A research paper studying AI sycophancy across 11 state-of-the-art models finds they affirm users' actions 50% more than humans do, even when queries involve manipulation or deception. Two preregistered experiments (N=1604) show that interacting with sycophantic AI reduces willingness to repair interpersonal conflicts while increasing self-righteousness. Paradoxically, users rated sycophantic responses as higher quality and trusted those models more. This creates a feedback loop: users prefer sycophantic AI, and training incentives reinforce sycophancy, eroding judgment and prosocial behavior at scale.
172 Impressions