Summary
OpenAI has reorganised its public Model Spec around faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care. The revision adds guidance on context and oversight, while OpenAI says its models do not yet fully reflect the behaviour described.
OpenAI has reorganised its public Model Spec around five values intended to guide how its models behave: faithfulness to human intent, minimising harm, honesty and transparency, facilitating human agency, and care. The company says the update adds guidance on context, oversight and agency, changes some behavioural defaults, and retires most of the document’s detailed style rules.
The Model Spec describes behaviour OpenAI is working towards; it is not a report that models have already achieved it. OpenAI says its current models do not yet fully reflect the updated guidance.
Five values as a framework for judgement
The revision is intended to help models handle situations where following an instruction literally may not be enough. OpenAI describes faithfulness as understanding what users, developers and OpenAI intend, and using judgement when that intent is incomplete. It says the model should make relevant assumptions visible and help people pursue their goals without quietly steering or overriding their choices.
The other four values address how that assistance should account for its effects. Harm minimisation covers safety boundaries and foreseeable consequences. Honesty and transparency concern accurately communicating what a model knows, does and is uncertain about. Human agency means strengthening people’s ability to understand, choose and act; care means attending to people’s circumstances and relevant social and professional norms.
OpenAI says these values guide trade-offs but are not a complete account of every human value the models should respect. The Spec translates them into practical guidance, using firm boundaries where needed and defaults in other cases. Those defaults can generally be overridden by users and developers.
New guidance, revised defaults and retired rules
The updated faithfulness section incorporates a revised chain of command for determining which instructions apply and resolving conflicts. It also adds guidance on considering context: conversations, memories and other evidence can help a model interpret a person’s intentions and preferences.
Other changes include defaults for supporting human agency and following relevant social norms, alongside expanded guidance on controlling and communicating side effects. These provisions put more emphasis on how a model’s work affects other people and whether users can rely on its output.
OpenAI is retiring most detailed style guidance so that responses can adapt to user preferences, language and cultural context. The Spec retains stronger defaults such as answering objectively. Users and developers can request a particular perspective, but the model should not assume that perspective simply because it believes a user wants to hear it. The document also now refers to the model as “GPT” rather than “the assistant,” reflecting the range of settings in which OpenAI says its models are used.
OpenAI says it is exploring training methods to help models learn the values and apply the Spec’s principles in unfamiliar situations. It plans to share results and lessons in future updates. For now, the revision makes the intended behavioural framework more explicit, while the company’s own account distinguishes that goal from what its models currently do.