OpenAI rewrites its Model Spec around five core values and retires style rules

OpenAI has rewritten its Model Spec, the public rulebook for how its models should behave, around five core values [1]. The update adds guidance on context, oversight and agency as models take more consequential actions, and drops most of the detailed style rules [1].

What are the five values?

OpenAI reorganised its Model Spec around five values on 29 October. They are faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care [1].

OpenAI calls faithfulness to human intent the foundation: going beyond literal instructions to understand what OpenAI, developers and users intend, and acting wisely when that understanding is incomplete [1]. The chain of command stays: OpenAI sets boundaries, developers work within them, and users direct the model within those constraints [1].

What changes in practice?

A new subsection tells models to consider context carefully, using conversations, memories and other evidence to understand people's intentions and preferences [1]. OpenAI also added defaults on supporting human agency and following social norms, and expanded guidance on side effects [1].

OpenAI is retiring most of the Spec's detailed style guidance. Style should adapt to what people prefer, the company says, while stronger defaults such as approaching questions objectively remain [1].

OpenAI now refers to the model as "GPT" rather than "the assistant" in the Spec. OpenAI says this reflects the growing range of settings in which its models are used [1].

Do the models already behave this way?

Not fully, by OpenAI's own account: the Spec describes the behaviour it is working toward, and its models do not yet fully reflect it [1]. The company says it is exploring new training methods to teach the values and will share results later [1].

Our read

Opinion: Moving from style rules to values is the right direction for agents that act on their own, but values are harder to audit than rules. The real test is whether OpenAI publishes evidence that its models follow the new Spec, which it says is still to come.

What we could not verify

• We reviewed OpenAI's announcement, not every changed section of the full Spec text.
• Which current OpenAI models have been trained on the updated Spec; OpenAI says they do not yet fully reflect it.
• Results of the new value-training methods, which OpenAI has not yet published.