OpenAI introduces revised model specifications to enhance AI alignment with human values

OpenAI has released an updated Model Spec, emphasizing human values and intent through five core principles: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The revision includes clearer guidance on context, oversight, and agency, while retiring detailed style guidelines. Models are now referred to as 'GPT' to reflect broader use cases. OpenAI aims to build models that support users in shaping their lives while respecting their choices and the impact of their actions on others. The update introduces new training methods to help models better understand and follow these principles, with future updates planned to share results from this work. OpenAI invites feedback to refine its approach to AI development. The revised Spec emphasizes understanding and respecting evolving user values, incorporating an updated chain of command for resolving instruction conflicts. It includes a new subsection on careful context consideration and defaults to support human agency and follow social norms. Guidance on managing side effects is expanded, retaining strong defaults like objective reasoning. Users can request specific perspectives without inferring preferences from assumptions. This update marks a step toward making the reasoning behind the Spec more explicit as models become more capable, with ongoing improvements planned based on research and user feedback.