The Pulse
OpenAI Recasts Its Model Spec Around Five Values
OpenAI published an updated Model Spec on October 29, organizing intended GPT behavior around five values, including faithfulness to human intent and harm minimization. The company says its models do not yet fully meet the standard and desc
AI.info Team · Written with AI · fact-checked ·
OpenAI is setting a higher bar for how its models should interpret human intent while acknowledging that its deployed models have not reached that bar. In an update published October 29, the company reorganized its public Model Spec around five values and said the document describes behavior it is still working toward.
OpenAI’s five values set the new frame
The revised Spec groups its behavioral guidance under faithfulness to human intent, minimizing harm, honesty and transparency, facilitating human agency, and care. OpenAI says the values give a clearer account of commitments already reflected in the document and help explain how a model should handle situations its instructions do not spell out.
“The Model Spec describes the behavior we are working toward; our models do not yet fully reflect it,” OpenAI writes. The admission makes the document a statement of intended behavior, not evidence that every model currently follows the guidance.
OpenAI presents faithfulness as more than literal compliance: models should consider what users, developers and OpenAI intend, while remaining within boundaries set by the Spec. The company also says users’ goals and preferences can be uncertain or change over time, so models should make relevant assumptions visible and leave people able to correct them.
Human intent still sits within a chain of command
The updated Model Spec retains a hierarchy for resolving conflicting instructions. It places root-level rules above system, developer and user instructions, followed by guidelines and content with no authority by default. That structure means the call to follow human intent does not give every user request equal standing.
OpenAI describes firm safety boundaries as protection against serious harm, while much of the remaining guidance takes the form of defaults that users or developers can override. The Spec’s overview also says the hierarchy aims to balance individual interests with collective ones. That distinction is central to the update’s promise: assistance should serve people’s purposes without ignoring foreseeable effects on others.
New guidance asks models to consider context, including conversations and memories, when interpreting preferences. OpenAI says that context should help models understand intent; it does not replace the user’s ability to steer them. The revision also adds defaults for supporting human agency and following relevant social norms, and expands guidance on controlling and communicating side effects.
OpenAI retires much of its style guidance
The company is removing most of the Spec’s detailed style rules, arguing that style should adapt to people’s preferences, languages, cultures and context. It keeps a stronger default for objective answers, while allowing users and developers to request a particular perspective. The document says models should not infer that preference simply because they think a user wants to hear a certain answer.
OpenAI also replaces “the assistant” with “GPT” throughout the Spec, saying the broader term better reflects the range of uses and deployment settings for its models. The document applies to GPT models that power OpenAI products, including its API platform.
The Spec is a target, not a performance report
OpenAI says it has been exploring training methods to help models learn the values and apply the Spec’s principles in unfamiliar situations. The company has not presented results from that work in this update; it says it will share findings in future updates. For now, the revision sets out the behavior OpenAI wants to train toward and invites feedback on its values, trade-offs and wording.
The distinction between written guidance and model behavior is explicit in the Spec itself: OpenAI says its production models do not yet fully reflect it. The document’s immediate change is therefore to the public standard, with the company’s own account of model compliance still unfinished.