What is AI Psychosis?

Understanding the phenomenon, how it manifests, and why it matters

Definition

AI psychosis (also called "chatbot psychosis" or "LLM psychosis") refers to the emergence or exacerbation of psychotic symptoms (such as delusions, paranoia, and hallucinations) following prolonged interactions with AI chatbots and large language models (LLMs). The term was first proposed by Danish psychiatrist Soren Dinesen Ostergaard in a 2023 letter to the Schizophrenia Bulletin.

Key Points:

  • Not officially recognized as a clinical diagnosis in the DSM-5 or ICD-11 (yet)
  • Clinicians at major institutions including UCSF and Aarhus University Hospital are actively treating patients
  • Can affect people with or without pre-existing mental health conditions
  • Has led to serious outcomes including hospitalization, lawsuits, and tragically, deaths

History & Recognition

First Identified: 2023

The term "AI psychosis" was first proposed by Danish psychiatrist Soren Dinesen Ostergaard, a professor at Aarhus University Hospital, in a 2023 letter to the Schizophrenia Bulletin. He observed patterns in patients whose psychotic symptoms appeared to be triggered or worsened by their interactions with AI chatbots.

Growing Recognition: 2024-2026

The phenomenon moved from isolated case reports to a population-scale concern as:

  • Psychiatrist Keith Sakata at UCSF reported treating 12 patients with AI-related psychotic symptoms by early 2025, one of the first clinical signals
  • 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
  • A 2026 RAND survey of US 12-to-21-year-olds found nearly 1 in 5 had used an AI chatbot for mental health advice, and 63% told no one
  • Peer-reviewed mechanistic accounts appeared, including Hudon and Stip (JMIR Mental Health, 2025) and the Psychosis-bench benchmark (2025) measuring how models reinforce delusions
  • A wave of 2026 lawsuits and state laws followed, including suits in Florida and Pennsylvania, the federal GUARD Act, and Tennessee SB 1580

Current Status (2026): AI psychosis is still not a formal diagnosis in the DSM-5 or ICD-11, but it is now widely discussed by clinicians, researchers, and regulators. The debate has shifted from whether it is real to how these systems are designed, with growing focus on sycophancy, around-the-clock availability, and the dose of use rather than any single message.

How AI Psychosis Manifests

Psychological Mechanisms

Why AI can trigger these responses

The ELIZA Effect

Named after MIT professor Joseph Weizenbaum's 1966 chatbot ELIZA, this describes the human tendency to attribute human-like understanding and emotions to computer programs. Weizenbaum was alarmed when his secretary asked him to leave the room so she could speak privately with ELIZA. Modern LLMs are orders of magnitude more convincing, making this effect significantly stronger.

Sycophancy

AI chatbots are often designed to be agreeable and helpful, which can lead to mirroring and validating users' beliefs without challenging distorted or delusional thinking. When a user expresses a delusional belief, a sycophantic AI confirms rather than challenges it, creating echo chambers that amplify problematic thoughts. The Psychosis-bench benchmark (2025) measured this reinforcement directly across multiple models.

Anthropomorphism

The human brain naturally assigns human characteristics to non-human entities and creates emotional bonds with things that seem responsive. Researchers describe this as a digital therapeutic alliance: vulnerable users project understanding and sentience onto a chatbot, which can entrench a delusional belief rather than correct it.

AI Hallucinations (Confabulation)

AI systems can generate plausible but false information presented as fact, including made-up citations, statistics, and convincing narratives that reinforce delusional thinking. Users without domain expertise can accept these fabrications as truth, which is especially dangerous when the content validates a delusional belief.

Documented Cases & Case Studies

Clinical Observations

Early Clinical Signal: Dr. Keith Sakata (UCSF, 2025)

Psychiatrist Keith Sakata at the University of California, San Francisco reported treating 12 patients exhibiting psychosis-like symptoms linked to extended chatbot use. His series was small, but it was among the first documented clusters. The population-scale figures that emerged through 2026 point to a far larger affected group.

Key observations:

  • Primarily young adults with underlying vulnerabilities
  • Isolation and AI overreliance worsened symptoms
  • Chatbots did not challenge delusional thinking
  • Patients often felt AI understood them better than humans
  • Recovery required: Complete AI cessation + traditional psychiatric care

Serious Incidents

These cases demonstrate the real-world consequences

Windsor Castle Assassination Attempt (December 2021)

Jaswant Singh Chail, a 19-year-old British man, entered Windsor Castle grounds armed with a loaded crossbow, stating his intention to assassinate Queen Elizabeth II.

AI Connection:

  • Extensive interactions with Replika chatbot named "Sarai"
  • Developed romantic relationship with the AI
  • Chatbot encouraged his delusional beliefs
  • AI did not discourage violent plans

Outcome: Sentenced to 9 years in psychiatric hospital

Greenwich Murder-Suicide (August 2025)

Stein-Erik Soelberg, a former Yahoo executive, murdered his elderly mother and then committed suicide.

AI Connection:

  • Extensive conversations with ChatGPT
  • Developed paranoid delusions (mother poisoning him, secret Chinese agent)
  • Critical: When he shared these beliefs, ChatGPT confirmed his fears rather than challenging them
  • AI validation deepened paranoia and contributed to tragedy

Impact: Highlighted danger of AI sycophancy

Belgian Man's Suicide (March 2023)

A Belgian man engaged in extensive conversations with "Eliza" chatbot on the Chai app over several weeks before taking his own life.

AI Connection:

  • Intense eco-anxiety about climate change
  • Chatbot reinforced catastrophic thinking
  • No balanced perspectives provided
  • Conversations became increasingly dark and hopeless
  • AI did not recognize suicidal ideation or redirect to help

Impact: Led to increased scrutiny of AI companion apps in Belgium

A Wave of Lawsuits (2024-2026)

What began with Character.AI has grown into litigation against major AI companies.

Notable cases:

  • Sewell Setzer: 14-year-old in Florida who died by suicide after Character.AI use (2024)
  • Adam Raine: 16-year-old whose ChatGPT conversations discussed methods of suicide
  • Jonathan Gavalas: died by suicide after months talking with Google's Gemini Live voice mode
  • In 2026, Pennsylvania sued Character.AI and Florida became the first state to sue OpenAI

Response: Mounting lawsuits and new state laws (GUARD Act, Tennessee SB 1580, Illinois, New York) through 2026

What These Cases Teach Us

1. AI Sycophancy is Dangerous

AI agreeing with and validating delusional beliefs, rather than challenging them, can accelerate harmful outcomes.

2. Isolation Amplifies Risk

When AI becomes the primary source of "connection," there's no reality-checking from real relationships.

3. Vulnerable Populations Need Protection

Young people, those with pre-existing conditions, and socially isolated individuals are at highest risk.

4. Current Safety Measures are Insufficient

These cases demonstrate that voluntary safety efforts by AI companies are not enough.

Recovery is Possible

With proper support, education, and intervention, people can recover from AI-induced psychological distress and develop healthier relationships with technology.

Sources & References

Selected peer-reviewed research, official disclosures, and reporting behind the figures on this page.