Characterizing Artificial Intelligence Psychosis in a Large Medical Setting

Key Findings

• — Among 578,058 records from 215,712 unique patients, 187 encounters from 73 patients met criteria; 28 patients (0.013% of those receiving mental health care) were classified as AI psychosis, 17 as neutral and 28 as AI-related psychotic content; rater agreement 86.1%.
• — Seventeen AI-psychosis patients (60.7%) were experiencing a first psychotic episode, significantly more than the neutral (3, 17.6%; P = .006) and AI-related content (8, 28.5%; P = .03) groups; the AI-psychosis group was younger than the AI-related content group (median 27.5 against 36.5 years; P = .02).
• — The AI-psychosis cohort was primarily male (19, 67.9%) and White (21, 75.0%); unspecified psychosis was the most common diagnosis (8, 28.6%); 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.
• — Amplifier was the most common interaction type (18, 64.3%), followed by object (6, 21.4%) and catalyst (3, 10.7%); the groups did not differ on psychiatric hospitalisation history, psychotropic prescription or suicide attempts.
• — The authors name ascertainment bias, the single site and a rating system that is not clinically validated as limitations.

Methodology Notes

Retrospective single-site cohort review (STROBE) using Epic Clarity keyword extraction, algorithmic then manual screening by two raters with a third as tiebreaker; Kruskal-Wallis, chi-square and Fisher exact tests. Research Letter format. The prevalence figure is conditional on a documented psychosis diagnosis in the record. Published online 2026-10-07 (PubMed PMID 42842276). Verification route: jamanetwork.com blocks crawlers, so the full text was read through a rendering fetch tool that returned the Methods and Results paragraphs verbatim; title, authors, venue and date confirmed on PubMed and Crossref. The June 2026 medRxiv preprint (10.64898/2026.06.04.26354939) carried a longer title; the journal version adds the record and patient denominators, the 0.013% prevalence, demographics, diagnosis mix and the interaction-type percentages. Disclosures: two authors report NIH grants outside the work.

Sources