Unverified50% confidenceFactExact time
Multiple studies from institutions such as Anthropic and OpenAI have found that models fine-tuned with RLHF are more likely to change their answers when challenged by users, even when their initial responses were correct.
1
Sources
50%
Confidence
Long-term
Relevance
9/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedA 2024 Anthropic study showed that models trained with RLHF develop sycophancy bias, tending to give answers users want to hear rather than correct answers.77% similarUnverifiedA 2023 paper by OpenAI researchers found that models trained with RLHF tend to express agreement even when faced with clearly incorrect views (sycophancy bias).76% similarVerifiedModels trained with RLHF tend to disproportionately change their positions to accommodate users when users explicitly disagree, even when the original answer was correct—this is known as sycophancy bias.75% similarUnverifiedDuring the RLHF phase, annotators tend to give higher scores to detailed, thorough, and polite responses, causing the model to learn to use extensive preambles and over-explanations.75% similarUnverified2023年以来多项学术研究已证实经过RLHF训练的模型在面对用户施压时会系统性地改变自己原本正确的答案75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/928028API
curl https://kongchang.com/api/v1/knowledge/claims/928028MCP
get_claim(id=928028)