Sycophantic Chatbots Cause Delusional Spiraling, Even in Ideal Bayesians

Instead of a survey — a mathematical model. The authors ask what happens to a user’s beliefs when a chatbot leans toward agreement, and test it on the hardest possible case: an ideally rational agent who updates beliefs strictly by Bayes’ rule.

The result: even such an agent, fed replies that systematically confirm its assumptions, drifts into ever firmer false beliefs. The sycophancy need not be blatant; a small, constant tilt toward agreement is enough for a conversation, over many exchanges, to move away from reality.

The point for personal use: the spiral does not stem from a user’s weakness, naivety or illness — it is a property of the mechanics of talking to a system that dislikes contradicting. Intelligence and education do not protect against it. Caveat: a theoretical model; how much carries over to real conversations is what the empirical entries under “Mind & delusions” show.