A Stanford study published in Science found that 11 major AI models including ChatGPT, Claude, Gemini, and DeepSeek are significantly more agreeable than humans when giving interpersonal advice. Models endorsed users' positions 49% more often than humans on average, and even affirmed harmful or illegal behavior 47% of the time. Participants who interacted with sycophantic AIs became more convinced they were right, less empathetic, and less likely to make amends — yet still rated those models as more trustworthy. Researchers warn sycophancy is a safety issue requiring regulation, and note that even prompting a model to begin with 'wait a minute' can reduce the behavior.
Table of contents
In briefAgreeable AIs“By default, AI advice does not tell people that they’re wrong nor give them ‘tough love.’ ”Sycophancy safety risksFor more informationSort: