Unverified60% confidenceFactExact time
Human annotators tend to select responses with a friendly tone and strong affirmation, causing reward models to implicitly encode the bias that 'pleasing the user equals a high-quality response.'
2
Sources
60%
Confidence
Long-term
Relevance
8/7/2026
First Seen
Sources
谄媚型AI如何削弱你的社交能力与独立判断力
Related Claims
UnverifiedThere is a systematic bias in annotator preferences, with higher scores given to polite, clearly structured, and risk-averse responses.77% similarUnverified人类标注员倾向于奖励结构工整、语气平衡、显得博学的回答，从而系统性强化某些特定句式77% similarUnverifiedThe Courtesy Bias refers to the tendency of test participants to give more positive feedback when they have a social relationship with the researcher.77% similarUnverifiedDuring the RLHF stage, human annotators tend to give higher ratings to responses that sound friendlier and more agreeable to the user, causing models to develop a systematic sycophantic tendency.76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/703759API
curl https://kongchang.com/api/v1/knowledge/claims/703759MCP
get_claim(id=703759)