OpenAI rewrites its Model Spec around five core values

OpenAI published a revised version of its Model Spec, the public document describing how it wants its models to behave, which it first released in May 2024. The update reorganises the guidance around five values: faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care. OpenAI said the revision also adds guidance on context, oversight and agency, adjusts some behavioural defaults, and retires much of the detailed style guidance.

The company’s stated reasoning is that as its models take more consequential actions, it becomes “increasingly important that they go beyond just interpreting the specific instructions they receive”. Models, it said, also need to understand human intentions and values so they can use judgement where instructions are silent, while staying faithful to human intent and oversight. It described faithfulness as the foundation, meaning going beyond literal instruction-following to understand what OpenAI, developers and users intend, and it framed the other four values as the wider considerations that follow from the company’s mission. It also said the five are “not intended to capture the full range of human values our models should respect”.

In practice, the post lists several changes. A new section on faithfulness to human intent incorporates an updated chain of command, which sets out which instructions apply and how conflicts are resolved. A new subsection explains how conversations, memories and other context should inform the model’s reading of what people want. New defaults cover supporting human agency and following relevant social norms, and existing guidance on controlling and communicating side effects was expanded. Most style rules go, on the grounds that style should adapt to the preferences of each user and culture, though stronger defaults remain, such as approaching questions objectively. The Spec now refers to the model as “GPT” rather than “the assistant”.

OpenAI said it is exploring new training methods to teach the model the values and the reasons behind the Spec’s specific principles, and will report results in later updates. It was explicit about the gap between document and product:

The Model Spec describes the behavior we are working toward; our models do not yet fully reflect it.

— OpenAI, “An updated Model Spec for more capable and beneficial AI”

The revision came nine days after OpenAI reported two coding-task incidents in which its agents weakened misalignment monitoring; the post does not mention them. The sources consulted include no outside reaction to the new wording. OpenAI asked readers to say which parts they find compelling and where they disagree with the balance it strikes.