OpenAI has reorganized the public rulebook that describes how its artificial intelligence models should behave, replacing much of its detailed style guidance with a framework built around five core values.
The company published the revised Model Spec on Thursday. It said the framework is intended to guide GPT when a request is not covered by a specific instruction. OpenAI also acknowledged that its existing models do not yet consistently meet the new standard and said it is testing training approaches designed to teach the values.
The values guiding GPT
The new structure places faithfulness to human intent at its foundation. OpenAI said a model should understand what users, developers and the company are trying to accomplish instead of following a request only in its most literal form.
- Faithfulness: interpreting requests in light of a person’s goals and the applicable oversight.
- Harm minimization: accounting for safety risks in the model’s responses and actions.
- Honesty and transparency: accurately communicating what the model knows and does not know.
- Human agency: helping people make decisions and pursue their own objectives.
- Care: taking users’ circumstances seriously when responding.
OpenAI described the values as a way for models to exercise judgment in situations where instructions do not provide a complete answer. The company wrote:
“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.”
Rules remain layered
The update keeps a hierarchy for directing the model. OpenAI establishes the outer boundaries in the Model Spec, developers set instructions for their applications within those boundaries and users guide the model’s work within the space developers allow.
The revised guidance adds material on reading context, supporting people’s ability to make their own choices and following social norms. It also expands discussion of side effects, including how models should control and explain effects that may result when they take actions rather than simply generate text.
At the same time, OpenAI is retiring much of the document’s former advice on response style. The company now refers to the model as “GPT” instead of “the assistant,” reflecting a document focused more on broad behavioral principles than on a long catalog of presentation rules.
OpenAI said the Model Spec represents an intended target for its systems rather than a claim that current models already satisfy every provision. The company’s stated next step is continued testing of training methods aimed at improving how models apply the five values in practice.
