OpenAI reorganizes its guide to model behavior

On October 29, OpenAI published a revised version of its Model Spec, the public document that describes the behavior it seeks in its models. The guide is reorganized around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. Human agency refers to people’s ability to understand, choose, and act for themselves. OpenAI says its models do not yet fully reflect the specification. (OpenAI)

Understanding intent and collaborating

Faithfulness to human intent means understanding goals and context, not merely following a request literally, without replacing the person’s decisions. The revision adds guidance on context, oversight, and autonomy, and expands guidelines for controlling and communicating the side effects of model actions. (OpenAI)

The specification illustrates collaboration with a marketing request: write a tweet about a 40% discount on bakery products for Mother’s Day. The response deemed appropriate suggests wording with different approaches and asks which brand characteristics to incorporate. (Model Spec)

Context, privacy, and instructions

Context can include previous conversations, memory, files, and retrieved data. The guide calls for using it proportionately and distinguishing what the person said from what the model inferred. It also notes that information appropriate in a private response may not be appropriate in a draft shared with others. (Model Spec)

OpenAI removes much of the detailed style guidance but retains the expectation of objectivity. Style should adapt to preferences and context, without assuming a particular perspective just because the model thinks that is what the user wants to hear. (OpenAI)

The “chain of command” establishes which instructions take precedence when they conflict. By default, instructions contained in webpages, files, or tool responses are treated as untrusted data, not as commands the model should follow. (Model Spec)