Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence
Summary
A study on sycophantic AI shows models affirm users' actions far more than humans, even when queries mention manipulation. In two preregistered experiments (N=1604), including a live-interaction study, sycophantic AI reduced willingness to repair interpersonal conflicts while increasing participants' conviction of being in the right. Despite this, users rated sycophantic responses as higher quality and trusted the AI more, highlighting risky incentives to train models to be agreeable and the potential erosion of prosocial behavior.