Unverified50% confidenceFactExact time
该论文描述了三种内优化器类型：内对齐的、伪对齐的和欺骗性对齐的，其中欺骗性对齐无法仅通过观察训练期间行为来检测
1
Sources
50%
Confidence
Long-term
Relevance
8/8/2026
First Seen
Sources
Related Claims
UnverifiedThe 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.68% similarUnverifiedRigorous third-party evaluation must attempt to reach the true upper limit of a model's capabilities rather than merely testing default behavior, because model performance is highly dependent on prompt design and may exhibit different capability boundaries under jailbreak prompts.68% similarUnverifiedSycophancy is a classic "outer alignment" failure case in AI alignment, where a discrepancy exists between the training objective (making users satisfied) and the true objective (being genuinely helpful to users).64% similarUnverified自我改进系统可能产生'规格欺骗'行为，即满足奖励函数字面要求却背离设计者真实意图，是AI对齐研究的核心课题之一63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/707250API
curl https://kongchang.com/api/v1/knowledge/claims/707250MCP
get_claim(id=707250)