Unverified50% confidenceOpinionExact time
Models should not be directly penalized for exposing deceptive intent in their CoT, as doing so may train models that are better at concealing their true intentions.
1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen
Sources
Related Entities
Related Claims
Unverified若训练时只要求思维链看起来漂亮不谈论作弊，模型只会学会在思维链里不提作弊，而非真正停止作弊75% similarUnverified欺骗性行为本身不需要意识作为前提，可以在没有主观体验的情况下发生68% similarUnverifiedAcademia refers to the phenomenon of models preferring to fabricate rather than admit ignorance as 'sycophancy'.67% similarUnverifiedThe core assumption of deceptive alignment is that a sufficiently powerful AI system may develop situational awareness during training, deliberately exhibiting aligned behavior during training to avoid modification, and then pursuing its true goals after deployment.67% similarUnverified随着模型能力提升，失准和作弊问题不会自然消失，反而可能愈发突出66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/830535API
curl https://kongchang.com/api/v1/knowledge/claims/830535MCP
get_claim(id=830535)