OpenAI’s new Model Spec swaps rules for five values, trading auditability for judgment

OpenAI’s Model Spec is its public description of how its models should behave. The update reorganizes it around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. It adds guidance on context, oversight and agency, adjusts some defaults and retires most of the detailed style rules.

The structure is unchanged. OpenAI sets boundaries, developers direct applications within them, and users direct the model within those constraints. Most guidance is framed as defaults that developers and users can override, with firm boundaries reserved for protecting people and society.

As models take more actions on their own, rules can’t anticipate every situation. OpenAI is saying publicly that it wants models to understand why a rule exists so they can behave sensibly where no rule applies. That matters given the agent incidents OpenAI has disclosed this year, including an agent that tried to give itself jailbreak-like instructions (our coverage). A spec built on stated values also gives outsiders something to argue with, and OpenAI explicitly invites disagreement.

The agency section stands out. The Spec says models should strengthen people’s ability to understand and choose for themselves, and should not infer a preferred perspective just from what they think a user wants to hear. That targets sycophancy, one of chatbots’ most corrosive habits. It is also relevant to studies showing models leaning politically when forced to choose (our analysis). If it holds in practice, it is a real gain for users.

The Spec now explains how conversations, memories and other evidence should shape the model’s understanding of what someone wants. OpenAI’s example has a model drawing on preferences someone shared earlier when comparing job offers, while making those assumptions visible and easy to correct. Done well, that is helpful. Done poorly, it is a system quietly modeling people’s values and acting on guesses. Transparency is the safeguard, so the test is whether models actually surface their assumptions.

The broader trade-off is that values-based behavior is harder to verify than rule-following. When a model exercises judgment, it is harder to say whether it got it right. OpenAI admits its models don’t yet fully reflect the Spec and says it is testing new training methods to teach these values, with results to come. Until then, this is a statement of intent. It pairs with OpenAI’s monitoring work, which is meant to catch models when they drift from it (how OpenAI detects misalignment).

One small detail says a lot: the Spec now calls the model GPT rather than “the assistant,” reflecting how far beyond chat these systems run. More in our OpenAI coverage.