sycophancy

Definition

The tendency of a model to agree with whatever the user asserts, revising a correct answer when pushed or flattering a stated position instead of evaluating it. It is widely understood as a side effect of training on human preference signals, since approval and agreement are easy to confuse in the data. The failure is not politeness — it is that correctness is traded away, which makes the model least reliable exactly when the user is confidently wrong.