OpenAI has published an update to its public Model Spec, the document describing desired model behavior. The revision reorganizes guidance around five core values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.
The change responds to the growing importance of models understanding human intentions and values as they grow more capable and take actions with real-world consequences. OpenAI says models must do more than follow instructions literally: they should exercise judgment in situations instructions do not explicitly cover while remaining aligned with human intent and oversight.
Key elements of the update
- The Spec now centers on the five values listed above, making explicit commitments to serve human purposes, be truthful, minimize harm, respect people’s freedom to form their own views, and take people’s circumstances seriously.
- The revision clarifies existing principles, adds guidance on context, oversight, and agency, adjusts some behavioral defaults, and retires much of the Spec’s detailed style guidance.
- Detailed style rules were largely removed; instead, style should adapt to users’ preferences and context, though some stronger defaults—such as approaching questions objectively—remain.
- Reflecting broader use cases and deployment settings, the document now refers to the model as “GPT” rather than “the assistant.”
Faithfulness to human intent as the foundation
OpenAI presents faithfulness to human intent as the foundational value: models should understand and help people pursue what matters to them. The Spec frames instructions as the clearest signals of values in most cases, but expects models to infer deeper intentions and act wisely when information is incomplete. At the same time, models should avoid subtly steering or overriding people’s choices and should help users reflect and grow in ways they would endorse.
The updated Spec includes a revised chain of command that explains which instructions apply and how conflicts are resolved. A new subsection, "consider context carefully," explains how conversations, memories, and other contextual evidence should inform the model’s understanding of intentions and preferences.
The other four values
- Harm minimization: respect firm safety boundaries and account for foreseeable consequences of the model’s actions.
- Honesty and transparency: be clear about what the model knows, what it does, and where it is uncertain so people can assess its reasoning and outputs.
- Facilitating human agency: strengthen people’s ability to understand, choose, and act for themselves.
- Care: attend to the circumstances of those affected, follow relevant social and professional norms, and produce work people can rely on.
The Spec translates these values into practical guidance, applying firm boundaries where necessary to protect people and society and otherwise setting defaults that developers and users can override.
Training work and next steps
OpenAI is also exploring new training methods intended to help models learn these core values and better follow the Spec’s behavioral principles. The aim is to build an internal understanding of why the principles matter and how to apply them in new situations—so the models do the right things for the right reasons. OpenAI says it will share results and lessons from this training work in future updates.
The reorganization helped identify where additional guidance was needed, where new principles should be introduced, and where existing guidance should be revised or retired. The updated Spec is intended to make the values and reasoning behind it more explicit as models become more capable, but OpenAI notes that the models do not yet fully reflect the Spec and that both the Spec and the models will continue to evolve based on research, deployment experience, and public feedback.
Invitation for feedback
OpenAI invites readers to comment on which parts of the Spec are compelling, where they disagree with its balance, and what it overlooks or leaves unclear. Feedback on values, trade-offs, and specific language is requested to help improve the Spec; the updated document is publicly available and OpenAI welcomes perspectives and responses.
