OpenAI updated its Model Spec today, adding guidance on how models should interpret user intentions, use context and communicate the consequences of their actions. The public guide to intended model behavior now centers on five values, while dropping most detailed rules governing writing style. OpenAI says its models still fall short of fully following the Spec, so the revision describes intended behavior rather than an improvement already achieved across its models.
For users, one change concerns how earlier conversations and memories should influence a model’s understanding of their preferences. OpenAI’s example involves comparing job offers: a model could consider a previously expressed preference for predictable hours, but should expose that assumption and let the user correct it. The company says models should help people reconsider their beliefs and priorities without covertly directing them or substituting the model’s choices for theirs.
Writing style is also becoming more dependent on individual preferences, cultural and language differences, and feedback from users. OpenAI is removing most prescriptive style guidance while keeping a stronger default toward objective treatment of questions. Users and developers can request a specific perspective, but models should not assume that perspective simply because they predict it would please the user.
The five organizing values cover fidelity to human intentions, reducing harm, truthfulness and openness, people’s ability to make their own decisions, and care for those affected. OpenAI says these commitments should help models handle unfamiliar circumstances and weigh competing considerations. Its guidance generally establishes defaults that developers and users can change, alongside firm boundaries intended to protect individuals and society.
The revision also updates the instruction hierarchy used to determine which directions take precedence when they conflict. OpenAI sets the boundaries, developers guide applications inside them, and users direct work within those limits. Other additions address social and professional expectations, helping users develop their own capabilities, and managing and explaining side effects of a model’s work.
OpenAI says it is exploring training approaches intended to teach models the reasons behind these principles and help them apply those principles in new situations. The company plans to publish findings from that work in later updates. The Spec also switches its term for the model from “the assistant” to “GPT,” reflecting what OpenAI describes as a wider range of uses and deployment settings.
