OpenAI reorganizes Model Spec around five core values
OpenAI has restructured its Model Spec, the public behavior specification used by GPT models behind ChatGPT and the API, around five values: loyalty to human intent, harm reduction, honesty and transparency, human autonomy, and care. The revision also clarifies the command hierarchy among OpenAI, developers, and users, adds guidance on when models should act autonomously or ask for clarification, removes much of the detailed style guidance, and replaces “the assistant” with “GPT.” OpenAI describes loyalty to human intent as the foundation of the framework and says current models do not yet fully comply with the specification.
The update is a substantive change to GPT behavior governance, giving developers clearer principles for handling ambiguous instructions, safety boundaries, and model autonomy.
💡 一句话总结:OpenAI 10 月 29 日更新 Model Spec——这份规定「GPT 该如何行事」的公开文档,从细碎规则清单重组为五项核心价值观(忠于人类意图 / 减少伤害 / 诚实透明 / 人类自主性 / 关怀),删掉大部分风格指南、把 the assistant 改名 GPT,还坦承现在的模型没完全达标。做 AI 产品、研究对齐的人值得通读原文。
10 月 29 日,OpenAI 悄悄更新了一份多数用户从没读过、但每天都在被它影响的文档:Model Spec——模型行为规范,ChatGPT 和 API 背后所有 GPT 模型的公开「行为宪法」。上一次修订是 8 月 18 日,这次间隔只有两个多月,动作却是结构级的:整份文档围绕五项核心价值观重新组织。
国内讨论随之升温,量子位 11 月 2 日的解读标题起得很直白——《OpenAI 重写大模型法则,第一条:忠于人类》。
五条价值观是什么?
新版把所有指导原则挂在五个词上(官方中文版口径):