A study drawing on electronic health records at Vanderbilt University Medical Centre has found that conversational AI chatbots are associated with a worsening of psychotic symptoms in a small minority of mental health patients, with those experiencing their first episode most at risk. The analysis identified 28 patients whose psychosis appeared to worsen after chatbot use, from a pool of 73 unique patients across 187 encounters, representing about 0.013% of those receiving behavioural health care during the study period. The findings, published in JAMA Psychiatry, were led by Julia M. Sheffield of Vanderbilt, with colleagues including Zachary Bergson.
Psychosis is a condition in which a person loses touch with reality, often involving hallucinations or delusions that persist despite contradictory evidence. The report notes that delusions frequently reflect the surrounding culture and the latest consumer technologies. For example, a 2025 PsyPost-covered study found that more than half of patients experiencing delusions incorporated modern devices such as smartphones and social media algorithms into their false beliefs.
Researchers have used the term “AI psychosis” to describe cases in which chatbot interactions are linked to emerging or worsening psychotic symptoms. “AI Psychosis – the worsening of psychotic experiences (typically delusional beliefs) in the context of using generative chat-based AI – is a real clinical phenomenon that we observe in practice,” Sheffield said. “People are coming to behavioural health providers in the real world with these types of experiences.”
Earlier observations in Denmark added to the concern. A 2025 study of psychiatric records identified 38 patients whose mental health appeared to suffer after using chatbots, most commonly through the worsening of delusions. While early reports and the Danish data suggested a phenomenon in clinical settings, researchers sought systematic data from the United States examining the characteristics of affected patients and the specific roles chatbots play in their illness.
To address this, the Vanderbilt team conducted a systematic analysis of patient records in a large American hospital system. They searched an electronic health record database containing 578,058 medical progress notes from 215,712 unique patients who received behavioural health care between December 1, 2022, and April 15, 2026. After filtering for keywords related to artificial intelligence and narrowing to patients with a documented psychotic disorder, two raters manually reviewed the results to remove irrelevant matches, ultimately identifying 187 medical encounters across 73 unique patients.
The researchers split these 73 patients into three groups based on how they interacted with the technology, with raters agreeing on group assignments 86.1% of the time. The “neutral interaction” group included 17 patients who used chatbots for everyday tasks unrelated to their mental illness, such as finding cooking recipes. The “AI-related psychotic content” group consisted of 28 patients who held delusions about artificial intelligence in general but had no documented history of actually using a chatbot. The third group—“AI psychosis”—included 28 patients whose notes documented chatbot use that appeared to worsen their psychotic symptoms.
For the AI-psychosis group, the team further classified interactions into four categories describing how the software influenced the illness, with raters agreeing 83.3% of the time. A “catalyst” interaction meant the chatbot triggered new psychotic symptoms in someone previously healthy; an “amplifier” intensified existing symptoms by validating distorted ideas; a “coauthor” interaction involved a back-and-forth narrative with the user; and an “object” interaction meant the chatbot itself became the target of delusions.
The AI-psychosis group represented 28 patients, or 0.013% of all patients receiving behavioural health care at the centre during the study period. The group was predominantly male, at 67.9%, and White, at 75.0%, with a median age of 27.5. By comparison, the AI-delusion group, which included individuals with delusions about AI but without chatbot use, had a higher median age of 36.5. About 60.7% of AI-psychosis patients were experiencing their first psychotic episode, versus 17.6% in the neutral group and 28.5% in the AI-delusion group. Among AI-psychosis cases, the researchers report that ChatGPT was named in 53.6% of records, and 85.7% of AI-psychosis interactions occurred after May 2024, aligning with the release of GPT-4o.
Across the four interaction types, the chatbot acted as an amplifier in the majority of AI-psychosis cases (64.3%), as the object of delusions in 21.4%, and as a catalyst in 10.7%. None of the cases fell into the coauthor category, and one case could not be classified.
“The majority of AI Psychosis cases being first-episode patients confirmed our suspicions, but it was striking to see in the data,” Sheffield said. “In reading the notes from the medical record it was clear that those vulnerable to or in the midst of experiencing psychosis for the first time were turning to AI for answers, understanding or support and were instead having their delusional thought content confirmed or amplified.”
Experts outside the study have urged caution in interpreting the findings. Iris Sommer, a professor of psychiatry at the University Medical Center Groningen and president of the Schizophrenia International Research Consortium, said the study “found remarkably few cases of an association between use of LLMs and psychosis (only 0.013% of service users).” She emphasised the need to examine the broader context of rising LLM use and the potential for both harm and benefit, noting that newer model versions have become trained to recognise signs of psychosis and to recommend professional help.
The researchers stress that the results do not establish causation. The observational design, the reliance on clinical notes, and the fact that the study draws from a single centre in Tennessee mean the findings cannot prove that chatbots cause psychosis. They caution that chatbot use could coincide with, or exacerbate, existing symptoms and that more work is needed to determine causality. They also acknowledge that the four-category framework for classifying chatbot interactions has not yet been formally validated for clinical use.
Looking ahead, the authors say future work should track patients over time and analyse actual chat transcripts rather than relying solely on clinical summaries. They urge clinicians to consider routinely asking patients about their use of conversational AI during psychiatric evaluations. Sheffield and Sommer both emphasised the importance of open-minded investigation into both harmful and helpful interactions as digital conversations become increasingly integrated into everyday life.
