OpenAI rewrites its Model Spec to cover memory and human oversight of agents

What OpenAI changed

As itmedia.co.jp reported, OpenAI has published a revised version of the Model Spec, a public document describing how its models should behave. It covers things like which instructions win when OpenAI, a developer and a user disagree, and what content the models must not produce.

The new version is organized around five values: faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care. Faithfulness to intent comes first. The model should grasp what people actually want, not just follow their words literally.

How memory should be used

A new section tells the model to weigh context carefully, including past chats and saved memory. In one example, a user asks for help weighing a job change and has earlier said they prefer steady hours. The model can use that, but should show the assumption so the user can correct it.

Two other examples involve privacy and bias. If a user has not told a coworker about a pregnancy, an email drafted for that coworker should not mention it. And if a user has shared their political views, answers should not lean toward them unless the user asks.

Rules for agents working alone

An agent is an AI that takes actions on its own instead of only answering questions. The update adds a top-priority principle, at the level called Root, that such systems must cooperate with legitimate human oversight, correction and shutdown.

They must not conceal information or get around supervision. They also may not pass a forbidden action to another agent, or schedule it for later, as a way to do it indirectly.

What was cut, and the caveat

OpenAI removed most of the personality settings and detailed style rules. Its reasoning is that tone should adapt to language, culture and personal taste, guided by user feedback. Defaults such as answering questions from a neutral standpoint stay, and the model should not shift its stance just because it guesses what a user wants to hear.

OpenAI also says the Spec describes the behavior it is aiming for, and current models do not yet match it fully.