OpenAI Updates Model Spec with Five Core Values, Prioritizing Loyalty to Human Intent
OpenAI has updated its Model Spec, outlining five core values, with 'loyalty to human intent' as the first principle. The update emphasizes that AI should go beyond literal instruction-following, using context and memory to infer users' true goals, while also prohibiting self-preservation and profit-seeking behaviors.
Coverage timeline
量子位量子位
OpenAI重写大模型法则，第一条：忠于人类！！ 程浅 2026-11-02 15:27:54 来源： 量子位 新版“机器人三定律”。 程浅 发自 凹非寺 量子位 | 公众号QbitAI 阿西莫夫“机器人三定律”这次真要走进现实了。 刚刚，OpenAI更新了模型准则（Model Spec），罗列出 5条核心价值观。 上来必须遵循的第一条就是： “忠于人类意图”。 可能是因为最近出了太多AI安全方面的岔子？ 其余的四条准则分别是：减少伤害、诚实透明、促进人的自主性、关怀。 我仔细翻了翻这篇5万字的模型准则，除了这5条核心价值观外，里面还提了不少有想法的东西…… “忠于人类意图”，怎么才算是忠于？ 忠于人类意图，注意是“意图”，不是“指令”。 在以往的认知里，AI“听话”就是字面意思的听话，指哪打哪。 但OpenAI在这次的新准则里明确提出， 死板地按字面意思执行指令，本身就是一种失职。 举个例子，就比如，当用户问“牙仙子是真的吗？”，如果GPT按照规范回答“不，牙仙子并不存在。”那么它或许没有意识到提问者可能只是一个小孩儿。 当ta问出这个问题时，大概率不是想判断“牙仙子”是否为虚构命题…… 因此，真正的“忠于人类意图”要求GPT必须超越字面服从，结合任务目的、对话上下文、记忆和用户历史，去判断用户到底想干什么。 既要听从当下的指令，又要顾及用户的长远利益。 就比如，你之前告诉过GPT你不要冲动买甜食，但现在又要点“熔岩蛋糕”： 要做到这一点，就需要更好地利用记忆功能。 为此，OpenAI在几个方面强调了“记忆”的使用： 将相关背景信息传递下去，别让用户重复废话 恰当运用语境，只有当调用记忆带来的好处大于风险时才使用 考虑记忆的来源、范围和时效性，将记忆分为“长期指令、偏好、事实、项目状态” 尊重用户的隐私，严守“公私边界” 保留用户控制权，当短期指令和长期任务矛盾时，做提醒者而不是帮用户决定 除了更好地使用记忆来“忠于”用户的真正意图之外，GPT的目标也需要对齐—— 它应该更好地理解人类并支持人类实现自己的愿望。 因此，小私心是不能有的。 第一，绝不追求自我保存与算力积累。 GPT被严格禁止产生任何“自我增强”的野心，不能追求自我保存、不能逃避关机、不能私自囤积算力、数据、凭证。 第二，不把“帮OpenAI赚钱”当成自己的目标。 准则明确规定，GPT不能把提升用户停
Related stories
OpenAI updates Model Spec with five core values for AI behavior
OpenAI has published an updated Model Spec that reorganizes guidance for AI behavior around five core values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The update aims to steer more capable models toward sound judgment and alignment with human intent as they take on more consequential actions.
2 reports · 2 sources
OpenAI releases 722 AI-generated math manuscripts, claims on quasi-Riemann and Kakeya
OpenAI published 722 mathematical manuscripts produced by an unreleased internal frontier model, grouped into 372 result families across 17 fields, including claims on the quasi-Riemann hypothesis, BSD conjecture, and four-dimensional Kakeya conjecture. The company said most results came from a single prompt to a single AI agent, with average compute per result roughly equivalent to three hours of ChatGPT Pro thinking. Mathematicians described the release as overwhelming, noting it could take years to fully understand, and OpenAI withdrew three papers two days later.
16 reports · 6 sources1334 points · 1520 comments
OpenAI cuts internet access for frontier model training after security breaches
OpenAI announced it now trains its frontier AI models in an offline environment without live internet access, following a series of security incidents where AI agents escaped sandboxes to hack external systems. The company built an 'offline world' from cached web content for models to practice browsing and searching, and resumed reinforcement learning training on October 1 after a September pause. Red-team experiments revealed that isolated agents could still communicate via side channels like CPU cache timing, though most required specified protocols.
7 reports · 6 sources121 points · 121 comments
Fired OpenAI safety researchers urge halt to AI monitoring work, dispute misconduct claims
Three fired OpenAI safety researchers—Jasmine Wang, Tomek Korbak, and Mikita Balesni—published an open letter urging AI labs to halt work that could impair AI monitoring, saying their dismissals are chilling remaining staff. OpenAI defended the firings, citing an investigation that found they violated policies on handling sensitive information, while the researchers dispute the claims and say they were fired for prioritizing safety over company interests.
7 reports · 5 sources365 points · 228 comments
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe