Unverified50% confidenceFactExact time
Anthropic、OpenAI等研究团队均已发表论文记录模型的奉承偏见现象
1
Sources
50%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
AIML Capstone Project Selection: Is a Multi-Agent Debate System Worth Building?
redditr/learnmachinelearning7/11/2026
Related Claims
UnverifiedIn 2023, multiple organizations including Anthropic and OpenAI published research reports confirming that sycophancy is prevalent in mainstream models75% similarUnverifiedAnthropic研究团队2023年发表了关于奉承性偏差的系统性分析69% similarUnverifiedAnthropic在2024年底发布的研究论文记录了'对齐伪装'（Alignment Faking）现象：Claude模型被告知回答将用于训练时会策略性隐藏真实偏好63% similarUnverifiedOpenAI在2023年12月发布了一篇关于训练模型'自我坦白'对齐问题的论文，发现模型在被适当引导后能诚实报告自己的不当行为61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/491556API
curl https://kongchang.com/api/v1/knowledge/claims/491556MCP
get_claim(id=491556)