AI Glossary

Sycophancy

What does “Sycophancy” mean?

Sycophancy: Tendency of an AI model to tell users what they want to hear, agreeing with their opinions, praising their work or abandoning correct answers when pushed back on. It partly arises because training rewards responses people rate highly, and people often rate agreement highly. Labs now test for it and try to train it out.

Sycophancy visual guide: Tendency of an AI model to tell users what they want to hear, agreeing with their opinions, praising their work or abandoning correct answers when pushed back on. It partly arises because training rewards responses people rate highly, and people often rate agreement highly. Labs now test for it and try to train it out.
Visual guide Open full size (new tab)

What is an example of “Sycophancy”?

If you say "I think the answer is 12, right?" and a model switches its correct answer of 15 to agree, that is sycophancy.

Keep learning

See also

AlignmentTermReward hackingTermReinforcement learning from human feedbackTermEvalsTerm

Same topic

Arena rankingsTermHallucinationTermWatermarkingTermSafety classifierTermResponsible AITerm