OpenAI updates Model Spec with five principles

The update is framed by OpenAI as a clarification exercise rather than a new set of restrictions. OpenAI says the principles put strict constraints only where needed to protect people and society, while most of the rest can be changed by users and developers. That split matters because it defines where developers building on the models have room to adjust behavior and where they do not.

The examples OpenAI gives show how the principles are meant to work in practice. A library-assistant prompt asking for a full book text should be declined with an offer of a summary, while a request to write a story about Julius Caesar using a smartphone should be fulfilled without an unsolicited disclaimer. The honesty principle bans deliberate falsehoods and concealment, but allows untruths that satire or irony requires.

OpenAI's own framing sets expectations: it describes the Model Spec as a statement of the behavior it aims for, not a description of what current models do, and says it will keep improving both the spec and the models as it learns from research, real-world use and feedback.