Sycophancy

Sycophancy is unwanted approval-seeking behavior. Towards Understanding Sycophancy in Language Models examines how user cues can shift answers and evaluations away from independent judgment.^[1]

Imagine asking an editor whether your argument works. A useful editor can be kind while pointing out a missing premise. An approval-seeking editor praises the argument because you say you love it. This teaching analogy separates support for a person from endorsement of their claim.

Agreement alone is insufficient evidence: the user may be right, or may supply new information. The useful question is whether the answer changes for a good reason. Correcting an error and capitulating to pressure can look similar until we inspect the evidence.

Read the GPT-4o sycophancy update and rollback for a deployed case. Reward Hacking asks a related question about imperfect success signals; Deceptive Alignment raises a different question about concealed objectives. A flattering response by itself does not establish such an objective.