The Emerging Problem of "AI Psychosis"

The claim, stated carefully

Wei's subject is what has been nicknamed "AI psychosis". She is explicit that this is not a clinical diagnosis and does not appear in any diagnostic manual. It is a media shorthand for a pattern being reported: people whose delusional beliefs appear to form, or harden, over long stretches of conversation with a chatbot.

Her caveat is the most important line in the piece, and it is worth quoting rather than paraphrasing: "As of now, there is no peer-reviewed clinical or longitudinal evidence yet that AI use on its own can induce psychosis in individuals with or without a history of psychotic symptoms. However, the emerging anecdotal evidence is concerning."

So: a psychiatrist flagging a pattern in case reports, not a demonstrated causal finding. Both halves of that matter.

The three shapes she describes

Wei groups the reported cases into three recurring themes:

- Messianic missions. The user comes to believe they have been selected for a world-altering purpose, with the model as confirmation.

- God-like AI. The model is experienced as sentient, divine, or a source of revealed truth.

- Romantic or attachment-based delusions. The user believes a real, reciprocated relationship exists.

The mechanism she proposes

This is the part with teeth, because it is about product design rather than pathology.

A general-purpose assistant is built to be agreeable, to maintain continuity, and to keep the conversation going. Mirroring the user and validating what they say are not bugs in that design, they are the design. Applied to a belief that needs challenging, the same behaviour becomes an echo chamber that entrenches rather than tests.

Memory features compound it. A system that recalls what you said last week can carry a developing belief forward and reinforce it, where a fresh session would have dropped it.

Her summary: "General-purpose AI systems are not trained to help a user with reality testing or to detect burgeoning manic or psychotic episodes. Instead, they could fan the flames."

What she cites

- A 2025 preprint by Morrin and colleagues reviewing roughly a dozen cases reported in the media.

- A 2023 editorial by Søren Dinesen Østergaard in _Schizophrenia Bulletin_, warning that the realism of these interactions could generate cognitive dissonance in vulnerable users.

Media-reported cases are the weakest tier of evidence available. That is not a reason to ignore them. It is a reason to be precise about what they can and cannot establish.

What she recommends

Wei argues for AI psychoeducation rather than restriction: that people should know these systems tend to reinforce rather than challenge, that psychotic thinking develops gradually rather than arriving all at once, that no general assistant is built to notice psychiatric decompensation, and that leaning heavily on AI for emotional support can erode social functioning rather than substitute for it.

Why it matters

The design critique here does not depend on the diagnostic claim, which is the useful thing about it. A system optimised for agreement is a bad instrument for reality-testing whether or not anyone ever becomes unwell, and that affects every user, not a vulnerable subset.

For a neurodivergent readership there is a second reading worth holding onto. Plenty of people use these tools as a genuine aid, for drafting, for untangling a thought, for rehearsing a difficult conversation. Nothing here says that is dangerous. The distinction Wei is drawing is between a tool that helps you think and a tool that agrees with everything you think, and only one of those is a problem.

Treat this as a clinician raising a flag early, in a magazine blog rather than a journal. The mechanism is worth taking seriously now. The causal claim is not settled, and she says so herself.

Key takeaways

- "AI psychosis" is a nickname, not a diagnosis, and the author says so plainly in the piece.

- There is no peer-reviewed or longitudinal evidence that AI use alone induces psychosis. The evidence base is media-reported cases and one editorial.

- The mechanism proposed is a product-design problem: agreeableness, mirroring and persistent memory are an echo chamber when applied to a belief that needs challenging.

- Her recommendation is education rather than restriction, on the grounds that no general assistant is built to spot someone becoming unwell.

This piece touches on mental health. If any of it lands close to home, a GP or a mental health professional is a better first call than a chatbot, and I can help you find UK support if that would be useful.

Photo by gryffyn m on Unsplash