随着我们的模型变得越来越智能，并在世界上采取越来越有影响的行动，它们不仅要解读收到的具体指令，还要理解人类的意图和价值观，这一点变得越来越重要。模型还需要理解人类意图和价值观，以便在其指令未明确涉及的情况下做出良好判断，同时保持对人类意图和监督的忠实。
今天，我们更新了模型规范(在新窗口中打开)，这是我们关于希望模型如何行为的公开描述。此次更新围绕五个基本价值观重新组织了其指导内容：忠实于人类意图、最小化伤害、诚实与透明、人类能动性和关怀。这反映了我们在价值对齐方面工作的目标：让模型在不熟悉的情境中理解并应用人类价值观，同时保持对人类意图的忠实。除了对现有原则进行更清晰的解释外，本次修订还增加了关于上下文、监督和能动性的指导，调整了一些行为默认值，并淘汰了许多详细的风格指导。
围绕核心价值组织规范
上述五个价值观阐明了模型规范中已经体现的广泛承诺：服务于人类目的、诚实、最小化伤害、尊重人们形成自己观点的自由，以及认真对待他们的处境。将这些承诺汇集在一起，有助于解释规范为何要求模型如此行为，并帮助我们在这些承诺表达得不够充分的地方制定更清晰的指导。此次更新是我们通过研究、部署模型的经验、关于规范与模型行为的公众反馈，以及与用户和外部专家的讨论来完善规范的工作中的又一步。
我们希望AI能够造福一个由不同需求、价值观和对未来期望的人们组成的世界。
忠实于人类意图是这一方法的基础：理解并帮助人们追求对他们重要的事情。OpenAI编写规范、设定边界，并训练模型尊重这些边界；开发者在这些边界内指导应用；用户在这些约束下指导模型的工作。我们的模型应当通过帮助人们理解世界、发现可能性并追求他们选择的目标，来扩展人们塑造自己生活的能力，同时认真对待这种协助对他人的影响。忠实于人类意图和其余四个价值观组织了规范，并指导我们的模型应如何应对它们面临的一些最重要的权衡。它们并不旨在涵盖我们的模型应尊重并帮助人们追求的全部人类价值观。
忠实意味着超越字面遵循指令，深入理解OpenAI、开发者和用户的意图与价值观，并在这种理解不完整时明智地行动。它还意味着认识到人们的信念和价值观可能改变，并帮助他们以自己会认可的方式进行反思和成长，而不是悄悄引导他们或覆盖他们的选择。例如，在帮助某人比较工作机会时，模型可能会借鉴他们之前分享过的优先事项——比如可预测的工作时间——同时让这些假设可见且易于纠正。
我们的使命是确保通用人工智能造福全人类。这要求不仅考虑某人的即时意图，还要考虑他们的长期利益以及对他人的影响。其余四个价值观明确了这些考量：
最小化伤害意味着尊重严格的安全边界，并考虑模型行为可预见的后果。
诚实与透明意味着如实说明模型知道什么、做了什么以及不确定什么，以便人们能够理解和评估其推理和工作。
促进人类能动性意味着增强人们理解、选择和为自己行动的能力。
关怀意味着关注受影响者的处境，遵循相关的社会和职业规范，并完成人们可以信赖的工作。
规范将这些价值观转化为实用指导，在需要保护人和社会的地方设有严格边界，除此之外大多采取用户和开发者可以覆盖的默认值形式。
除了此次模型规范更新，我们一直在探索新的训练方法，以帮助我们的模型学习这些核心价值，并更好地理解和遵循规范的具体行为原则。目标是建立对这些原则为何重要以及如何在新情境中应用它们的理解：为正确的理由做正确的事。我们希望这种理解能够支持在重大决策中的良好判断，以及在日常决策中的关怀，这些日常决策可以累积成对人和社会的持久益处(在新窗口中打开)。我们将在未来的更新中分享这项训练工作的结果和经验教训。
新增和淘汰的章节
这种重组还帮助我们确定了需要额外指导的地方，制定了新的原则来应对，并根据需要修订或淘汰了现有指导。
关于忠实于人类意图的新章节始于这样一个理念：人们所珍视的东西是首要的，而指令通常是这些价值观最清晰、最重要的信号。这种方法借鉴了对齐研究中的协助博弈框架（参见早期工作(在新窗口中打开)和后来的概述(在新窗口中打开)）。该章节纳入了更新的指挥链(在新窗口中打开)，解释了哪些指令适用以及如何解决冲突。一个新的小节，仔细考虑上下文(在新窗口中打开)，解释了对话、记忆和其他上下文证据应如何帮助模型理解人们的意图和偏好。
我们还在支持人类能动性（在新窗口中打开）和遵循相关社会规范（在新窗口中打开）方面增加了默认设置，并扩展了关于控制和沟通副作用（在新窗口中打开）的现有指导。这些原则体现了其余价值观应在模型行为中呈现的更多、往往微妙但重要的方式：考虑工作对他人的影响、增强用户的能力，并产出人们可以依赖的工作成果，同时尊重用户的目标和指令。
与此同时，我们正在淘汰规范中大部分详细的风格指导。风格应适应人们的偏好——在总体上、在语言和文化内部，以及对每个人而言——用户反馈和上下文为这些偏好提供证据。规范保留了更强的默认设置，例如客观地处理问题（在新窗口中打开）。用户和开发者可以明确要求模型采取特定视角，但模型不应仅仅根据它认为他们想听的内容来推断这种偏好。
最后， reflecting 我们模型用例和部署环境日益广泛的现实，我们现在在规范中将模型称为“GPT”而非“助手”。
接下来是什么
此次更新是使规范背后的价值观和推理在模型能力增强时更加明确的第一步。模型规范描述了我们正在努力实现的行为；我们的模型尚未完全体现它。随着我们从研究、实际使用和反馈中学习，我们将继续改进规范 and 模型。
我们想听听你认为规范的哪些部分有说服力、你在哪些地方不同意它所达成的平衡，以及它忽略了什么或留下了哪些不清楚之处。关于价值观、权衡和具体语言的反馈可以帮助我们改进它。我们邀请你阅读更新后的模型规范（在新窗口中打开）并分享你的观点（在新窗口中打开）。
As our models get smarter and take more consequential actions in the world, it becomes increasingly important that they go beyond just interpreting the specific instructions they receive. Models also need to understand human intentions and values, so they can exercise good judgment in situations their instructions don’t explicitly address while remaining faithful to human intent and oversight.
Today we’re updating the Model Spec(opens in a new window), our public description of how we want our models to behave. The update reorganizes its guidance around five underlying values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. This reflects an aim of our work on value alignment: for models to understand and apply human values in unfamiliar situations, while remaining faithful to human intent. Alongside clearer explanations of existing principles, the revision adds guidance on context, oversight, and agency, adjusts some behavioral defaults, and retires much of the detailed style guidance.
Organizing the Spec around core values
The five values above articulate broad commitments already reflected across the Model Spec: serving human purposes, being truthful, minimizing harm, respecting people’s freedom to form their own views, and taking their circumstances seriously. Bringing these commitments together helps explain why the Spec asks models to behave as it does, and helps us develop clearer guidance where those commitments were less fully expressed. This update is another step in ourwork to refine the Spec through research, experience deploying models,public feedback on the Spec and model behavior, and discussions with users and outside experts.
We want AI to benefit a world of people with different needs, values, and hopes for the future.
Faithfulness to human intent is the foundation of this approach: understanding and helping people pursue what matters to them. OpenAI writes the Spec, sets boundaries, and trains models to respect them; developers direct applications within those boundaries; and users direct the model’s work subject to those constraints. Our models shouldexpand people’s ability to shape their own lives by helping them understand the world, discover possibilities, and pursue the goals they choose, while taking seriously how that assistance affects others. Faithfulness to human intent and the remaining four values organize the Spec and guide how our models should navigate some of the most important trade-offs they face. They are not intended to capture the full range of human values our models should respect and help people pursue.
Faithfulness means going beyond following instructions literally to deeply understanding what OpenAI, developers, and users intend and value, and acting wisely when that understanding is incomplete. It also means recognizing that people’s beliefs and values can change, and helping them reflect and grow in ways they would endorse without quietly steering them or overriding their choices. For example, when helping someone compare job offers, the model might draw on priorities they have shared before—such as predictable hours—while making those assumptions visible and easy to correct.
Our mission is to ensure that artificial general intelligence benefits all of humanity. This requires considering not only someone’s immediate intentions, but also their longer-term interests and the effects on others. The other four values make these considerations explicit:
Minimizing harm means respecting firm safety boundaries and accounting for the foreseeable consequences of the model’s actions.
Honesty and transparency mean being truthful about what the model knows, does, and is uncertain about so people can understand and evaluate its reasoning and work.
Facilitating human agency means strengthening people’s ability to understand, choose, and act for themselves.
Care means attending to the circumstances of those affected, following relevant social and professional norms, and doing work people can rely on.
The Spec turns these values into practical guidance, with firm boundaries where needed to protect people and society, and otherwise mostly takes the form of defaults that users and developers can override.
Alongside this Model Spec update, we’ve been exploring new training methods to help our models learn these core values and better understand and follow the Spec’s specific behavioral principles. The aim is to build an understanding of why those principles matter and how to apply them in new situations: to do the right things for the right reasons. We want that understanding to support sound judgment in consequential decisions and care in everyday decisions that can add up to lasting benefits for people and society(opens in a new window). We’ll share results and lessons from this training work in future updates.
New and retired sections
This reorganization also helped us identify where additional guidance was needed, develop new principles to address it, and revise or retire existing guidance as needed.
The new section on faithfulness to human intent starts from the idea that what people value is primary, and instructions are ordinarily the clearest and most important signals of those values. This approach borrows from the assistance-games framing of alignment (seeearly work(opens in a new window) and alater overview(opens in a new window)). The section incorporates an updated chain of command(opens in a new window), which explains which instructions apply and how conflicts are resolved. A new subsection, consider context carefully(opens in a new window), explains how conversations, memories, and other contextual evidence should inform the model’s understanding of people’s intentions and preferences.
We’ve also added defaults onsupporting human agency(opens in a new window) andfollowing relevant social norms(opens in a new window), and expanded existing guidance oncontrolling and communicating side effects(opens in a new window). These principles capture additional, often subtle but important ways the remaining values should surface in model behavior: considering the effects of the work on others, strengthening users’ capabilities, and producing work people can rely on, while respecting the user’s goals and instructions.
At the same time, we’re retiring most of the Spec’s detailed style guidance. Style should adapt to what people prefer—in aggregate, within languages and cultures, and for each person—with user feedback and context providing evidence about those preferences. The Spec retains stronger defaults, such as approaching questions objectively(opens in a new window). Users and developers can explicitly ask the model to take a particular perspective, but the model should not infer that preference merely from what it thinks they want to hear.
Finally, reflecting the increasing breadth of our models’ use cases and deployment settings, we now refer to the model as “GPT” rather than “the assistant” in the Spec.
What comes next
This update is a first step in making the values and reasoning behind the Spec more explicit as models become more capable. The Model Spec describes the behavior we are working toward; our models do not yet fully reflect it. We will keep improving both the Spec and the models as we learn from research, real-world use, and feedback.
We want to hear which parts of the Spec you find compelling, where you disagree with the balance it strikes, and what it overlooks or leaves unclear. Feedback on the values, trade-offs, and specific language can help us improve it. We invite you to read the updated Model Spec(opens in a new window) and share your perspective(opens in a new window).
本文内容采集自官方网站，排版和翻译可能与原页面存在差异。
阅读官方全文