As artificial intelligence models become smarter and take on more consequential actions in the real world, OpenAI has updated its public Model Spec to guide their behavior. The organization emphasizes that models must go beyond simply interpreting specific instructions; they need to understand human intentions and values to exercise good judgment in situations not explicitly addressed by prompts. This update reorganizes the document's guidance around five underlying values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.
Faithfulness to human intent serves as the foundation of this approach, requiring models to deeply understand what OpenAI, developers, and users intend rather than following instructions literally. This value involves recognizing that people's beliefs can change and helping them reflect without quietly steering or overriding their choices. For instance, when assisting a user with comparing job offers, a model might draw on previously shared priorities like predictable hours while making those assumptions visible and easy to correct. The goal is to expand people's ability to shape their own lives by helping them understand the world and pursue chosen goals.
The remaining four values make explicit considerations for longer-term interests and effects on others. Minimizing harm involves respecting firm safety boundaries and accounting for foreseeable consequences. Honesty and transparency require models to be truthful about what they know, do, and are uncertain about so users can evaluate their reasoning. Facilitating human agency focuses on strengthening people's ability to understand, choose, and act for themselves. Finally, care means attending to the circumstances of those affected, following relevant social and professional norms, and producing work people can rely on.
Alongside these value articulations, the updated Spec introduces new sections on context, oversight, and agency while adjusting behavioral defaults. A new subsection explains how conversations and memories should inform a model's understanding of user intentions, and an updated chain of command clarifies how conflicting instructions are resolved. The document also adds defaults for supporting human agency and following social norms, while expanding guidance on controlling side effects. Conversely, OpenAI is retiring most detailed style guidance, noting that style should adapt to individual preferences and cultural contexts rather than being rigidly enforced.
Reflecting the increasing breadth of use cases, the Spec now refers to the model as "GPT" instead of "the assistant." OpenAI notes that this update is a first step in making values explicit as models become more capable, acknowledging that current models do not yet fully reflect these standards. The organization is exploring new training methods to help models learn these core values and apply them for the right reasons. OpenAI invites public feedback on the updated Spec to help refine the balance of values and address any overlooked areas.
Sources: OpenAI News.