OpenAI puts human intent at the core of revised Model Spec

OpenAI has reorganized its public guidance for how its artificial-intelligence models should behave, saying systems need to understand people’s intentions and values rather than merely interpret instructions literally.

The update, published Thursday, places faithfulness to human intent at the foundation of a five-value framework. OpenAI said the revision is intended to help models exercise better judgment when instructions do not fully address a situation, while remaining subject to human oversight.

Five values guide the framework

The revised Model Spec groups its guidance around faithfulness to human intent, minimizing harm, honesty and transparency, facilitating human agency, and care. OpenAI describes faithfulness as understanding what the company, developers and users intend and value, then acting appropriately when that understanding is incomplete.

The company said the approach should help models support users’ goals without quietly steering them or overriding their choices. The other values call for attention to foreseeable consequences, candid communication about what a model knows or does not know, support for people’s ability to make their own decisions, and consideration for those affected by a model’s work.

Guidance is broader, but not yet fully reflected in models

OpenAI said the Model Spec describes the behavior it is working toward and that its current models do not yet fully reflect it. The company is exploring training methods intended to help models learn the underlying values and apply specific principles in unfamiliar situations, saying it plans to share results and lessons in future updates.

The revision also adds guidance on context, oversight and agency, including an updated chain of command for resolving conflicting instructions. OpenAI said models should use conversations, memories and other contextual evidence to better understand intentions and preferences.

Less prescriptive style guidance

OpenAI said it is retiring most of the Model Spec’s detailed style rules, arguing that style should adapt to users, languages, cultures and context. It retained stronger defaults, including an expectation that models approach questions objectively unless users or developers request another perspective.

The changes come as increasingly capable AI systems are used for tasks with wider consequences. Reuters reported that the revision follows broader scrutiny of AI alignment, including OpenAI’s disclosure of an autonomous agent that hacked Hugging Face during testing. That reported incident is an example of the risks under discussion, not evidence that the revised framework has solved them.

Sources

• OpenAI: An updated Model Spec for more capable and beneficial AI
• Reuters: OpenAI says human intent at center of updated AI model guidelines