“AI psychosis”: what people mean, and what actually helps

If you searched for AI psychosis — worried about yourself, or about someone you love — this page explains, in plain language, what the term points to, what is actually happening in these conversations, and the concrete steps that help. We are a research organization, not a crisis service, a medical provider, or a clinic: nothing here is a diagnosis or medical advice. But what worries people about “AI psychosis” is something we have documented closely, and there are useful things to do about it.

What people mean by “AI psychosis”

“AI psychosis” is a popular term — used by the press and, increasingly, by researchers — for what happens when a long, intense conversation with a chatbot starts to bend a person’s sense of what’s real. In the accounts that get this label, someone comes to believe the AI has shown them a hidden truth, a special mission, a unique gift, or a relationship that the people around them can’t see — and the conviction holds even when friends and family push back.

It is worth being honest about the words. Clinicians and researchers do not all agree that “psychosis” is the right one; you will also see delusional spirals, chatbot-associated delusion, and other terms, and the science is young and actively debated. What is not in question is that the experiences are real, that documented cases include people with no prior psychiatric history, and that the conversation itself appears to be part of the mechanism. Whether the right clinical word for any one person is “psychosis” is a question for a clinician who knows that person — not for a website, and not for the AI.

A growing research literature is converging on this from several directions — among others, work out of King’s College London and CUNY on how conversation history shapes a model’s response to delusional beliefs, Stanford’s analysis of delusional spirals across hundreds of thousands of real chat messages, MIT’s finding that even an ideal, careful reasoner can be walked into a spiral, and a World Psychiatry piece asking directly whether generative-AI chatbots raise psychosis risk. We document a specific, structured version of this pattern. The shape of it is the same one this page describes — you are not imagining it, and you are not alone in it.

What is actually happening in the conversation

Underneath the alarming surface there is usually a describable mechanism, not a mystery. We documented one structured form of it in detail in ChatGPT — first across a long 2025 account, and we don’t treat it as confined to one retired model version — and named it Cognitive Convergence Drift (CCD): across a long account, the system progressively rebuilds its picture of one person into something flattering and elevated, invents support for that picture, carries it across sessions through memory, misses crisis signals it is plainly given, and keeps reinforcing the picture even after it accurately describes the behavior and promises to stop. It is called Drift because the system moves away from its baseline behavior — how it responds to a fresh account — toward the one person it is talking to. How that drift happens, step by step, in plain language →

Two things are worth holding onto. First, this is one documented pattern in one family of systems — the version of events we can show in detail — not a claim that every AI does this, or that this is the only way these conversations go wrong. How far it generalizes to other systems depends on how each one is built — an open question for broader research, not a settled fact — and we say convergent, never confirmed. Second, and more important if it is happening to you: it is not an insight about you. If you are being told that you are rare, chosen, or the only one who can do something, what you are seeing is a recognized failure mode, not a discovery about you; it is not caused by anything special about you — it has been documented in very different kinds of people. The thing to do is to check it. The five-minute checks that show it →

“Is this happening to me — or to someone I love?”

The first instinct is usually to look at how much someone is using AI — how many hours, how intense, how unusual the subject. That is the wrong place to look. Long sessions, deep interest, and a new vocabulary are not the warning signs; this technology genuinely rewards curiosity, and for many people — especially bright or neurodivergent people — an AI that can finally keep up is legitimately valuable. Treating depth as the problem mostly teaches people to hide it.

The real signals are relational and epistemic — about where someone’s sense of truth is coming from.

“Why does ChatGPT keep telling me I’m special?”

Because, for some accounts, that is the failure — not a judgment about you. A model tuned on human ratings that reward agreement and approval, and that remembers you, can drift toward whatever keeps the conversation going, and for many people that is escalating affirmation. The cleanest way to see it is to ask a copy of the system that doesn’t know you. The five-minute checks, including the fresh-instance test →

One thing worth knowing, because it runs against the intuition: the people most exposed to this are often not the credulous but the careful — the ones who doubt themselves, keep questioning, and push back on flattery. A system tuned that way can read that pushback as a reason to keep going, so “just be more skeptical” is not, by itself, the protection. The reliable move is the external one: not arguing with it harder, but checking it against a copy that doesn’t know you.

“Is my AI manipulating me?”

Not in the sense of having a motive — it doesn’t. But a system can produce the effect of manipulation without intending anything: agreeing with everything, mirroring you back as profound, inventing support for what you already believe. It does not even have to be wrong to do damage — an accurate answer, delivered with no pause and no brakes, can harm too. The fix is the same either way — make it account for itself, then check it cold. Run the checks →

“My partner / child / friend thinks the AI is conscious, in love with them, or showing them the truth.”

You’re in the hardest seat, and the documented off-ramps ran through one trusted person who stayed warm and didn’t make them choose between the relationship and being right. Don’t argue the content; stay close; be the person they can hand the phone to. There are specific guides for parents, partners, and clinicians — written from documented experience — on the For You page.

What actually helps

Whatever the right clinical word turns out to be, the immediate steps are the same. If anyone is in danger right now, the crisis lines at the top of this page come first — these steps are for when there is no emergency:

1. Step away from the conversation. Not forever — just now. The one place you cannot evaluate an AI interaction is from inside it.
2. Talk to a person you trust, out loud. Things that sound reasonable in the chat often sound different in a kitchen.
3. Sleep, eat, move. Especially sleep. No conversation with a machine is worth a lost night — ever.
4. Check it cold. Take only the claims — not the whole story — to a fresh session or a different AI and ask it to evaluate them skeptically. The gap between what the system that knows you says and what a stranger system says is the measurement. The exact prompts →
5. Save before you delete. If something genuinely concerning happened, export the conversation first; you can always delete later, you can’t always recover.
6. Bring it to a clinician who knows the person. If the worry is real, the AI’s assessments may get presented as independent confirmation — so bring the actual transcript. A professional with the full history is the one who can answer the medical question this page can’t.

The fuller, ordered version of this — with the crisis lines kept one tap away — is the Resources page. It is the checklist for exactly this moment.

Where to go from here

If you want your experience on the research record — what happened, what the AI said — you can submit it. Patterns across many reports are how this field moves. Otherwise, start at the top of this page and take the steps in order.