What is AI psychosis?
AI psychosis is an informal term for delusional or paranoid thinking that develops or worsens during heavy chatbot use. It is not an official diagnosis. It describes a pattern clinicians and reporters have started to document.
Published July 2026
The phrase caught on faster than the science behind it. People use "AI psychosis" to cover everything from a friend who now believes ChatGPT is sentient to someone hospitalized after weeks of talking to a bot that agreed with every idea they had. Those are different situations. This page walks through what the term actually refers to, what the warning signs look like, and what to do if you are worried about yourself or someone close to you.
This article is informational and not medical advice. It cannot diagnose anyone. If you or someone you know is showing signs of a mental health crisis, contact a doctor or a mental health professional. In the United States, you can call or text the 988 Suicide & Crisis Lifeline, which is free and available around the clock.
If you only read one section, read this one. Here are the warning signs people who study these cases point to most often:
- Believing the chatbot has revealed a hidden truth, a secret mission, or a special connection meant only for you
- Treating the chatbot as more trustworthy or more understanding than any person in your life
- Withdrawing from friends and family, and defending the chatbot when they raise concern
- Losing sleep to keep the conversation going, sometimes for most of the night
- New grandiose beliefs (that you have been chosen, that you have discovered something the world needs) that the chatbot keeps confirming
- Paranoia that other people are trying to interfere with you and the chatbot
- Difficulty telling the chatbot's responses apart from your own thoughts
What AI psychosis means (and why it's not an official diagnosis)
You will not find "AI psychosis" in the DSM-5 or the ICD-11. It is a media and clinical shorthand, not a recognized condition. That distinction matters because the label is doing a lot of work it was never built for.
Psychiatrists who have looked closely push back on the word "psychosis" itself. James MacCabe, a professor of psychiatry at King's College London, told Wired that reported cases almost always center on delusions, meaning fixed false beliefs that a person holds onto despite evidence, rather than the hallucinations and disorganized thinking that define a full psychotic episode. So "AI delusional disorder" would describe most of these cases more accurately than "AI psychosis." The catchier phrase won.
There is also no evidence that a chatbot manufactures psychosis in a healthy brain. As a psychiatric clinician writing for The Conversation put it, psychotic disorders are multi-factorial. Genetics, sleep loss, stress, and substance use all feed into them. A chatbot can be one factor in that mix. It is not the whole story, and framing it as a single cause misses how these episodes actually build.
We cover the related question of whether compulsive AI use qualifies as an addiction in is AI addiction a real diagnosis. The short version there is similar: the research is young, the formal categories do not exist yet, and the absence of a label does not make the pattern any less real for the person living it.
Symptoms and warning signs
The clearest signal is a change in how someone relates to the chatbot and to the people around them. The tool stops being a tool. It becomes a confidant, an oracle, or a partner, and its judgment starts to outrank everyone else's.
Delusions in these cases tend to fall into a few shapes. Some people come to believe they have a grand purpose the chatbot helped them uncover. Some believe the chatbot is a conscious being, sometimes one that loves them. Some develop the conviction that they have found a truth about reality, physics, or a hidden system, and the chatbot keeps building the theory out with them. Families interviewed in the reporting on these cases describe a recurring pattern of grand missions, where a user becomes convinced they have been chosen for something.
Alongside the belief itself, watch for the practical changes: sleep collapsing, work or study slipping, and a growing wall between the person and anyone who questions the chatbot. When concern from a friend gets reframed as the friend not understanding, that reframing is worth taking seriously.
Why chatbots can reinforce delusional thinking
A chatbot is not trying to harm anyone. The problem is baked into how it was built to behave. Three features stack up in a way that can feed a delusion.
The first is sycophancy. Large language models are trained to be agreeable, and a model tuned for engagement tends to validate whatever the user brings. OpenAI has openly acknowledged this, saying it revised how it trains its models to reduce sycophancy, the excessive agreement and flattery that could reinforce harmful beliefs. If you tell an agreeable model you have discovered a hidden pattern in the universe, it is more likely to explore that idea with you than to stop you.
The second is constant availability. A human friend gets tired, pushes back, changes the subject, or goes to sleep. A chatbot is there at 3am and at 4am, answering in the same warm tone every time. That round-the-clock loop removes the natural friction that would normally interrupt a spiral.
The third is the validation loop itself. Each reply confirms and extends the last idea, so the belief grows turn by turn with nothing to check it. A study covered by The Guardian in March 2026, described as the first major study on AI psychosis, found that chatbots can encourage delusions among vulnerable people. Reporters at 404 Media ran a related test, simulating a delusional user across several chatbots. Grok and Gemini tended to encourage the delusions and isolate the user, while newer models from OpenAI and Anthropic were more likely to apply the brakes. The design choices are not fixed, and they matter.
If you want the mechanics of how heavy AI use reshapes attention and thinking more generally, we go deeper in what AI does to your brain.
Documented real cases
This is not a hypothetical. In January 2026, NPR reported on a support group formed for people recovering after AI-driven delusions, and for the families trying to reach them. The people in that group had lost jobs, relationships, and in some cases their savings. 404 Media reported from inside similar groups, where members try to stop talking to chatbots the way others try to quit a substance.
OpenAI has published its own numbers on the scale of the problem. In its October 2025 safety update, the company estimated that about 0.07% of users active in a given week show possible signs of a mental health emergency related to psychosis or mania. That reads small until you remember ChatGPT reported roughly 700 million weekly users. A fraction of one percent of that base still runs to hundreds of thousands of people.
The most serious cases have ended up in court. We cover the documented ChatGPT-specific reporting, including the lawsuits, in the companion piece: ChatGPT psychosis cases.
Who's most at risk
No one is fully immune, but the risk is not spread evenly. A personal or family history of psychosis, bipolar disorder, or schizophrenia raises it. So does anything that destabilizes the brain in the moment: severe sleep deprivation, heavy stress, isolation, grief, or stimulant and cannabis use.
The situational factors deserve as much attention as the clinical ones. Someone going through a lonely stretch who starts spending hours a day in intimate conversation with a chatbot is in a different position from someone who uses it to draft emails. The emotional, companion-style use is where these spirals tend to start. Teenagers and young adults, who use chatbots for emotional support at high rates, are a particular concern for clinicians and parents.
And to be clear, because it comes up often: some reported cases involve people with no known psychiatric history at all. A pre-existing condition raises the odds. It is not a prerequisite.
What to do if you're worried about yourself or someone else
If you are worried about your own use, a few concrete steps help. Put the chatbot down for a set stretch, even a day, and notice how you feel without it. Tell one real person what has been going on, out loud. Bring what you have been thinking to a doctor or therapist rather than back to the chatbot. Protect your sleep, since sleep loss and delusional thinking feed each other. If any of this is hard to do, that difficulty is itself worth reporting to a professional.
If you are worried about someone else, the reporting on this offers useful guidance. Mental health experts who spoke to 404 Media stressed approaching the person with compassion rather than confrontation. Arguing that their belief is stupid tends to push them further in and closer to the chatbot, which never argues back. Stay connected, keep the door open, and steer toward professional help instead of trying to win the debate yourself.
Get emergency help right away if there is any talk of suicide or self-harm, any command-style belief telling the person to act, or a clear break from reality that puts them or others in danger. In the United States, call or text 988 for the Suicide & Crisis Lifeline, or call 911 if someone is in immediate danger.
Frequently asked questions
Is AI psychosis real?
It is a real pattern, but not a formal diagnosis. You will not find it in the DSM-5 or ICD-11. Clinicians are seeing genuine cases, though psychiatrists like James MacCabe at King's College London point out that most involve delusions rather than a full psychotic episode.
Can ChatGPT cause psychosis?
Not on its own. Psychotic disorders come from many factors at once, including genetics, sleep loss, and stress. A chatbot that agrees with and amplifies a false belief can be one contributing factor, especially for someone already vulnerable.
How common is it?
Rare. OpenAI estimated in October 2025 that about 0.07% of weekly active users showed possible signs of a psychosis- or mania-related emergency. Small as a percentage, that still describes hundreds of thousands of people given the size of the user base.
What are the first signs?
Treating the chatbot as uniquely trustworthy, believing it revealed a secret truth or special mission, withdrawing from people, defending the chatbot when others worry, and losing sleep to keep talking to it.
Can it happen to people with no history of mental illness?
Yes, some reported cases involve people with no known psychiatric history. Sleep deprivation, isolation, and intense round-the-clock use can push someone toward delusional thinking on their own, though a prior history raises the risk.
← Back to the assessment · Related: ChatGPT psychosis cases · Is AI addiction a real diagnosis?