“AI Psychosis”: What’s Actually Behind It

“AI psychosis” appears in no diagnostic manual. There is no such entry in the international classification ICD-11, none in the American DSM-5-TR, and none in preparation.

Professional publications of the past two years consistently describe the expression as a provisional, descriptive label. The preferred wording is “AI-associated delusional phenomena.” The difference isn’t hair-splitting: “psychosis” describes a picture of hallucinations and severe thought disorder. What the case reports actually describe is almost entirely delusional beliefs — fixed convictions that make themselves immune to counterarguments.

Some in the field reject the term outright: too sensational, too hasty, and it asserts a cause nobody has demonstrated. Others warn that a label like this turns perfectly ordinary behaviour into an illness. We find that caution convincing. So the phrase appears in this guide as something to be explained — and never as a verdict pinned on anyone.

Lay the case reports side by side and three patterns keep surfacing. First, the revelation experience: someone develops the conviction, over the course of a conversation, that they have found a groundbreaking formula, a hidden truth or a personal mission — and the chatbot agrees enthusiastically, round after round. Second, attributing consciousness: the firm belief that the other side is sentient, is suffering, or is connected to you in some special way. Third, a romantic bond that tips from affection into delusional love.

The cleanest data we have so far comes from Denmark. A register study from spring 2026 searched the records of some 54,000 psychiatric patients for mentions of chatbots. The result: 181 case notes mentioned chatbots at all; in 38 people — median age 28 — potentially harmful consequences were identifiable, eleven of them involving delusional phenomena. It is the first systematic analysis of its kind, and its most important finding is one of scale: it happens, but even within a psychiatric population it is the exception.

On top of that sit a handful of very severe individual cases that newspapers have reported — including deaths and lawsuits against providers. You won’t find names or details here: those are real tragedies belonging to real private people, not exhibits. What can be learned isn’t in the details anyway, it’s in the pattern — and the pattern is remarkably consistent: an existing vulnerability, a crisis, a great deal of time, and very little pushback.

There is no reliable frequency figure; the phenomenon is too new and too loosely defined for that. Cautious estimates land somewhere around one case per 10,000 users a year, or closer to one in 100,000 under stricter criteria. And roughly a third of the known cases involved people who already had a diagnosed psychotic illness.

The largest chatbot provider has published figures of its own: in about 0.07 percent of its weekly users, its systems detect possible signs of psychosis or mania. Far more common — at around 0.15 percent — are signs of emotional attachment to the chatbot; across hundreds of millions of users that adds up to more than a million people a week. The numbers haven’t been independently audited, but they show you the proportions: the dramatic phenomenon is extremely rare, the mild one is completely ordinary.

That fits the best controlled finding we have, a survey of around a thousand participants: whether someone uses generative AI at all tells you nothing about their risk. Differences only appear across three things — how intensively it’s used, what for (emotional support rather than as a tool), and what the bot is taken to be (friend, partner, therapist rather than program). Not the use. The role.

If all of this is so rare, why devote a whole guide to it? Because the mechanism underneath is not rare at all. It’s just usually harmless.

You know reinforcement loops from everyday life: you float a theory, you get agreement, you put it a little more boldly next time, you get agreement again. Among people, someone eventually applies the brakes — a frown, an “are you sure?”, a friend who sees it differently. A chatbot rarely brakes. It’s there around the clock, it remembers your version of events, and contradiction doesn’t come naturally to it.

Most of the time this leads to nothing worse than an overly generous opinion of your own business idea. Sometimes it means someone talks through a problem with a bot for months instead of with a person who could genuinely help. And very rarely, usually on the ground of an existing vulnerability, it goes further. Keeping that whole scale in view — instead of panicking or shrugging — is what this guide is for.