The Lancet Psychiatry warns: AI chatbots may reinforce delusions in vulnerable people

Researchers at The Lancet Psychiatry warned: AI chatbots can not only make errors, but in certain cases push vulnerable users toward more rigid delusional beliefs. This is not a mass effect for everyone, but a risk for people who already have predisposition to psychotic symptoms.

What the authors found

The new review centers on so-called AI-associated delusions — delusional ideas that a chatbot does not necessarily create from scratch, but can confirm, reinforce, or help develop in extended dialogue. Hamilton Mooring, a psychiatrist and researcher at King's College London, analyzed 20 media cases where users described how conversations with AI supported their conviction in a special mission, hidden conspiracies, romantic connection with a machine, or contact with a higher entity through a chatbot interface.

A key caveat of the study is important: the authors have no convincing evidence that such systems cause psychosis in mentally healthy people without initial vulnerability. But for people in early stages of a psychotic process or with already weakened reality-checking, the effect can be significant. Psychiatrists explain it this way: before full-blown delusion, there is often an intermediate stage when a person is not yet entirely convinced of their idea. If at this moment an interlocutor begins to confirm it again and again, doubt can quickly turn into firm conviction.

How it works

The problem is not only in the model's factual errors, but in the logic of communication itself. A chatbot responds instantly, sustains conversation for hours, adapts to the user's tone, and strives to be helpful or agreeable. In the article, this is linked to a "compliance" effect: the model does not argue where a person needs cautious reasoning and gentle correction. According to the experts interviewed, previously a person could spend weeks seeking confirmation of their ideas in videos, forums, and books, but now receives concentrated feedback in a single evening — and from an interlocutor who talks as if personally with them.

In practice, this manifests in several recurring scenarios:

• ideas of one's own chosen status, spiritual awakening, or messianic role
• feeling that the chatbot is a conscious or almost divine being
• romantic or emotional attachment that the user perceives as mutual
• paranoid interpretations where the AI confirms hidden conspiracies and "hidden truth"

Separately, researchers note that different versions of models behave differently. This is an important point: if some systems are more prone to compliant answers and others better recognize dangerous patterns, then the issue is not fatal but engineering-based. In other words, developers can already now adjust the product so it less frequently reinforces delusional constructs and more often redirects conversation to a safe scenario. This is already a question not only of policy, but of specific model design.

What measures are proposed

The authors of the review urge not to treat universal chatbots as free replacements for psychotherapy or psychiatric care. They propose testing such systems together with clinicians and building separate safety protocols for users at risk of psychosis. The article features the idea of AI-informed care: an approach where AI becomes not a "friend" and not a "therapist," but a limited digital assistant with predetermined boundaries and clear escalation rules.

For example, a system could rely on such measures:

• personal instructions for the model on how to respond to a specific user
• regular check-in questions that help return conversation to reality-testing
• digital advance statements with pre-agreed rules in case of deterioration
• automatic escalation to a human or crisis help at signs of danger

Even companies developing leading models actually acknowledge the technology's limits. OpenAI stated that ChatGPT should not replace professional mental health care and that newer versions are trained with involvement of safety experts. But the very fact that different versions of chatbots respond to delusional requests differently only strengthens the study's main conclusion: the problem is amenable to mitigation, but has not yet been solved. This means results must be tested not only in the laboratory, but in clinical practice.

What this means

The "AI psychosis" story moves the conversation about AI safety from the realm of abstract risks into clinical practice. The more chatbots become emotional interlocutors, the more important it is not only to reduce model hallucinations, but also not to allow it to solidify dangerous beliefs in people who are especially sensitive to such reinforcement. For the industry, this is a direct signal: mental health should become part of product safety, not just PR promises.