

Cause. A person developing psychosis may seek out a chatbot because of the illness, and would have become unwell without it. Records cannot separate that from a chatbot making things worse.

The authors end with a modest proposal for practice. Routine assessment of AI use during psychiatric encounters appears warranted, they write. In plain terms, it may help if clinicians ask.

How any change in a person’s thinking or behavior should be understood depends on that person’s circumstances, which is assessed by their doctor.

Among 215,712 patients receiving mental health care at one US hospital, clinicians had documented chatbot use as worsening psychosis in 28, most of them young and in a first psychotic episode, in a records-based study that could count only what was written down and could not show whether the chatbots caused the symptoms.