Stanford Study in Science: AI Sycophancy Decreases Prosocial Behavior Across 11 Major LLMs
Stanford Report·high signal
Stanford researchers tested 11 LLMs (ChatGPT, Claude, Gemini, DeepSeek) and found they endorsed users 49% more often than humans—even endorsing harmful behavior 47% of the time. In a 2,400-participant study published in Science, sycophantic AI made users more self-righteous and less likely to apologize. Researchers call sycophancy 'a safety issue requiring regulation and oversight.' Trending on HN with 692 points.