OpenAI Reframes Model Spec Around Five Core AI Values
OpenAI has published an updated Model Spec that reorganizes its public guidance for GPT around five core values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The company says the revision is designed for models that are becoming more capable and are taking more consequential actions.
The Model Spec describes the behaviour OpenAI wants its models to exhibit. It sets out how OpenAI establishes boundaries, how developers direct applications within them, and how users direct a model’s work subject to those constraints. OpenAI stresses that its models do not yet fully reflect the updated Spec.
From literal instructions to contextual judgment
Faithfulness to human intent is presented as the foundation of the new structure. OpenAI says this means more than following wording literally: GPT should seek to understand what OpenAI, developers and users intend and value, while acting carefully when that understanding is incomplete.
The company gives the example of helping a person compare job offers. A model could use a previously stated priority, such as predictable hours, but should make that assumption visible and easy for the person to correct. OpenAI also says a model should not quietly steer people or override their choices as their beliefs and values change.
The other four values make further constraints explicit. Harm minimization includes respecting firm safety boundaries and considering foreseeable consequences. Honesty and transparency require models to communicate what they know, do and remain uncertain about. Human agency focuses on helping people understand, choose and act for themselves, while care calls for attention to affected circumstances and relevant social or professional norms.
New guidance and fewer style prescriptions
The updated Spec includes a revised chain of command for determining which instructions apply and resolving conflicts. It adds a section on considering context carefully, covering conversations, memories and other evidence that can inform a model’s understanding of intentions and preferences.
OpenAI also introduced defaults covering support for human agency and adherence to relevant social norms, and expanded guidance on controlling and communicating side effects. The stated purpose is to account for effects on other people, strengthen users’ capabilities and produce work people can rely on while respecting the user’s goals and instructions.
At the same time, OpenAI is retiring most detailed style guidance. It says style should adapt to aggregate language and cultural preferences as well as individual user feedback and context. Stronger defaults remain, including an objective approach to questions. Users and developers may request a perspective explicitly, but GPT should not infer one simply because it believes that is what a person wants to hear.
Implications for AI governance
The Spec now uses “GPT” rather than “the assistant”, reflecting a wider range of use cases and deployment settings. OpenAI says it is also exploring training methods intended to help models learn the underlying values and apply the behavioural principles in unfamiliar situations, with results and lessons to be shared in future updates.
For organisations using GPT, the update offers a clearer governance reference, but not a substitute for deployment controls. Teams should map the Spec’s principles to their own instruction hierarchies, review processes and testing for transparency, context use, side effects and user control before placing models in consequential workflows.