OpenAI updates AI behavior standards with five guiding values
The revised Model Spec removes most style requirements and expands guidance on context and personal choice, but OpenAI says its models still fall short.
OpenAI says it has overhauled the Model Spec, its public guide to AI behavior, placing five values at the center of the document. The revision introduces instructions about context and personal choice while removing most rules governing style.
The company identifies faithfulness to human intent as the foundation. It says models should understand the goal behind a request, not merely carry out its exact wording. The other values cover care, human agency, minimizing harm, and honesty and transparency.
Additional guidance addresses memories, earlier conversations and social norms. Users and developers can still modify most default rules, OpenAI said, while firm restrictions remain where protecting people requires them.
The standards discourage responses designed simply to please users. They allow models to adopt a position on request but otherwise call for objectivity. The document now uses the name GPT in place of its previous label for the assistant.
OpenAI acknowledged that its models have not fully met the standards. It said it is experimenting with training methods to instill the values and invited public feedback on the trade-offs reflected in the revised guidance.
