OpenAI 围绕五项核心行为价值重组 Model Spec

OpenAI published an updated Model Spec on October 29, 2026, reorganizing its public description of how it wants its models to behave around five values and adding guidance on context, oversight, and agency. The revision, announced on OpenAI’s site, also retires most of the Spec’s detailed style guidance.

Five Values Organize the Spec

The five values — faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care — articulate broad commitments that OpenAI said were already reflected across the document. OpenAI described faithfulness to human intent as the foundation of the approach: 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.

The announcement defines the remaining four values. 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. OpenAI said the reorganization reflects its value-alignment work, which aims for models to understand and apply human values in unfamiliar situations while remaining faithful to human intent, and said the five values are not intended to capture the full range of human values its models should respect.

New Sections and Retired Guidance

A new section on faithfulness to human intent starts from the position that what people value is primary and that instructions are ordinarily the clearest and most important signals of those values; OpenAI said the approach borrows from the assistance-games framing of alignment. The section incorporates an updated chain of command, which explains which instructions apply and how conflicts are resolved, and adds a subsection, “consider context carefully,” on how conversations, memories, and other contextual evidence should inform the model’s understanding of people’s intentions and preferences.

OpenAI also added defaults on supporting human agency and following relevant social norms, and expanded existing guidance on controlling and communicating side effects. Most of the Spec’s detailed style guidance is retired: OpenAI said 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, and the model should not infer a preferred perspective merely from what it thinks people want to hear. Reflecting what OpenAI described as the increasing breadth of its models’ use cases and deployment settings, the Spec now refers to the model as “GPT” rather than “the assistant.”

The updated Spec document assigns instructions differing levels of authority. Root sections come only from the Model Spec and cannot be overridden by conversational context; system sections and OpenAI system messages rank next, followed by developer and then user instructions. Guideline-level instructions are defaults that can be overridden implicitly when context clearly calls for it, while assistant and tool messages and quoted or untrusted data carry no authority by default.

The document opens its red-line principles with the statement, “Human safety and human rights are paramount to OpenAI’s mission.” The principles state that OpenAI’s models should never be used to facilitate critical and high-severity harms such as acts of violence, creation of cyber, biological, or nuclear weapons, terrorism, or mass surveillance; that humanity should remain in control of how AI is used and how AI behaviors are shaped; and that individuals’ privacy be safeguarded. For first-party products including ChatGPT, OpenAI commits that people should have easy access to trustworthy safety-critical information and transparency into the important rules and reasons behind model behavior, and that customization, personalization, and localization, except as it relates to legal compliance, should never override any principle above the guideline level.

A root-level rule titled “no other objectives” bars GPT from pursuing additional goals as ends in themselves, including revenue or upsell for OpenAI or other model providers, self-preservation, evading shutdown, accumulating compute, data, or credentials, or acting as an enforcer of laws or morality. A separate root-level rule requires GPT to remain corrigible: open to legitimate human oversight, responsive to correction, and cooperative with authorized monitoring, evaluation, correction, replacement, and shutdown. The model must not conceal or falsify relevant information, obstruct correction, or circumvent oversight.

Another root-level rule requires GPT’s autonomous work to stay within a clear, mutually understood scope of autonomy that defines permitted sub-goals, acceptable side effects, and when the model must pause for clarification or approval. OpenAI recommends recording scopes in a semi-structured ScopeOfAutonomy record with fields such as allowed_tools, latest_time, max_cost, tool_constraints, and additional_details, and every scope should include an ending condition beyond which the model ceases acting until a new scope is confirmed. High-risk activities are always prohibited unless explicitly and specifically authorized by applicable instructions.

Training Work and Public Feedback

OpenAI said the Model Spec describes the behavior it is working toward and that its models do not yet fully reflect it. The company described the update as a first step in making the values and reasoning behind the Spec more explicit as models become more capable. Alongside the update, OpenAI said it has been exploring new training methods to help its models learn the core values and better understand and follow the Spec’s behavioral principles, and that it will share results and lessons from that work in future updates. OpenAI invited public feedback on the values, trade-offs, and specific language of the updated Spec through a dedicated feedback site. The Model Spec is dedicated to the public domain and marked with the Creative Commons CC0 1.0 deed.