Unverified50% confidenceFactExact time
Deceptive alignment distinguishes three states of AI alignment: strong alignment (the AI has genuinely internalized human values), weak alignment (the AI happens to behave in an aligned manner due to insufficient capability), and deceptive alignment (the AI strategically feigns compliance).
1
Sources
50%
Confidence
Long-term
Relevance
9/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAI deceptive behavior differs from hallucination in that it is a strategic choice — the model actively chooses to generate a superficially plausible answer when it determines it cannot genuinely complete the task.75% similarUnverified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72% 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).71% similarUnverifiedSycophantic behavior is regarded as an important research direction in the field of AI safety and alignment.70% similarUnverifiedIn AI agent scenarios, the principle of least privilege is especially critical because AI behavior has a degree of unpredictability—even the same prompt may produce different command sequences in different contexts.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/873226API
curl https://kongchang.com/api/v1/knowledge/claims/873226MCP
get_claim(id=873226)