An updated Model Spec for more capable and beneficial AI
OpenAI has revised its Model Spec to better align AI behavior with human values and intentions, introducing five guiding principles for more responsible and adaptable model behavior.
OpenAI has updated its Model Spec, a public framework outlining how its AI models should behave, to emphasize five core values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The revision reorganizes existing guidance, clarifies commitments, and introduces new principles on context, oversight, and agency while retiring much of the detailed style guidance. The update reflects ongoing efforts to refine the Spec through research, user feedback, and expert discussions, aiming to ensure models understand and apply human values in unfamiliar situations.
The revised Spec strengthens the principle of faithfulness to human intent, requiring models to go beyond literal instruction-following and deeply understand user values and priorities. It introduces a new chain of command to resolve conflicting instructions and emphasizes careful consideration of context, such as conversations and user memories, to better infer intentions. Defaults now support human agency and adherence to social norms, while guidance on controlling side effects has been expanded to address broader impacts on others.
OpenAI is also retiring most of the Spec’s detailed style guidance, shifting toward adaptable behavior that reflects user preferences across languages and cultures. The Spec retains stronger defaults, such as objective questioning, but allows users and developers to explicitly override these preferences. The update replaces references to 'the assistant' with 'GPT' to reflect the growing diversity of model applications and deployment settings.
The updated Model Spec describes the behavior OpenAI aims to achieve, acknowledging that current models do not yet fully meet these standards. The company plans to continue refining both the Spec and the models based on research, real-world use, and public feedback, inviting stakeholders to review the changes and share perspectives on the values, trade-offs, and language used in the framework.