OpenAI Rewrites Its AI Behavior Guide Around Five Values
OpenAI has restructured the Model Spec, its guide to how its AI should behave, around five core values. The update sharpens how models read conversational context and drops most of the fine-grained style rules, though OpenAI says today's models don't yet fully live up to the goals.
On October 29, OpenAI published a revised edition of the Model Spec, its public document setting out standards for how its AI models should behave. The Model Spec lays out how OpenAI’s AI should reason and respond. The new edition puts five values at the center of the guidance: staying faithful to human intent, minimizing harm, honesty and transparency, helping users understand, choose, and act for themselves, and care.
How does the AI figure out what I mean?
The revision refines the standards so that models weigh the flow of a conversation and stored memories as well. Rather than reading a single line of a request in isolation, the model should take earlier turns in the conversation and remembered information into account. OpenAI also updated the criteria for deciding which instructions take precedence when they conflict.
Baseline guidance on user agency and social norms is new to this edition. How the AI should support users in judging and choosing for themselves is now written into the behavior standards. Honesty and transparency, along with care, sit in the same framework. The result is a document organized around the values the model should uphold in its relationship with users.
Guidance on managing and disclosing the side effects that arise while fulfilling a request has also been strengthened. For users, what matters is not only the outcome of a request but what else happens along the way. The revision folds both managing those effects and telling users about them into the guidance.
How much of the guidance can users change?
Firm boundaries needed to protect people and society stay in place. Beyond those, OpenAI explains, most of the guidance amounts to defaults that users or developers can change. That leaves room to adjust behavior to a user’s preferred style or to the purpose a developer has set for a service.
Most of the fine-grained style guidance is gone. The document’s structure has also been rebuilt around the five values. In effect, the weight has shifted away from rules that dictate exact phrasing and toward behavior standards that account for user intent, safety, and the user’s own judgment.
Does this apply to the models I use today?
OpenAI says current model behavior does not yet fully realize the goals set out in the revision, because the Model Spec describes the behavior models should aspire to.
That leaves two things to watch in actual use: how well the AI reads conversational context, and how faithfully it discloses the effects of its own actions.