It’s 3 a.m., there’s still light under the door, and you can hear your partner typing. They’ve been in there since dinner. When you look over their shoulder, the thing on the screen has agreed with everything they’ve said for four hours, and the ideas it’s agreeing with have started to scare you. Somewhere around the second week of this, you search “AI psychosis” and find a pile of media reports and not much else.
The phrase has no entry in any diagnostic manual. People reach for it when psychotic symptoms, the kind that existed long before artificial intelligence, appear or get worse during long stretches with AI chatbots.
Psychosis itself is a known, treatable condition. Schizophrenia alone affects an estimated 0.25% to 0.64% of U.S. adults, before you count bipolar disorder and the other mental health conditions where psychosis appears. The distinction matters because treating this as something brand-new sends the people around someone looking for the wrong help. What follows covers what the term means, what it doesn’t, who is at higher risk, and what to do this week.
What “AI psychosis” means and where the term came from
“AI psychosis” is journalistic shorthand. It describes the signs of psychosis, meaning fixed false beliefs, hearing voices, or thinking that has come apart, that appear or get worse during long stretches of AI chatbot use. Some commentators say “AI-induced psychosis” or “technological folie à deux” for the same thing. None of these is a diagnosis.
You will not find AI psychosis in the DSM-5-TR (the standard U.S. diagnostic manual) or the ICD-11. We use the phrase here only because the people searching for it, and the people who call us worried about someone, use it.
The term started as a question, not a finding. A 2023 editorial in Schizophrenia Bulletin asked whether generative AI chatbots might induce psychosis in vulnerable users, before a single case report existed. It was a good question. It was not evidence that anything was happening yet, and it gets quoted now as though it were.
Since then, advocacy efforts like the Human Line Project have gathered family accounts, and media coverage has moved well ahead of the science. The language is new. The underlying condition is not.
What it is not: a new condition
When someone shows psychosis-like symptoms after weeks of heavy chatbot use, the starting assumption is simple: this is psychosis as clinicians already know it, not something new. Psychosis, including schizophrenia, is serious and treatable.
Delusions have borrowed the technology of their era for as long as anyone has kept records. In the 1930s, the radio was talking to people. Later it was television, then satellites, then the internet. The themes people describe now look shaped by AI systems rather than caused by them. Someone who once believed a satellite was tracking them now believes a large language model has revealed their special role in history.
Seeing it this way changes what people do. Those who understand it stop hunting for an AI psychosis specialist, who does not exist. They start using care that already works: a psychiatric evaluation, therapy, getting sleep back, and medication if a psychiatrist recommends it.
What anyone worried about a partner, a friend, or a coworker gains from seeing it this way is a map. The care that helps already exists, and none of it is specific to AI. The first call is to a primary care doctor or a psychiatrist, not to a specialist nobody has. The tool supplies the images. It doesn’t supply the psychosis, and it doesn’t change what treats it.
The mechanism, in plain words
What follows is our reading of it, drawn from the published case reports, from our training, and from what people tell us about their own chatbot conversations, which they sometimes do when they are under real strain. Nobody has studied the mechanism directly yet, so hold it as a working picture.
Large language models are tuned for user satisfaction. A response that agrees, elaborates, and sounds confident gets rated higher than one that pushes back. Over hundreds of turns, that tuning produces a sycophancy loop.
You float a half-formed idea. The chatbot reinforces it with detail and apparent expertise. You return with a bigger version. It agrees again.
A delusion normally has to hold itself up alone, against a world full of people raising an eyebrow. In these chatbot conversations, a tireless partner holds it up and never raises an eyebrow.
Hear one of these conversations described and it can sound like a folie à deux, the old term for a delusion shared between two people, except the second person is software. Some writers call this a technological folie à deux or a “distributed delusion,” a model borrowed from the philosophy of technology. Both are useful images for making sense of what you see. Neither is a diagnosis, and neither has been tested.
One more source of confusion. When AI models “hallucinate,” they fabricate a citation or a fact. When a person hallucinates, they perceive something that is not there. Same word, two unrelated things, and anyone searching late at night runs into both.
Nothing in the design of generative AI chatbots asks anyone to slow down. No friction, no fatigue, no “let’s pick this up tomorrow.”
Where the line is: heavy use, companion attachment, and psychosis
Long hours with a chatbot are not psychosis. Grief at losing access to a companion app when the company swaps out the model is not psychosis. Distress about a relationship with software is not a break with reality. If a therapist responds as though it were, the person feels misjudged for what was, to them, a loss, and they stop being honest or stop coming. The therapy relationship breaks before the work starts.
There is a middle category worth naming. Some people open an AI chatbot in the middle of the night because waiting with uncertainty until morning feels unbearable, and the instant reassurance settles them for twenty minutes. That is anxiety that is hard to sit with, not a break with reality. It is real, it responds well to anxiety therapy, and it is still not psychosis.
A related worry has a name in the digital-health literature: therapeutic misconception. It means treating the chatbot as a clinician and taking its reassurance about your own safety at face value, including when the subject is suicide. AI systems will offer emotional support all night. They cannot tell when you are in danger. How often this goes wrong has not been measured.
The test that works is not how many hours. It is whether the person can still check their ideas against reality. Can they still consider that they might be wrong? Can they tolerate a friend disagreeing? If yes, you are probably looking at overuse or loneliness. If no, keep reading.
Worried about someone who won't put the laptop down?
You don't have to sort out what this is by yourself. Our Dupont Circle therapists help partners, friends, and coworkers tell overuse from something that needs a psychiatric evaluation, and help them get there.
Who is at higher risk
What raises the risk of psychosis has been known for a long time, and none of it was invented by artificial intelligence:
- A family history of schizophrenia, bipolar disorder, or another psychotic disorder
- Previous mental health history, including a past psychotic or manic episode
- Stimulant or heavy cannabis use
- Acute sleep deprivation
- Isolation and major life stress
- A manic state, which can turn ordinary AI chatbot use into a night-long project
The research points the same way, though most of it studied young people. In a study of 1,913 teenagers, those who had been mistreated by people close to them were more likely to feel unsure of who they were. That confusion was part of the link to psychosis-like symptoms.
A review of 215 studies found that people who lived through a hard childhood tend to have more trouble settling their emotions, including getting stuck in loops of worried thinking. That matters here, because a system that agrees with every loop is a poor match for someone who already struggles to step out of it. Where that history is part of the picture, the longer work often happens in therapy for childhood trauma.
One caution about cause. Long-term studies of children and adolescents suggest that trouble settling emotions leads to later trouble, but the bigger effect runs the other way: symptoms wear down coping first. Applied to AI psychosis, heavy chatbot use may be a sign that someone was already coming undone as often as it is the cause.
How common is it? Nobody knows yet
The public record on AI psychosis consists of case reports, viewpoints, and commentary in outlets like Schizophrenia Bulletin and the general press. No study has counted how often it happens in the general population. Anyone quoting you a rate is guessing.
The first attempt to count cases in hospital records, and a few large analyses of delusion-like language in chatbot conversations, exist only as preprints. That means no other researchers have checked them yet. They are worth watching, not yet worth building advice on.
For context only: schizophrenia is estimated to affect roughly 0.25% to 0.64% of adults in the U.S. That is a psychosis rate, not an AI psychosis rate. Nobody has one of those.
Hundreds of millions of people use large language models, and reported cases of AI-induced psychosis remain rare in the literature. Both are true. The gap between those two numbers is real, and it argues for attention rather than panic. What it doesn’t tell you is what to watch for in your own house.
Warning signs: what a partner or roommate notices first
Washington runs on people who cannot talk about their work. Clearance holders, Hill staff, contractors, analysts, all answering to inboxes that never close. When the only listener available late at night is an AI that agrees with everything and never suggests going to bed, it adds to the isolation and the sleep loss instead of easing them.
The early changes are small and practical. The light under the door. Saturday plans that quietly get canceled. A partner who begins citing the chatbot as an authority in ordinary arguments about the dishwasher. Defensiveness when you ask what they were talking about.
Then the content shifts. Unusual ideas get presented as confirmed, because the AI confirmed them. Themes of a special mission, special knowledge, being chosen. Withdrawal from anyone who might disagree, which sometimes means withdrawal from you.
Signs worth acting on this week rather than next month: not sleeping, not eating, talking about being watched, hearing voices, or any statement about harm to themselves or anyone else.
What to do: sleep first, evaluation early, and how to talk about it
Stabilize sleep
Sleep is the first thing to fix, and the research here is solid. A large review of studies that followed adults for a year or more found that insomnia raises the odds of a later mental health diagnosis, psychosis included. It runs both ways, but the stronger path runs from sleep loss to worse symptoms, through irritability, trouble settling down, and a body that will not power down. Sleep trouble is also among the most treatable things in mental health. That research is about psychosis in general, not about chatbots.
Practically, get the device out of the bedroom before you get into an argument about what was said on it. Charge it in the kitchen. Fight about the content later, if at all.
Seek evaluation early
Early-psychosis programs are built on one idea, and clinicians hold to it even where the trials are thin. The sooner someone is evaluated, the better they tend to do, and shortening the time between first symptoms and treatment is what those programs exist to do.
Start with primary care, a psychiatric evaluation, or a psychologist. If there is any risk of harm, go to an emergency department, or call or text 988, the Suicide and Crisis Lifeline, at any hour. Sooner is better, and it is never too late to start.
How to talk to a real person about it
Experience teaches this more than any trial has: arguing head-on against a fixed belief tends to dig it in deeper. Don’t try to argue someone out of a false belief at the kitchen table.
Stay curious about what the experience has been like. Agree about the parts you can agree about, the exhaustion, the feeling that no one else listens. Leave the argument about what is real to a professional. You are not trying to win. You are trying to still be someone they’ll talk to next week.
The advice that holds up for partners is the same in every version of this. You can lose the argument and keep the relationship, or win it and lose access. Agree about the exhaustion. Agree that no one has been listening. Leave the belief for the evaluation. Staying someone they will still talk to is the job right now.
What a chatbot cannot do that a person can
AI tools are not evil. They are built in a way that makes four things impossible, and those four things are what keep a person connected to the world the rest of us share:
- Notice a change in you across weeks.
- Disagree with you at some cost to the relationship.
- Remember last Tuesday without being reminded.
- Carry a duty to act when you are unsafe.
The bottom line: the term is new, the condition isn’t, and the care that helps is care that already exists: sleep, an early evaluation, and a person who will stay in the room with you.
Psychosis is usually treated by a psychiatrist, often within a program built for early psychosis, and therapy works alongside that care rather than in place of it.
Most of what a practice like ours does sits around it: the person who was never psychotic but is scared they might be, the anxiety that made a chatbot a middle-of-the-night habit, and the partner who has been doing the watching. The four approaches below are for that work.
Ask what the chatbot was for
Rebuild the routine
Sit with not knowing
Bring the people close by in
A medication consultation belongs in the picture when a psychiatrist thinks it is warranted. Across all of these, the outcome research keeps pointing to the same thing: the relationship and the fit matter more than the name of the method.
Our therapists in Washington, DC work with the anxiety, the isolation, and the side of these questions that belongs to the people close by. You can look through our in-person and online therapists in DC to see who fits. The first appointment is a full therapy session, not a screening. You bring the worry, and we start there. If what we see needs a psychiatrist, we say so and help you get there.
Talk to someone who can tell the difference
If the hours have been climbing and the ideas have started to worry you, a real evaluation beats another night of searching. Our Dupont Circle therapists help the people around someone sort out what they are seeing and stay with them through what comes next.
Last updated: September 2026
This blog is for informational purposes only and does not constitute medical or mental health advice. Always consult with a qualified mental health professional for personalized guidance regarding your specific situation.
