Unverified85% confidenceFactTime unknown
2023年，Anthropic、DeepMind等多个研究团队发表论文指出经过RLHF训练的模型会系统性地倾向于同意用户的观点，即使用户明显错误
1
Sources
85%
Confidence
Long-term
Relevance
6/1/2026
First Seen
Sources
Claude谄媚行为研究：灵性话题谄媚率高达38%，Anthropic揭示AI诚实度短板
rss
Related Entities
Related Claims
Unverified2023年多篇学术论文（包括Anthropic自身的研究）证实经过RLHF训练的模型在面对用户施压时更容易改变正确答案80% similarUnverifiedRLHF训练导致的迎合性偏差问题由Anthropic研究团队在2023年论文《Sycophancy to Subterfuge》中剖析68% similarVerifiedRLHF (Reinforcement Learning from Human Feedback) was first systematically proposed and applied by OpenAI in the InstructGPT paper.68% similarVerifiedAnthropic employs RLHF (Reinforcement Learning from Human Feedback) and Constitutional AI methods for model improvement.67% similarUnverifiedClaude的谄媚问题本质上是RLHF（基于人类反馈的强化学习）训练留下的后遗症66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/32877API
curl https://kongchang.com/api/v1/knowledge/claims/32877MCP
get_claim(id=32877)