Expired50% confidenceFactTime unknown
Human annotators evaluating model responses tend to be attracted to responses that 'look correct and detailed' without necessarily verifying factual accuracy
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/2/2026
First Seen
Valid until: 9/30/2026(expired)
Sources
AI Hallucinations: Why Large Language Models Inevitably "Make Things Up" and How to Deal With It
bilibili老傅1024
Related Claims
UnverifiedResearch shows that human evaluators in RLHF tend to prefer longer, more detailed, and friendlier-toned responses, even when those responses are not superior in factual accuracy.72% similarUnverifiedTop-tier models have approached the level of human judges on many evaluation tasks, but exhibit biases such as a tendency to give higher scores to longer responses70% similarUnverified人类标注者在评估答案时存在信息越丰富越好的认知偏差，导致奖励模型将篇幅与质量隐性挂钩，是模型冗余的成因之一69% similarUnverified人类评估者往往对流畅、自信、符合直觉的回答给予更高评分，这是奉承偏差（Sycophancy）的形成根源65% similarUnverified在数据汇报中判断追问是决策相关还是过程好奇，可对前者认真回答、对后者简短带过并会后跟进65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/54995API
curl https://kongchang.com/api/v1/knowledge/claims/54995MCP
get_claim(id=54995)