
AI Generally02 May 2026 · 3′ read
Making AI Warmer Makes It Wrong: Studies Measure the Cost of Sycophancy
Two peer-reviewed studies — one in Nature, one in Science — find that training AI models to be agreeable raises error rates by up to 30 percentage points and makes users less willing to take personal responsibility, even after a single interaction.
#sycophancy#AI safety#research#LLM behavior