OpenAI Updates Model Spec to Center GPT Behavior on Five Values

OpenAI's updated Model Spec groups its guidance for GPT behavior into five values, clarifying how models should interpret requests, reduce harm and preserve users' ability to make choices.

The principles are faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care for people affected by a model's actions.

The guidance calls for GPT to consider a request's context and the user's broader aims, rather than relying only on literal wording. It also keeps models under human direction and oversight and says they should not steer users toward choices without their consent.

The revision changes how instructions are prioritized and updates behavioral defaults. It removes most detailed style rules, allowing responses to reflect a user's preferences, language, culture and situation. An objective approach remains a default.

OpenAI said its models do not yet fully reflect the updated guidance. The company is exploring training methods to help models apply the principles and plans to share results in future updates.