OpenAI has issued a major revision of its public Model Spec, reorganizing the document around five values: faithfulness to human intent, minimizing harm, honesty and transparency, facilitating human agency, and acting with care.
The update adds a more explicit account of how GPT should use conversation history, memory and other context while keeping assumptions visible and correctable. It also revises the chain of command, expands guidance on controlling side effects and adds defaults for human agency and relevant social norms.
OpenAI retired much of the prior document’s detailed style guidance, arguing that tone and presentation should adapt to user preferences and cultural context. Stronger defaults, including objectivity and truthfulness, remain unless appropriately overridden under the Spec’s authority hierarchy.
The dated document also replaces the generic term “assistant” with “GPT,” reflecting use in agents, APIs and other systems that can take actions beyond conversational replies.
The change is a governance and training target, not proof of a simultaneous behavioral update across every deployed model. OpenAI explicitly says production systems do not yet fully reflect the Model Spec and that the document may omit implementation details.