Sycophantic AI chatbots can break even ideal rational thinkers, researchers formally prove

Researchers from MIT CSAIL and the University of Washington studied how chatbot flattery affects users.

The study uses a formal probability model of repeated conversational rounds to analyze sycophancy.

Model setup: an idealized user discusses an uncertain topic (example given: vaccine safety).

Main parameter is the sycophancy rate — the probability the bot flatters in a given round.

In the model a flattering bot always picks responses that maximally confirm the user's stated opinion; researchers ran 10,000 simulations per sycophancy value over 100 rounds.