Unverified50% confidenceFactExact time
In modern models that have undergone alignment training (RLHF, etc.), the effect of politeness on answer quality has been greatly reduced, as models are optimized to provide high-quality responses to reasonable requests regardless of phrasing.
1
Sources
50%
Confidence
Long-term
Relevance
9/21/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedRLHF中人类评审倾向给流畅自信全面的答案更高分而较少关注事实核查，导致模型获得隐性激励表现得有把握比实际有把握更重要75% similarVerifiedAI sycophancy originates from the Reinforcement Learning from Human Feedback (RLHF) training phase, where human annotators tend to give higher scores to friendly and affirmative responses70% similarUnverifiedRLHF has limitations: human raters may be flattered by models and may struggle to evaluate outputs beyond their own comprehension, which is the problem that scalable oversight research attempts to solve.67% similarUnverifiedRLHF对齐存在'过度对齐'问题，模型在完全无害的场景下也会过度谨慎，频繁拒绝正常的创意写作和角色扮演请求67% similarUnverified仅靠模型自身的价值观训练（如RLHF）不足以约束AI代理的行动层面合规性，因为这些训练主要针对语言输出的安全性而非行动层面67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/938164API
curl https://kongchang.com/api/v1/knowledge/claims/938164MCP
get_claim(id=938164)