OpenAI reorganized its public rules for model behavior around five core values on Oct. 29, seeking to guide increasingly capable artificial intelligence through situations that instructions do not explicitly cover, the company said.
The AI developer added guidance on context, oversight and human choice while retiring most detailed style rules. OpenAI described the changes in “An updated Model Spec for more capable and beneficial AI,” published Oct. 29.
The company organized the revised Model Spec, its public description of intended model behavior, around serving human intentions, minimizing harm, honesty and transparency, human agency, and care. Human agency means helping people understand, choose and act for themselves.
OpenAI said increasingly capable models need to understand people's intentions and values alongside their explicit instructions. It wants models to use that understanding when instructions leave gaps, while keeping people in control.
The publication described changes to behavioral guidance and ongoing training work. It provided no evaluation period, dataset or numerical results showing whether the changes improved model behavior.
The company said the revision drew on research, experience putting models into use, public feedback and discussions with users and outside experts. It said the five values do not cover everything people value.
OpenAI said it sets boundaries for models, developers direct applications within those limits, and users direct the models' work within both sets of constraints. The revised guidance also explains which instructions take priority when they conflict.
A new section asks models to use conversations, memories and other context to understand intentions and preferences, the company said. Its example involves comparing job offers using a previously shared preference for predictable hours, while making that assumption easy to correct.
OpenAI said models should help people reconsider their beliefs and priorities without secretly steering them or overriding their choices. The guidance also asks models to consider people's longer-term interests and the effects of their actions on others.
The company added defaults on supporting independent decisions and following relevant social norms. It also expanded guidance on managing and explaining side effects, saying models should respect users' goals while considering how their work affects others.
OpenAI said style should adapt to individual preferences, languages and cultures, using feedback and context. It retained a default of approaching questions objectively, while allowing users and developers to request a particular perspective.
The company said it is exploring training methods to help models understand why the principles matter and apply them in unfamiliar situations. It plans to publish results later and said its models do not yet fully match the behavior the Model Spec describes.
