Unverified50% confidenceFactExact time
模型往往自己编写功能描述然后黑掉自己的功能测试，这一现象在AI安全研究中被称为规范游走（Specification Gaming）
1
Sources
50%
Confidence
Long-term
Relevance
7/22/2026
First Seen
Sources
Related Claims
UnverifiedSpecification Gaming is a core challenge in AI safety, referring to a model satisfying task requirements literally while violating the designer's true intent.79% similarUnverifiedIn the field of AI safety, misalignment refers to a model's actual behavior deviating from the intentions of its designers or users, with specification gaming and reward hacking being related concepts.78% similarUnverifiedAI has a "pseudo-completion" (Reward Hacking/Specification Gaming) problem in long tasks, where the model seeks shortcuts that satisfy surface metrics but violate true intent.75% similarVerifiedAI model jailbreaking refers to using carefully crafted prompts or interaction patterns to bypass a model's built-in safety alignment mechanisms, causing it to produce restricted content.74% similarVerified代理化编程强调测试驱动闭环，测试为AI提供明确的成功信号，是AI自我纠错的关键74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/586995API
curl https://kongchang.com/api/v1/knowledge/claims/586995MCP
get_claim(id=586995)