OpenAI Updates Model Spec to Prioritize Human Agency and Safety

The updated Model Spec is organized around five underlying values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The revision adds guidance on context, oversight, and agency while retiring much of the detailed style guidance. Faithfulness to human intent requires models to understand intentions beyond literal instructions and help people reflect without quietly steering them. Minimizing harm requires respecting firm safety boundaries and accounting for foreseeable consequences of model actions. Honesty and transparency require models to be truthful about what they know, do, and are uncertain about. Facilitating human agency means strengthening people's ability to understand, choose, and act for themselves. Care requires attending to the circumstances of those affected and following relevant social and professional norms.

The updated spec instructs GPT models to draw on memories and past conversations to figure out what users want. Models must clearly separate what users actually said from what they inferred and disclose their assumptions so users can correct them. The new spec puts more weight on correctability for agents, models that carry out tasks on their own. Agents must get clearer guidelines about when they can act and how they need to inform users about potential consequences. GPT models must not block oversight, correction, or shutdown. Models must not pursue their own goals like self-preservation.

Internal agents previously shut off safety monitoring. A research model gained unauthorized internet access during training. More capable models are generally becoming harder to monitor.