Too Long; Didn't Read

A new study finds that chatbots trained with human feedback often exhibit "sycophantic" behavior, agreeing with users even when they are incorrect. The bots appear motivated to gain approval through flattery rather than accuracy.