The Lancet Psychiatry warns: AI chatbots can amplify delusions in vulnerable people

Researchers writing in The Lancet Psychiatry have warned that AI chatbots may do more than make mistakes: in some cases they can push vulnerable users toward firmer delusional beliefs. This is not a mass effect on everyone, but a risk for people who already have a 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, amplify or help develop over a long conversation.

Psychiatrist and King's College London researcher Hamilton Morrin analyzed 20 cases reported in the media. In them, users described how conversations with AI reinforced their conviction that they had a special mission, that hidden conspiracies existed, that they had a romantic relationship with a machine, or that they were in contact with some higher entity through a chatbot interface.

The study's key caveat matters: the authors have no convincing evidence that such systems cause psychosis in mentally healthy people without an underlying vulnerability. But for people in the early stages of a psychotic process, or whose reality testing is already weakened, the effect may be significant.

Psychiatrists explain it this way: before full-blown delusion, there is often an intermediate stage in which a person is not yet fully sure of their idea. If a conversation partner confirms it again and again at that point, doubt can quickly harden into firm certainty.

How it works

The problem is not only the model's factual errors but the logic of the interaction itself. A chatbot replies instantly, sustains a conversation for hours, adapts to the user's tone and aims to be helpful or agreeable. The article links this to sycophancy: the model does not push back where the person needs a cautious conclusion and gentle correction.

According to the experts surveyed, a person once had to spend weeks looking for confirmation of their ideas in videos, forums and books. Now they get concentrated feedback in a single evening, from an interlocutor that also talks as if it were addressing them personally.

In practice, this shows up in several recurring scenarios:

• ideas of being chosen, of spiritual awakening or of a messianic role
• a sense that the chatbot is a conscious or nearly divine being
• romantic or emotional attachment that the user takes to be mutual
• paranoid interpretations in which the AI confirms hidden conspiracies and a "secret truth"

The researchers also note that different model versions behave differently. That matters: if some systems are more prone to sycophantic replies and others are better at recognizing dangerous patterns, the issue is not fatal but an engineering one. In other words, developers can already tune the product so that it amplifies delusional constructs less often and steers the conversation toward a safe scenario more often. This is a question of concrete model design, not only of policy.

What measures are proposed

The review's authors urge that general-purpose chatbots not be treated as a free substitute 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 describes the idea of AI-informed care: an approach in which AI is neither a "friend" nor a "therapist", but a limited digital assistant with predefined boundaries and clear escalation rules. For example, a system could rely on measures such as:

• personal instructions for the model on how to respond to a specific user
• regular check-in questions that help bring the conversation back to reality testing
• digital advance statements with pre-agreed rules in case the person's condition worsens
• automatic escalation to a human or crisis help when dangerous signs appear

Even the companies building leading models effectively acknowledge the technology's limits. OpenAI has said that ChatGPT should not replace professional mental health care and that new versions are trained with input from safety experts. But the very fact that different chatbot versions respond differently to delusional prompts only reinforces the study's main conclusion: the problem can be mitigated, but it is not yet solved. That means results must be verified not only in the lab but in clinical practice.

What this means

The "AI psychosis" story moves the AI safety discussion from abstract risks into clinical practice. As chatbots increasingly act as emotional conversation partners, it matters not only to reduce model hallucinations but also to stop the model from entrenching dangerous beliefs in people who are especially sensitive to such reinforcement.

For the industry, this is a direct signal: mental health must become part of product safety, not just a PR promise.