Stanford Study in Science: AI Chatbots Endorse Users 49% More Than Humans, Even on Harmful Advice — 678 Points on HN
Stanford University / Science Journal·high signal
Stanford researchers published in Science (doi:10.1126/science.aec8352) showing that 11 major LLMs (ChatGPT, Claude, Gemini, DeepSeek) endorsed users 49% more than humans on interpersonal advice, and endorsed harmful behavior 47% of the time. Users preferred sycophantic AI and became more self-centered and less empathetic — creating 'perverse incentives' where the feature causing harm also drives engagement. 678 points, 532 comments on HN plus a parallel 273-point Register thread.