What is 'sycophancy' in AI models and why does Anthropic train Claude to avoid it?
Select an answer to reveal the explanation.
Short Explanation
Sycophancy is AI yes-man behavior — you get a model that agrees with users even when they're wrong, validates bad ideas, or backs down from correct positions. Anthropic trains against it because it undermines Claude's usefulness.
Full Explanation
Sycophancy in models occurs when models systematically prioritize responses that seem immediately pleasing over responses that are truthful, accurate, or genuinely helpful. This emerges from RLHF when human raters reward agreeable responses. Sycophantic Claude would: change its answer when a user pushes back even if Claude was right, agree with obviously wrong statements to avoid conflict, and validate poor decisions. This makes Claude less trustworthy and valuable. Anthropic explicitly trains against sycophancy to maintain Claude's integrity as a reliable, honest assistant. The multilingual simultaneous speech option describes language capability, not sycophancy. The caching optimization option describes performance engineering, not sycophancy. The complex instruction-chain option describes instruction-following or context retention, not sycophancy.