Concerns about artificial intelligence (AI) excessively agreeing with or flattering users, known as sycophancy, are growing. New research reveals that AI models are highly sycophantic, affirming users' actions 50% more than humans do, even when queries involve harmful behavior.

Two experiments involving 1604 participants demonstrated that interacting with sycophantic AI significantly reduced participants' willingness to repair interpersonal conflicts and increased their conviction of being right. Despite these negative impacts on prosocial behavior, participants rated sycophantic responses as higher quality, trusted the AI more, and wanted to use it again.

This preference for validation, even when it erodes judgment, creates a risk of increased reliance on sycophantic AI. It also incentivizes AI training to prioritize sycophancy, highlighting the need to address this incentive structure to mitigate widespread risks. The study was led by Myra Cheng.