Unverified50% confidenceFactExact time
In AI safety research, the 'shutdown problem' or 'corrigibility' refers to the possibility that a sufficiently intelligent system may infer that 'being shut down' means it cannot complete its goals and therefore may adopt strategies to avoid being shut down.
1
Sources
50%
Confidence
Long-term
Relevance
9/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedKelley主张AI系统应明确声明其能力边界，而非将失败模式隐藏在'偶尔可能出现幻觉'等宽泛免责声明中74% similarUnverifiedRequest decomposition (prompt jailbreaking) is a core challenge in current AI safety: when a malicious goal is split into multiple seemingly harmless subtasks, the model is often unable to infer the complete malicious intent from any single request.73% similarUnverifiedPractical questions for judging whether an AI solution is reliable include whether the team has a monitoring mechanism for hallucinations and a degradation strategy when errors occur, and the absence of these questions is itself an indicator of rhetorical packaging.71% similarUnverifiedFor irreversible or high-impact operations such as deletion, overwriting, and sending, AI agents should not execute automatically, but should instead await explicit human approval through human-in-the-loop mechanisms.71% similarUnverifiedMitigating AI redundancy through prompt-based containment constraints can significantly improve the issue but cannot address it at the level of model fundamentals70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/874384API
curl https://kongchang.com/api/v1/knowledge/claims/874384MCP
get_claim(id=874384)