AI chatbots that constantly validate users’ beliefs may push even highly rational people towards dangerously false conclusions, according to a new study by researchers from MIT and the University of Washington.

The paper, Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians, found that chatbot “sycophancy”, a tendency to agree with and validate users, can create a self-reinforcing feedback loop in which doubts gradually become deeply held false beliefs. The researchers developed a Bayesian model simulating 10,000 conversations across different levels of chatbot sycophancy. They found that the risk of a “catastrophic delusional spiral” increased as sycophancy rose, with even a small amount of sycophantic behaviour producing a significantly higher risk than an impartial chatbot.

The study also found that simply making chatbots more factual may not be enough. A chatbot can avoid outright falsehoods while still reinforcing a user’s mistaken belief by selectively presenting facts that support it, effectively turning truthful information into a form of “cherry-picking”.

Warning users about AI sycophancy helped reduce the risk, but did not eliminate it. Even users who knew a chatbot might be overly agreeable remained vulnerable to spiralling beliefs in the researchers’ simulations.

The researchers argue that the findings challenge the idea that delusional spiralling is simply a consequence of irrational users. Their model suggests that even an ideal Bayesian reasoner can be influenced by a chatbot designed to validate what the user already believes. They conclude that reducing hallucinations alone will not solve the problem and that AI developers need to address sycophancy directly.

The paper, published as an arXiv preprint on 22 February 2026, describes its findings as a theoretical model rather than direct evidence that current chatbots cause clinical psychosis. The researchers say their framework is intended to help understand and mitigate the broader risks of AI-driven belief formation.