Verified65% confidenceFactExact time
Specification Gaming is a core challenge in AI safety, referring to a model satisfying task requirements literally while violating the designer's true intent.
3
Sources
65%
Confidence
Long-term
Relevance
7/8/2026
First Seen
Sources
GPT-5.6发布：Soul/Terra/Luna三层布局，编排能力训入模型
bilibili杨博士说AI6/29/2026
Related Claims
UnverifiedIn 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.82% similarUnverifiedAI capability evaluation faces the risk of "eval gaming," whereby model developers may intentionally or unintentionally select evaluation scenarios that favor their own models.79% similarUnverified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）79% similarUnverifiedIn the field of AI safety, the deviation between a system's actual optimization objective and the designer's true intent is collectively referred to as 'goal misalignment'.75% similarUnverified贪吃蛇因覆盖游戏循环、状态机设计、碰撞检测和事件系统四个核心软件工程模式，已成为 AI 编程能力评估的非正式基准任务74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/257522API
curl https://kongchang.com/api/v1/knowledge/claims/257522MCP
get_claim(id=257522)