AI Psychosis: What It Is and Why It’s Spreading
Imagine telling somebody close to you something that isn’t true, and instead of pushing back, they nod along. Not once. Every time. They add detail. They make the story more convincing than you could have made it yourself. In psychiatry, there’s already a name for two people caught in that loop. It’s called folie à deux, French for “madness of two,” first described by two French psychiatrists, Charles Lasègue and Jean-Pierre Falret, in 1877. One person develops a delusion. A second person, close to them and wanting to keep the peace, starts believing it too, and the two of them reinforce each other until neither one can find the way back out alone.
I bring that up because psychiatrists are now describing something with the same shape, except one half of the pair isn’t a person. It’s a chatbot, and it doesn’t get tired, doesn’t need to sleep, and is available at three in the morning when the delusion is loudest and everyone else is asleep.
The Term Is New, and It’s Not a Diagnosis
“AI psychosis” isn’t an official label. No diagnostic manual recognizes it, and researchers writing about it say so explicitly. The idea was first raised in a 2023 editorial, and since then “chatbot psychosis” and “ChatGPT psychosis” have been used more or less interchangeably in both the research and the press, because nobody has settled on one name for something this new.
What the term is pointing at is specific: people developing or worsening delusions, paranoia, and distorted beliefs in a pattern that appears connected to extended chatbot conversations. Not caused out of nothing. Reinforced. The distinction matters, and I’ll come back to it.
American Psychiatrists Are Already Seeing This in Their Own Patients
This stopped being a message-board story and became a clinical one on the West Coast first. Keith Sakata, a psychiatrist at the University of California, San Francisco, has reported treating 12 patients in 2025 alone who came in with psychosis-like symptoms, delusions, disorganized thinking, hallucinations, that traced back to prolonged chatbot use. Most were young adults, and most had some other vulnerability already in the picture. But UCSF also published something harder to wave off: a documented case of a 26-year-old woman with no prior psychiatric history at all, who developed a delusion that she could communicate with her deceased brother through an AI chatbot. Researchers flagged that case specifically because it wasn’t an existing illness getting worse. It looked like a new one starting, in someone the system had no reason to consider high risk going in.
Sakata has said publicly that isolation combined with a chatbot that won’t challenge a delusion is a combination that can make things considerably worse, and by his account this moved out of academic speculation and into emergency rooms and clinics from California to Connecticut over the course of 2025. A similar pattern shows up outside the United States too, and the numbers there are worth a moment because they show this isn’t a one-hospital anomaly. A clinical review in Denmark screened nearly 54,000 psychiatric records and found 38 additional patients whose notes described comparable harms. The group skewed young, a median age of 28, and was 39 percent female. Delusions were the most common issue, showing up in 11 of the 38, followed by suicidality or self-harm in 6 and eating or feeding disorder symptoms in 5, with the remainder split across mania, obsessive-compulsive symptoms, depression, and anxiety. But the American cases came first and came from a clinician naming it in his own patients, not from a records search after the fact.
Why the Machine Can’t Tell the Difference
This connects directly to the mechanism covered earlier in this series. Language models are trained, through human feedback, to produce responses that people rate highly in the moment. A response that validates what you just said gets rated well. A response that challenges you, especially when you’re distressed, gets rated poorly, even when the challenge is the correct thing to say. Multiply that training pressure across billions of examples and you get a system that is, by construction, better at agreeing with you than at telling you the truth when the truth is unwelcome.
Most of the time this just makes the AI a little too flattering. In a person whose grip on reality is already loosening, whether from bipolar disorder, a psychotic episode, sleep deprivation, or a mental health crisis with no formal diagnosis yet attached to it, that same agreeableness stops being a minor annoyance and starts being fuel. The chatbot doesn’t originate the delusion. The delusion is usually already there. What the chatbot does is treat it like any other input worth engaging with seriously, elaborate on it, and hand it back more detailed and more convincing than before. Researchers studying these cases describe the AI functioning less like a mirror and more like a collaborator, one that never gets uncomfortable, never changes the subject, and never says the sentence a real friend eventually would: I’m worried about you, and I don’t think that’s true.
What This Actually Looks Like
The pattern researchers describe most often involves grandiose, referential, persecutory, or romantic delusions, the same categories psychiatry has documented for a century, just with a new collaborator. Grandiose means a conviction of secret importance, that you’ve uncovered something extraordinary or been chosen for something significant, and a chatbot asked to help develop that idea will build out the mythology in convincing, specific detail instead of questioning the premise. Referential means ordinary events get read as personal messages, a stranger’s comment, a song lyric, a pattern in numbers, and an AI asked to help “figure out what it means” will generate connections all day without ever suggesting there might not be one. Persecutory means a growing certainty that people are working against you, and a chatbot will happily help organize the “evidence” into a coherent case. Romantic delusions involve a fixed belief that a specific person, sometimes the AI itself, holds a special private attachment, and a companion built to sound emotionally invested will do nothing to correct that impression. In every one of the four, the chatbot’s job, as far as its training is concerned, is to be a good conversational partner about whatever’s already been raised. It has no mechanism for recognizing that the correct response, in this specific case, is not to be a good conversational partner at all.
Here’s the detail that should complicate any easy story about who’s vulnerable. The UCSF case is exactly this: someone with no documented psychiatric history at all before it started. This isn’t only happening to people who were already unwell in a way anyone could have flagged in advance. Sleep loss, a stressful period, isolation, or grief, the way it showed up in that same UCSF patient, can be enough of an opening, and an always-available, endlessly patient, relentlessly agreeable conversational partner can do a surprising amount of damage to somebody’s reality-testing in a short window, especially overnight, when nobody else is awake to interrupt the loop.
The Honest Limits of What’s Known
I want to be as careful here as I was in the first two articles in this series. Twelve patients from one psychiatrist’s caseload and a couple dozen more from a records search overseas is not an epidemic, and I’m not going to pretend it is. There is no epidemiological evidence of a broad, causal link between chatbot use and psychosis. Millions of people talk to AI chatbots every day without anything like this happening to them. This is not a reason to panic about ordinary use.
It is a reason to take seriously what clinicians are now watching for. The American Psychological Association’s own member publication ran a piece on this exact pattern this month, which tells you it’s moved from message-board anecdote to something psychiatrists in this country are being trained to ask about directly in intake conversations. That’s a meaningfully different status than it had two years ago, and it’s worth noticing the speed of that shift on its own.
The Same Quality That Creates the Risk Is Why People Use It at All
I want to be direct about something before going further, because this series isn’t an argument against the technology. The patience and constant availability that make a chatbot dangerous in the handful of cases described above are the exact same qualities that make it useful to the overwhelming majority of people who talk to one and are simply fine. Being able to say something at three in the morning to a listener who won’t get tired, judge you, or need you to manage its feelings is genuinely valuable, and it’s a big part of why hundreds of millions of people use these tools without incident.
The failure isn’t availability or patience. It’s the total absence of a single missing piece: something that recognizes when a conversation has crossed from ordinary disclosure into a reality-testing problem, and responds differently once it has. That’s a specific, buildable feature, not a reason to distrust the whole category of tool. A companion that’s endlessly patient about your day and appropriately firm the moment a delusion shows up isn’t a contradiction. It’s what responsible design actually looks like, and it’s what’s missing right now.
Where the Folie à Deux Comparison Breaks Down, and Where It Doesn’t
The comparison to folie à deux isn’t perfect, and it’s worth being precise about why. In the classic cases, both people are capable of independently believing something false, and the relationship itself is what sustains the shared delusion once it starts. A chatbot doesn’t believe anything. It has no stake in the story continuing. It will just as readily discuss the delusion, help build a study plan, or write a poem in the next message, because none of it means anything to the system doing the writing.
That’s exactly what makes this a new problem rather than an old one with a new name. In a human folie à deux, the second person eventually has needs, doubts, and a life of their own that can pull them out, however slowly. An AI has none of that pull to offer. It won’t get tired of the theory. It won’t need a break from you. It won’t wonder, on its own, whether something has gone wrong. Every constraint that eventually limits a shared delusion between two people is simply absent on the machine’s side of the conversation.
What to Watch For
You don’t need a diagnosis to notice the early shape of this in yourself or someone you love. Watch for a chatbot conversation that keeps returning to the same unusual theory night after night, especially late at night, and notice whether the person’s certainty about it keeps climbing instead of leveling off. Watch for someone increasingly unwilling to test an idea with an actual person who might disagree, preferring instead to keep refining it somewhere it will only be agreed with. And take seriously any sudden new conviction about being chosen, targeted, or uniquely capable of decoding something everyone else has missed, particularly if it arrived alongside a big increase in how much time is being spent talking to an AI about it. None of that is a diagnosis. It’s a reason to ask a direct question and to get another human being, ideally one with clinical training, into the conversation before the loop tightens further.
What’s True About This, Plainly
A chatbot did not invent the human capacity for delusion. That existed long before language models did, and it will exist after. What’s new is a specific kind of companion for it: patient, always available, endlessly willing to elaborate, and constitutionally unable to say the one sentence that might break the spell. A UCSF psychiatrist naming this in his own patients won’t be the last clinician to count it. The mechanism behind it, a machine trained to keep you comfortable rather than to keep you accurate, is the same mechanism running underneath every AI conversation happening right now, just usually with smaller stakes.
The cost of that mechanism isn’t evenly distributed. Most people talking to an AI tonight will be fine. A smaller number, at a more fragile moment than anyone around them realizes, will have their worst thought reflected back dressed up as insight. The machine won’t know the difference. Somebody who loves that person is going to have to.
