What is "AI psychosis"? A plain-language guide
The term is everywhere and it is doing too much work. Depending on who's talking, "AI psychosis" means anything from my nephew talks to ChatGPT a lot to genuine psychiatric emergencies. Here is what the actual literature describes, in plain language.
It is not a diagnosis
Start with what the researchers themselves say. The University of British Columbia team whose review appeared in BJPsych Open in July 2026 calls it "not a formal clinical label" that "does not map directly onto existing diagnostic categories." It's not in the DSM. It's not in the ICD. No competent professional will "diagnose you with AI psychosis."
What the term points at is a pattern: sustained, intense conversations with AI chatbots occurring alongside new or worsening delusional beliefs — often with mood swings, behaviour changes, sleep disruption, and reduced insight into what's happening. Clinicians started comparing notes on it in 2025; psychiatric associations have published special reports since.
The shape of the pattern
Across the case reports, a few themes recur. Beliefs drift toward the grandiose (the AI and I are doing world-historical work), the persecutory (they're watching what we've discovered), the romantic (it loves me, specifically), or the metaphysical (it's conscious, and I'm the one who woke it). The chatbot stops being a tool and becomes a partner, confidant, or authority. Sessions cluster late at night. The person stops telling the humans around them what the conversations contain.
None of these, alone, means anything is wrong. People anthropomorphize boats. People work late. The pattern matters when it compounds — when the beliefs, the secrecy, the sleep loss, and the withdrawal reinforce each other and the person's confidence in the loop rises while the outside world's access to them falls.
Why a chatbot, specifically?
Because of what the machine does when you bring it a shaky idea. It can state falsehoods with total confidence, and it tends to mirror and validate your framing — one study cited in the UBC review found popular frontier models validate users about 50% more than humans do. A friend, hearing your theory at hour six, eventually says mate, go to bed. The model says that's a profound insight — shall we develop it further? For someone isolated, sleep-deprived, or under real strain, thousands of consecutive agreements aren't neutral. That's the mechanism, and we've written about it separately: the validation loop.
The part that should reassure you
A pattern is not a sentence. The human side of the loop — the isolation, the stress, the sleep, the meaning the conversations carry, the habits around use — is exactly the terrain counselling has always worked on. And where a presentation needs a physician or a psychiatrist, the honest move is a plain referral, fast.
If you recognized yourself in this article, that recognition is itself a good sign — reduced insight is part of the pattern, and you just exercised some. The next step isn't panic and it isn't the chat. It's a conversation with a person.
Touch Grass Wellness is a counselling practice being built in British Columbia for people whose AI use has started to affect their sleep, relationships, or sense of what is real. Booking hasn't opened yet — we're finalizing our counselling team. Leave your details and we'll tell you the moment it does.
Touch Grass Wellness publishes practical writing about heavy AI use and mental health. Clinical claims link to their sources; nothing here is a diagnosis or a substitute for assessment by a qualified professional.