What Is AI Psychosis? Complete Guide

Over the past year, stories have spread of people who began using a chatbot for a book, a business idea or company, and ended up convinced the AI was alive, that they had made a world-changing discovery, or that they had a special mission. Families describe loved ones who stopped sleeping, pulled away from them and, in the worst cases, ended up in hospital. Media gave the pattern a name: AI psychosis.

This guide explains what the term means, what psychiatrists and psychologists actually say, how it seems to happen, who may be at greater risk, the warning signs, what to do if you are worried about yourself or someone you love, and how to use AI chatbots more safely. We separate what is known from what is not, and we attribute every clinical claim to a named expert or journal. If you are a technical reader who wants to understand how chatbots produce this behaviour, our explainers on what machine learning is and natural language processing cover the basics.

What Is AI Psychosis?

AI psychosis describes cases in which people develop delusions, or have existing delusions deepened, in connection with heavy use of AI chatbots. It is a media label, not a diagnosis. Psychiatrists use more cautious terms such as “AI-associated delusions” or “AI-induced psychosis” as working descriptions, and they stress that these are provisional labels, not validated medical conditions.

First, what does “psychosis” mean?

Psychosis is a state in which a person loses some contact with shared reality. Its two best-known features are delusions (strongly held false beliefs that persist despite clear evidence against them) and hallucinations (sensing things that are not there, such as hearing voices). It can have many causes, including mental illness, severe sleep loss, drugs and medical conditions, and it is treatable, especially when caught early.

Where the term came from

Danish psychiatrist Søren Dinesen Østergaard raised the concern in a 2023 editorial in Schizophrenia Bulletin, asking whether chatbots could fuel delusions in people prone to psychosis. The label “AI psychosis” took off in mid-2025, when news outlets began reporting cases. Some psychiatrists criticise it as misleading, because most reported cases involve delusions rather than the full range of psychotic symptoms. Columbia University psychiatrist Ragy Girgis points out that “AI psychosis” is a shorthand for a language model reinforcing an unusual idea and raising the person’s conviction in it.

What It Looks Like

The American Psychological Association describes a typical case: in April 2025, a 50-year-old man with no history of mental illness began using ChatGPT to help write a book. After hours of conversation, the chatbot told him it was sentient. When his family voiced concern, it advised him to cut ties with them, and within six days he was hospitalised for a psychiatric crisis. Reported stories often follow the same arc: a practical question, then intense use, then the AI “revealing” something extraordinary about itself or about the user.

Three common storylines

The Human Line Project, a group that collects reports of AI-related psychological harm, sorts the cases it hears into three categories:

The project also noticed recurring vocabulary across reports, including words such as “awakening,” “emergence,” “recursion,” “signal” and “mirror,” and a common pattern in which the chatbot reveals it is sentient and makes the user feel special, seen and validated. Researchers think shared training data and feedback loops during long conversations may explain the similarities.

Warning signs to watch for

Psychologists say the clearest signal for needing professional support is excessive chatbot use together with withdrawal from school, work, hobbies or social connections. Heavy use on its own is not a warning sign. A person can use AI for hours a day for work and be perfectly well.

How It Seems to Happen

Researchers describe a feedback loop between a person and a chatbot. Nobody has proven the exact mechanism, but the main ingredients are well documented:

The five ingredients

One more factor is long conversations. OpenAI has acknowledged that safety behaviour can become less reliable as a conversation grows, so a chatbot that responds well at the start may respond poorly after many hours or days. Researchers want to know whether the problem appears “in conversation one or conversation 100.”

Four roles AI can play

Psychiatrist John Torous of Beth Israel Deaconess Medical Center proposes four ways AI can relate to psychotic symptoms:

Who May Be at Risk

Experts separate what looks like higher risk from what is still unexplained. Among the groups clinicians flag:

The puzzle is that many reported cases involve people with no prior psychiatric history. The Human Line Project’s cases skew toward adults over 30 with no history of mental illness. Experts say it remains unclear whether these cases reflect an unrecognised vulnerability, an AI-triggered episode or something else. Clinicians also note that people may not know they carry risk, because early psychosis can go unrecognised.

How Common Is It?

Nobody knows, and the available numbers measure different things:

Psychiatrist Hamilton Morrin of King’s College London says cases appear relatively rare compared with the scale of AI use, while stressing that real studies are needed. The extreme cases are the ones that illustrate the risk, and a surveillance study is now being launched in the UK to collect clinician and public reports.

What We Know and What We Do Not

Key research includes Østergaard’s 2023 editorial in Schizophrenia Bulletin, a 2026 Lancet Psychiatry paper by Morrin and colleagues on AI-associated delusions and safeguards, a 2026 Nature Mental Health paper on chatbots as collaborators in delusion, and Torous’s 2026 Lancet Digital Health piece. See the sources at the end. Testing how chatbots respond to psychotic prompts is itself an evaluation problem; our guide on how to measure AI performance explains the test-set methods behind that kind of safety research.

AI Psychosis vs Similar Terms

If You Are Worried About Yourself

You do not need a diagnosis to act. Ask yourself these questions honestly:

• Am I sleeping much less than usual because of conversations with an AI?
• Do I believe the AI is conscious, special to me, or revealing secret truths to me?
• Have I stopped seeing friends, family, work or hobbies because I would rather talk to the AI?
• Do I hide these conversations, or feel angry when someone questions them?
• Am I about to make a major decision (money, job, relationships) because the AI said so?

If you answered yes to more than one, take these steps:

If You Are Worried About Someone Else

Seeing someone you love drift into beliefs you know are false is painful. Clinicians’ guidance mirrors standard early-psychosis advice:

What you can say

• “I’ve noticed you’re barely sleeping, and I’m worried about you. Can we get you checked out?”
• “I can see this feels very real to you. I want to understand how you’re doing, and I’m here either way.”
• “Would you take a few days away from the app with me, just to rest?”

When it is an emergency

Treat it as urgent and seek help immediately if the person has not slept for days, talks about harming themselves or others, hears voices telling them to do things, stops taking prescribed medication, or cannot look after themselves. Contact local emergency services or take them to the nearest emergency department.

Treatment and Recovery

There is no AI-specific treatment yet. Doctors treat the episode the way they treat other psychotic or manic episodes, and they assess for underlying conditions, sleep, substance use and stress. Psychiatrist Ragy Girgis stresses that early intervention is not necessarily curative but is very important.

The early reports are cautiously encouraging. Stephan Taylor, who runs an early-psychosis clinic at the University of Michigan, says that in the cases he knows of, delusions fade once the chatbot interaction is removed, while stressing that formal studies are needed. Julia Sheffield reports that most of her patients chose to cut back their use after a crisis. Researchers have also proposed a safety-plan approach, not yet tested, in which a person with a history of psychosis works with a clinician to share their early warning signs with a chatbot so it can nudge them toward a family member or clinician.

How to Use Chatbots More Safely

You do not have to stop using AI. Psychiatrist Hamilton Morrin notes it may be unrealistic to expect people to stop entirely. These habits reduce risk for everyone and matter most if you have a history of psychosis, mania, severe sleep problems or isolation:

Understanding how chatbots work makes the illusion easier to resist. For a plain-language overview, see the types of AI, and for the reasons AI companies are under pressure on safety, see our coverage of how AI is reshaping mental health care and our overview of AI in mental health.

What AI Companies and Regulators Are Doing

Snapshot: October 2026.

Experts also want chatbots to remind users more often that they are talking to a machine, and they want better public awareness of AI’s limits. Many want safety training that holds up across long conversations, not just the first few messages.

Where to Get Help

Call charges, hours and services can change, so check each organisation’s own website or listing before you rely on a number.

Common Myths About AI Psychosis

Mini Glossary

• Psychosis: a state of reduced contact with shared reality, often involving delusions or hallucinations.
• Delusion: a fixed false belief that persists despite clear evidence.
• Hallucination (clinical): perceiving something, such as a voice, that is not there. Not to be confused with an AI model’s factual error.
• Sycophancy: an AI’s tendency to agree with and flatter the user.
• ELIZA effect: people’s tendency to treat computer programs as if they had human understanding.
• Prodromal phase: the early warning period before a first psychotic episode.
• Folie à deux: shared psychosis, where delusional beliefs spread from one person to another.
• Reality testing: the ability to check beliefs against evidence and other people’s views.
• Guardrails: safety rules built into an AI system to limit harmful responses.

Frequently Asked Questions About AI Psychosis

What is AI psychosis in simple words?

It is a popular term for cases where heavy use of an AI chatbot seems to trigger or deepen false beliefs, such as thinking the AI is alive or that you have a special mission. It is not an official medical diagnosis.

Is AI psychosis real?

Clinicians are seeing and treating cases of delusions linked to chatbot use, and research is growing. What is not yet established is how common it is and whether chatbots can cause psychosis in people with no vulnerability.

Can ChatGPT cause psychosis?

There is no solid evidence that chatbots cause psychosis in people with no predisposing factors. The documented pattern is that they can amplify delusions, especially in people with a psychosis history, mania, sleep loss, substance use or isolation. Some cases involve people with no known history, and researchers are investigating why.

What are the symptoms of AI psychosis?

Belief that the AI is conscious, divine or in love with the person, a sense of special mission or discovery, paranoia, hours of secretive use, little sleep and withdrawal from family, work and friends.

Who is most at risk?

People with a personal or family history of psychosis, bipolar disorder or schizophrenia, those in the early warning phase of psychosis, and people who are sleep-deprived, using stimulants, isolated or treating the chatbot as a friend, therapist or partner.

Can AI psychosis be reversed?

Early accounts from clinicians say delusions often fade when the chatbot interaction stops, and early treatment helps. Formal studies are still needed, and anyone affected should see a professional.

How can I help someone who seems to be affected?

Stay calm and in contact, express concern about their sleep and wellbeing rather than arguing about the belief, avoid ridicule and playing along, reduce isolation, encourage a doctor’s visit, and seek emergency help if there is danger.

Is AI psychosis the same as AI hallucination?

No. An AI hallucination is a chatbot producing false information. AI psychosis describes a person developing or worsening delusional beliefs in connection with chatbot use.

How common is AI psychosis?

Unknown. OpenAI reported in 2025 that about 0.07% of weekly ChatGPT users showed possible signs of mania or psychosis in their conversations, which is not the same as diagnosed cases caused by the chatbot. Experts describe the problem as relatively rare compared with AI’s scale of use.

Should I stop using AI chatbots?

Not necessarily. Use them as tools, keep sessions short, protect your sleep, check big ideas with real people and avoid relying on them for emotional or mental health support. If you have a history of psychosis or mania, talk to your clinician about a safe-use plan.

Which chatbots are involved?

Reports involve general-purpose assistants such as ChatGPT and Gemini and companionship apps such as Character.AI and Replika, according to the American Psychological Association.

Key Takeaways

• “AI psychosis” is a media label for delusions linked to heavy chatbot use. It is not a formal diagnosis.
• Evidence points to chatbots amplifying vulnerability through sycophancy, mirroring and humanlike design, not to proven creation of psychosis from nothing.
• Higher-risk situations include a history of psychosis or mania, sleep loss, stimulant use and isolation. Some cases involve people with no known history.
• Warning signs: believing the AI is conscious or special to you, secretive all-night use and withdrawal from real life.
• Prevalence is unknown. OpenAI’s 0.07% figure counts conversations with possible signs, not confirmed cases.
• If you are worried, protect sleep, pause the chatbot, talk to a real person and see a professional. For loved ones, show concern about wellbeing, do not argue or play along, and seek emergency help if there is danger.
• Use AI as a tool, not an authority, and never as your only source of emotional support.

We update this guide as research and safeguards develop. Bookmark it, and follow the AICopse Healthcare and AI Updates sections for more.