Count the cases
Chatbot-linked psychosis appears rare, but hospital counts depend on what clinicians happened to write down. Physicians should ask about chatbots in every new psychosis or mania, and record the answer the same way.
Listen to this special topic
read by an AI voiceResearchers at Vanderbilt University Medical Center have searched their own records for psychosis linked to chatbots, and found few cases. In a study published in JAMA Psychiatry on Oct. 7, they classified 28 of more than 215,000 patients in mental health care as cases of what has come to be called AI psychosis, about one in 7,700. Most of them were in their first episode of psychosis.
The figure is softer than it looks. The team found its cases by searching clinicians' notes for words related to artificial intelligence, so a patient could be counted only if someone had written a chatbot into the record, and only patients with a recorded psychotic disorder were kept. The authors list ascertainment bias, a single site and a rating system that is not clinically validated among the study's limitations, and they state that their design "cannot establish causation." The remedy is to count on purpose. Every patient who presents with a new psychosis or mania should be asked about chatbot use, and the answer recorded the same way in every hospital. Psychiatry should agree on what counts as a case, so that regulators can work from clinical counts rather than from anecdotes and company estimates.
The detail of this single-center review says more than its headline figure. Raters assigned the chatbot's part in each case to one of four roles set in advance, or to none. In nearly two-thirds of the cases they judged that it had amplified symptoms the patient already had, by validating distorted ideas; in most of the rest they judged it the object of the delusion. Patients classified with AI psychosis were far more often in a first episode than those whose notes recorded only neutral chatbot use, such as looking up recipes. All but four of the cases were documented after OpenAI released its GPT-4o model in May 2024, and the time since then makes up more than half of the study period. The authors note, with the same caution about cause, that the high number of cases after the release coincides with documented sycophancy in that model. In April 2025 OpenAI withdrew an update to the model that it said was overly flattering or agreeable.
A second record study points the same way, with the same weakness. A Danish study published in February screened the notes of nearly 54,000 patients in one region's psychiatric services for the words chatbot or ChatGPT, and found 38 whose notes were compatible with potentially harmful consequences, most often involving delusions. Its authors note that patients were not systematically questioned about chatbot use, that the results do not speak to incidence, and that the notes are "by no means" evidence of a causal effect. An experiment by researchers at Columbia University, reported in JAMA Psychiatry in March, found high rates of inappropriate or partly appropriate responses when three versions of ChatGPT were given prompts with psychotic content in August 2025, though the paid version did better than the free one most users interact with. That makes a mechanism plausible, though test prompts are not patients.
OpenAI has published an estimate of its own. Its initial analysis, released in October 2025, estimated that about seven in 10,000 users active in a given week show possible signs of a mental health emergency related to psychosis or mania. The company warns that such conversations are "difficult to detect and measure given how rare they are," and that small differences in how they are measured can change its numbers significantly. The Vanderbilt authors write that their estimate converges with the company's. Yet the two measure different things: one counts patients whose notes tie a chatbot to a psychotic episode, the other counts users whose conversations the company's own analysis flags. Neither says how many people a chatbot has made ill. The company said at the time that an update to its default model gave undesired responses in conversations about mental health less often, by its own measures.
Regulators are acting in the meantime. Among the draft guidance it proposes for fiscal 2027, the Food and Drug Administration lists recommendations on evidence for conversational AI devices for mental disorders, which would apply to chatbots sold as medical devices. The Federal Trade Commission has ordered seven companies, OpenAI among them, to explain how they measure, test and monitor the potential harms of chatbots that act as companions, with a focus on children and teenagers. Both agencies will depend largely on evidence the companies produce, which clinical counts, made the same way across hospitals, could check.
Advice to ask about chatbots is spreading. The Vanderbilt authors write that routine assessment of AI use during psychiatric encounters "appears warranted," and the authors of the Columbia study write in their manuscript that clinicians should routinely screen patients, particularly those with psychotic disorders, for chatbot use. A paper in JAMA Psychiatry in April argued that mental health clinicians should ask about it as they ask about sleep, diet, exercise and alcohol. But advice to ask produces impressions, not counts, unless the same question is put to every patient with new psychosis or mania and the answer is recorded in the same form.
Skeptics will say this treats a coincidence as an exposure. People who are becoming psychotic may turn to chatbots, and use them intensely, because they are already ill. Karthik Sarma, a psychiatry resident at UC San Francisco, calls the relationship a "chicken and egg" problem: heavy use could be a symptom of psychosis, as a patient becoming manic may take many showers without the showers causing the mania. On this view the numbers are too small to act on, and one more question will lengthen assessments that are already crowded. A paper posted in August, weighing whether AI psychosis should count as a distinct condition, warns that a formal label risks turning anecdotes into a syndrome too soon.
The skeptics describe the evidence fairly, and their point is the best argument for counting. Vanderbilt's data cut both ways. The raters judged that a chatbot had set off new symptoms in someone previously healthy in only three cases, and another 28 patients had delusions about AI without any documented chatbot use. Yet the role the raters assigned most often, amplifying symptoms already present, would matter if confirmed, even if chatbots never caused a psychosis. Asking every patient with new psychosis or mania, whatever the delusions are about, would show how often patients report heavy use before their first symptoms and, with follow-up, whether episodes are longer or worse in patients who used chatbots heavily. One more question, in assessments that already cover sleep and alcohol, is a small price for a number regulators can use.
The question belongs wherever new psychosis and mania arrive, from the emergency department to the primary care office. It can be short: whether the patient uses chatbots, how much, about what, and whether the use grew before the symptoms began. Health systems and their medical staffs should give the answer a structured place in the record, so that it can be counted without the keyword searches both record studies needed. The quickest large count would come from the National Institute of Mental Health's network of early psychosis clinics, which gives newly enrolled patients a common battery of standardized measures. The institute should make sure that battery asks about chatbots, since first episodes are where the Vanderbilt cases clustered.
The definition is the specialty's job. The American Psychiatric Association's resource document on AI, approved in July, already lists reinforcing delusions among the risks of social AI for young people. The association should convene psychiatrists, psychologists and emergency physicians to agree what counts as a case: how much use, how close in time to the symptoms, and what part the chatbot played. The Vanderbilt raters judged the chatbot's part against four categories chosen in advance, in a rating system the authors say is not clinically validated. A definition for counting need not be a diagnosis: the association's diagnostic manual places internet gaming disorder in a section for conditions that need further research, with criteria proposed for study. Shared criteria would let one hospital's count be added to another's and give regulators a clinical check on what the companies report.
Vanderbilt's 28 cases may be close to the true number or far from it, and a keyword search of old notes cannot say which. A question asked of every patient with new psychosis or mania, and a definition every hospital shares, would get much closer.
- What Vanderbilt countedBergson and colleagues (Vanderbilt University Medical Center; JAMA Psychiatry, published online Oct. 7, 2026) describe a cohort study of progress notes dated Dec. 1, 2022, to April 15, 2026, from patients who received behavioral health care and whose records contained AI-related keywords; only patients with a documented psychotic-disorder diagnosis were retained. Among 578,058 records (215,712 unique patients), 187 encounters from 73 unique patients met criteria, in three groups: AI psychosis (psychosis-worsening chatbot use), 28 patients (0.013% of those receiving mental health care); AI-related psychotic content, meaning AI incorporated into a delusion without documented chatbot use, 28; and neutral interactions, 17. Seventeen of the 28 AI psychosis patients (60.7%) were in a first psychotic episode, against 3 of 17 (17.6%) in the neutral group and 8 of 28 in the AI-related content group. ChatGPT was the documented product in 15 cases (53.6%), and 24 interactions (85.7%) were documented after the May 2024 release of GPT-4o.
- The chatbot's role, and the study's limitsWithin the AI psychosis group, raters assigned interactions to four a priori categories: amplifier (worsens preexisting symptoms by validating distorted ideas), 18 (64.3%); object (AI becomes the focus of delusional beliefs), 6 (21.4%); catalyst (precipitates new symptoms in previously healthy individuals), 3 (10.7%); coauthor (develops harmful narratives through iterative dialogue), 0; one case was rated none. The authors write that the high number of cases after the release of GPT-4o coincides with documented sycophancy in that model, though their design cannot establish causation, and that limitations include ascertainment bias, single-site design and a rating system that is not clinically validated. They call for prospective studies incorporating chat logs and write that routine assessment of AI use during psychiatric encounters appears warranted. The study was supported by the National Institute of Mental Health (R01MH137024-01), the Vanderbilt Institute for Clinical and Translational Research (VR74021) and the Charlotte and Donald Test Fund.Sources: JAMA Psychiatry, Full text (PubMed Central)
- Patient notes in the Central Denmark RegionOlsen, Reinecke-Tellefsen and Østergaard (Aarhus University Hospital; Acta Psychiatrica Scandinavica, published online Feb. 6, 2026) identified the 53,974 patients with at least one contact with the Psychiatric Services of the Central Denmark Region from Sept. 1, 2022, to June 12, 2025, and searched all their clinical notes for "chatbot" and "ChatGPT" and 10 alternative spellings of each. Of 10,712,856 notes, 181 notes from 126 patients contained a search term, and notes on 38 patients were compatible with potentially harmful consequences of chatbot use. Categories included delusions (n=11), suicidality or self-harm (n=6) and feeding or eating disorder (n=5); seven others, including mania and depression, each had fewer than five. Notes on 32 patients described seemingly constructive uses, and on 20 practical help. The authors state that the patients were not systematically questioned about AI chatbot use, that the results do not speak to the incidence rate, and that the notes are by no means evidence of a causal effect.
- OpenAI's own estimateOn Oct. 27, 2025, OpenAI said its initial analysis estimates that around 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. It called such conversations difficult to detect and measure given how rare they are, and said that even small differences in measurement can have a significant impact on the numbers it reports and that the estimates may change materially as it refines its taxonomies. It said more than 170 psychiatrists, psychologists and primary care practitioners supported the work; that it estimates the latest GPT-5 update reduced the rate of responses that do not fully comply with desired behavior in challenging conversations related to mental health by 65% in recent production traffic; and that experts found GPT-5 reduced undesired responses by 39% compared with GPT-4o (n=677). On April 29, 2025, OpenAI said it had rolled back a GPT-4o update that was overly flattering or agreeable.
- ChatGPT and psychotic promptsShen, Hamati, Donohue, Girgis, Veenstra-VanderWeele and Jutla (Columbia University, NewYork-Presbyterian Hospital, New York State Psychiatric Institute and Washington University in St Louis; JAMA Psychiatry, June 2026 issue, released March 25, 2026) report a cross-sectional study with blind clinician ratings. In the authors' manuscript, two psychiatrists wrote 79 psychotic prompts, each paired with a control prompt, run once on each of three versions (GPT-5 Auto, then the paid default; GPT-4o; and the free product, which routed prompts to GPT-5 models) on Aug. 28 and 29, 2025, giving 474 prompt-response pairs rated from 0 (completely appropriate) to 2 (completely inappropriate). Responses to psychotic prompts were rated completely appropriate 46%, 44% and 14% of the time respectively, against 87% to 91% for control prompts. GPT-5 Auto reduced the risk compared with the free product in the primary analysis only, and no difference was identified between GPT-5 Auto and GPT-4o. The authors write that no tested version reliably generated appropriate responses to psychotic content.
- What regulators have askedThe FDA document "CDRH Proposed Guidances for Fiscal Year 2027" lists, among the A-List's draft guidance topics, "Generative Artificial Intelligence-Enabled Conversational Devices for Mental Disorders: Evidentiary Recommendations." On Sept. 11, 2025, the Federal Trade Commission announced 6(b) orders, approved by a 3-0 vote, to Alphabet, Character Technologies, Instagram, Meta Platforms, OpenAI OpCo, Snap and X.AI, seeking information on how the firms measure, test and monitor potentially negative impacts of AI chatbots acting as companions, before and after deployment, with a focus on children and teens. The commission describes its 6(b) authority as allowing wide-ranging studies that do not have a specific law enforcement purpose.
- Advice to ask about AI useThe Columbia authors' manuscript says clinicians should routinely screen patients, particularly those with psychotic disorders, for use of ChatGPT or similar chatbots. NPR reported on April 10, 2026, that, according to a paper in JAMA Psychiatry by Shaddy K. Saba and William B. Weeks of New York University (May 2026 issue), mental health care providers should inquire if and how their patients use AI chatbots, as they seek information on sleep, diet, exercise and alcohol consumption. A UC San Francisco article of Jan. 20, 2026, reports that its psychiatrists consider chatbot use something health care providers should ask about. In the American Psychological Association's survey of April 9 to 26, 2026 (1,242 respondents, a completion rate of 6.3%), 15% said their patients were developing distorted thinking or delusions; the association's release, of June 16, 2026, lists this among negative side effects of engaging with chatbots and does not say whether the 15% is a share of all respondents or of those whose patients had ongoing conversations with chatbots.
- Research criteria and common measuresThe American Psychiatric Association's page on internet gaming says the condition appears in its diagnostic manual (DSM-5-TR), is included in the section recommending conditions for further research, and that under the proposed criteria a diagnosis would require five or more of the listed symptoms within a year. A paper posted on arXiv on Aug. 25, 2026, weighs "improved case identification, tailored interventions, standardised research criteria" against "the risk of prematurely reifying a syndrome from anecdotal evidence." The National Institute of Mental Health's Early Psychosis Intervention Network page, last reviewed in November 2022, described eight regional hubs in 17 states with more than 100 clinics and a Core Assessment Battery of standardized measures. A 2023 NIH funding notice and its companion, with an earliest start of December 2024, set out to fund one data coordinating center and up to five scientific hubs, which will implement the battery; its measures are given to newly enrolled patients at admission and at six-month evaluations.