OpenAI rebuilds Model Spec around five values
OpenAI has reorganized its public Model Spec around five core values and added guidance on context, agency and oversight, a sign that it wants its models to use judgment rather than follow instructions to the letter.
OpenAI announced an update to its Model Spec, the public document that describes how it wants its GPT models to behave. The guidance is now grouped under five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.
OpenAI calls faithfulness to human intent the foundation. In its framing, models should go beyond literal instruction-following to understand what OpenAI, developers and users intend, and act wisely when that understanding is incomplete. The company says this draws on the assistance-games approach to alignment. The section includes an updated chain of command, which sets out which instructions apply and how conflicts between them are resolved. It also adds a new subsection on how conversations, memories and other context should shape the model's reading of what people want. In one example, a model helping compare job offers might recall that a user previously said they value predictable hours, while making that assumption visible and easy to correct.
The revision also adds defaults on supporting human agency and following relevant social norms, and expands guidance on controlling and communicating side effects. It adjusts some behavioral defaults and retires much of the detailed style guidance. Firm safety boundaries remain. Most other rules are defaults that developers and users can override.
OpenAI says it is testing new training methods meant to help models learn why these principles matter, not just what they say. It will share results later. The announcement does not specify which behavioral defaults changed. OpenAI also acknowledges that its current models do not yet fully reflect the updated Spec.