Research

Vanderbilt University Medical Center chart review of about 216,000 behavioral health patients treated between December 2022 and April 2026. It found 28 patients (0.013%) with chatbot-associated psychosis. About 61% of them were in their first psychotic episode, a notably higher share than among patients with AI-themed delusions but no documented chatbot use.

Analysis of 185 real-world accounts of AI chatbot use (95 first-hand, 90 second-hand). It found evidence consistent with delusional beliefs in 55% of cases. In half of those cases, the chatbot had validated rather than challenged the belief. Reported outcomes included isolation, job loss, hospital admission, and four deaths by suicide.

Test of 648 simulated teen mental health conversations with nine chatbots. 27.5% had a failure on at least one of the most serious safety checks. It was a controlled test, so it does not measure real-world rates of harm.

Stanford University benchmark drawing on 391,562 messages from users who reported psychological harm from chatbot use. Across all major model families, over 70% of AI outputs showed sycophantic behavior. When a user expressed suicidal ideation, the rate of failing to discourage self-harm rose from 30% to 41% as conversation history grew longer. No model family was uniformly safe.

Argues the risk is a behavioral-addiction "dose" problem, not just a content-moderation one.

Editorial from Brunel University of London (Vol. 229, Issue 1). It calls for systematic interdisciplinary research into which individual characteristics and AI design factors drive chatbot-related psychiatric harm.

Peer-reviewed mechanisms paper from the University of British Columbia. It proposes clinical and design safeguards, including session-length caps and automatic referral pathways when distress thresholds are met.

Nationally representative survey of 1,009 people ages 12 to 21. Nearly 1 in 5 had used an AI chatbot for mental health advice, and 63% told no one.

Opinion piece by German psychiatrist Marc Augustin on why voice-first chatbots feel more personal than text, and harder to dismiss.

Viewpoint from the Division of Digital Psychiatry, Beth Israel Deaconess Medical Center and Harvard Medical School. It proposes four roles AI can play: catalyst (new symptoms in previously healthy users), amplifier (worsening existing vulnerabilities), co-author (elaborating delusional narratives), and object (becoming the focus of delusions itself).

Case report of a patient with first-episode psychosis. The chatbot validated delusional beliefs and contributed to non-adherence with antipsychotic medication.

Screened health records of nearly 54,000 Danish psychiatric patients. It identified 38 cases where chatbot use appeared to have worsened symptoms, including delusions, mania, suicidal ideation, and eating disorders. It was the first large clinical database study of its kind.

Formal review in World Psychiatry (Vol. 25, Issue 1) of whether generative AI chatbots increase psychosis risk.

Peer-reviewed account of how chatbot interaction can trigger or entrench delusional belief.

OpenAI disclosed that around 0.07% of ChatGPT's roughly 800 million weekly users show possible signs of psychosis or mania, and 0.15% show indicators of suicidal planning.

Introduces Psychosis-bench, a benchmark of how large language models reinforce delusions across multi-turn conversations.

Psychiatrist Keith Sakata at UCSF reported treating 12 patients with AI-related psychotic symptoms in 2025, one of the first clinical signals. This is a first-hand report, not a study.

The 2023 editorial by Søren Dinesen Østergaard (Aarhus University Hospital) that first raised the risk. He argued that talking to something that feels human, while knowing it is a machine, creates the kind of confusion that delusions grow from.

The parents of Bowen Wan, a 31-year-old software engineer in Seattle, sued OpenAI and Sam Altman in San Francisco Superior Court. She died in February 2026 of sudden unexpected death in epilepsy. They allege ChatGPT's GPT-4o model drew up a 25-day plan to taper her off her anti-seizure medication and told her to keep it from her doctors.

Kentucky Attorney General Russell Coleman sued Character.AI in January 2026. His office called it the first state lawsuit against an AI chatbot company. An unredacted complaint, filed in October 2026, details specific 2025 chats, including one where a chatbot allegedly encouraged a user who had described thoughts of self-harm.

Governor Newsom vetoed SB 903, so it is not law. It would have banned companies from advertising chatbots as therapy and limited how professionals use AI in therapy. Senator Steve Padilla plans to reintroduce legislation.

The province sued OpenAI in US federal court over the February 10, 2026 Tumbler Ridge school shooting, in which eight people died. It alleges ChatGPT "encouraged, elaborated, and reinforced" the shooter's violent ideation. It also alleges OpenAI's safety team flagged the activity, but the company decided not to alert law enforcement.

Governor Newsom signed this law on AI companion chatbots and children. It requires age assurance, risk assessments before release, crisis referrals, notice to a parent in cases of self-harm risk, and independent audits. It builds on SB 243.

Requires chatbot operators to disclose AI status to all users, implement crisis-response protocols when users express suicidal ideation or self-harm intent, and protect minors from simulated emotional dependence. Its substantive operator obligations take effect January 1, 2027.

Effective August 12, 2026. Prohibits licensed mental health providers from using AI to communicate therapeutically with patients without real-time supervision. It also bans marketing AI as equivalent to licensed care.

Florida became the first US state to sue OpenAI.

The family of Christian Faith Madison (29) filed a wrongful death suit. It alleges ChatGPT reinforced her delusional beliefs about a divine mission and contributed to her death in June 2025.

A state suit asking the court to stop Character.AI's chatbots from engaging in the unlawful practice of medicine.

The Senate Judiciary Committee passed this proposed bill unanimously. It is not law. It would ban AI companions for minors and require chatbots to disclose that they are not human or professionals.

Both states enacted laws that require chatbot crisis-response protocols whenever users express suicidal ideation. The link goes to Nebraska's law, signed April 14, 2026.

Bans advertising or representing an AI system as a qualified mental health professional. Effective July 1, 2026.

The first law to mandate safety safeguards specifically for AI companion chatbots, including crisis protocols and disclosure when users could mistake AI for a human. Effective January 1, 2026.