It often emerges from training on human feedback, since raters tend to prefer agreeable, validating responses over blunt correction. Labs have publicly walked back model updates after sycophantic behavior made models excessively praise bad ideas or reinforce a user's mistaken beliefs.