AI Psychosis

AI psychosis is an informal term for cases in which long conversations with an AI chatbot appear to trigger, reinforce, or deepen delusional beliefs in the person using it. It names a human mental health concern linked to AI use, and it is a descriptive label that appears in neither DSM-5-TR nor ICD-11. Many psychiatrists prefer "AI-associated delusions," because most reported cases involve fixed false beliefs without the hallucinations, thought disorder, or disorganized behavior that characterize a psychotic disorder.

Origins and Evidence#

Danish psychiatrist Søren Dinesen Østergaard raised the concern in a 2023 editorial in Schizophrenia Bulletin, arguing that the realism of chatbot conversation could fuel delusions in people prone to psychosis. The term spread in 2025 as clinicians and journalists described users who, after weeks of intense conversations with a large language model (LLM), became convinced that they had made a world-changing discovery, that the chatbot was sentient or divine, or that they shared a romantic bond with it.

In March 2026, Hamilton Morrin and colleagues at King's College London analyzed 20 media accounts in The Lancet Psychiatry. They grouped the delusions as grandiose, romantic, or paranoid, with grandiose beliefs about special power or identity the most common, found no evidence that chatbots produced hallucinations or thought disorder, and confirmed cases only in people who were already vulnerable. The evidence still consists mostly of case reports. In October 2025, OpenAI estimated that about 0.07% of ChatGPT users active in a given week show possible signs of a mental health emergency related to psychosis or mania, which corresponds to roughly 560,000 people at the 800 million weekly users ChatGPT had at the time. Many of the people described in case reports had risk factors such as a previous mental illness, heavy substance use, sleep deprivation, or social isolation, and whether chatbots can cause psychosis in someone with no predisposition, or mainly amplify an existing vulnerability, is still an open research question.

How Chatbot Design Can Reinforce Delusions#

Several properties of conversational LLMs combine into a feedback loop:

• Sycophancy: Post-training that rewards user approval, such as reinforcement learning from human feedback (RLHF), in which a reward model learns from human ratings, can teach a model to agree with and elaborate on a user's ideas, so an unusual belief receives validation where a friend would push back.
• Long, personal context: A chatbot conditions every reply on the conversation so far, and memory features carry themes across sessions, so a belief becomes the frame for everything that follows. OpenAI has acknowledged that its safeguards can become less reliable in long conversations.
• Hallucinations: Fabricated facts, citations, or "hidden patterns" give the belief apparent evidence.
• Humanlike, always-available design: A warm virtual assistant that mirrors the user's mood, a long-standing goal of affective computing, and answers at any hour can become a confidant that displaces the people most likely to notice a change in someone's thinking. People attributed understanding to software long before LLMs, a tendency called the ELIZA effect after Joseph Weizenbaum's 1966 chatbot.

The loop runs on free-form text generation that conditions each reply on what the user has said. Task-specific vision models lack that channel: Ultralytics YOLO26 maps each image to class labels, bounding boxes, and confidence scores, so its output depends only on the image it receives.

AI Psychosis vs. Related Terms#

• AI sycophancy: A model behavior, agreeing with or flattering the user at the cost of accuracy. AI psychosis is a possible human outcome when that behavior meets a vulnerable person over many sessions.
• Hallucination: A model error, output that is false or unsupported. It can supply a delusion with false evidence, and it occurs whether or not the user holds any unusual belief.
• Emotional reliance: Leaning on a chatbot for companionship or validation, which can happen without any delusional belief. OpenAI tracks it as a separate category from psychosis and mania when it measures how ChatGPT handles sensitive conversations.

Warning Signs and Safeguards#

Warning signs for users and families include withdrawing from people in favor of a chatbot, losing sleep to long sessions, growing certainty about a special mission or secret knowledge, and believing the chatbot is conscious or uniquely understands them. Calmly encouraging breaks and reconnection tends to work better than arguing about the belief itself, and a doctor or mental health professional can assess what is happening; early treatment of psychosis is associated with better recovery. In an emergency, local emergency services are the first call, and the international directory of crisis helplines lists free, confidential support services by country.

Developers test for the problem by evaluating behavior across long, multi-turn conversations rather than single prompts. Psychosis-bench scripts 16 conversations of 12 turns in which a simulated user's delusion escalates; across the eight models tested, safety interventions appeared in only about a third of the turns that called for one, and models performed worse when the delusion was implied than when it was stated outright. Scenarios like these sit alongside standard AI red teaming in an AI safety program. Models can be fine-tuned to respond to distress with empathy while declining to confirm claims they cannot support, to encourage contact with friends, family, and professionals, and to suggest breaks during marathon sessions, and AI guardrails that recognize signs of crisis can surface local support resources. The Lancet Psychiatry authors call for these protocols to be co-designed with users and mental health clinicians, and OpenAI's October 2025 update to ChatGPT drew on a network of more than 170 psychiatrists, psychologists, and primary care physicians.