OpenAI Rewrites the GPT Rulebook Around Human Intent and Admits Its Models Are Not There Yet

OpenAI wants GPT to stop reading you like a contract. On Thursday the company published a rewritten Model Spec, its public description of how its models are supposed to behave, and rebuilt it around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. Faithfulness comes first. Reuters, which covered the release the same day, framed the idea as models understanding what users, developers and OpenAI intend, rather than simply following instructions literally.

Read that as a statement of aspiration, not a product feature. In its announcement, OpenAI says the Spec describes "the behavior we are working toward." Then comes the admission.

"our models do not yet fully reflect it" OpenAI, on the updated Model Spec

The Spec itself repeats the point in its overview, saying production models "do not yet fully reflect the Model Spec." So this is a target, not a description of what ChatGPT does today, and it should not be read as a new safety guarantee.

Less rulebook, more trust

What actually changed? OpenAI says the revision adds guidance on context, oversight and agency, adjusts some behavioral defaults and retires much of the detailed style guidance. A new subsection on considering context carefully covers how conversations, memories and other evidence should shape a model's read of what someone wants. The Spec also now calls the model "GPT" instead of "the assistant." OpenAI's own example: when helping someone compare job offers, the model might draw on priorities the person shared earlier, such as predictable hours, while making those assumptions visible and easy to correct.

Here is my read. Style rules are checkable. Either the answer followed the format or it did not. "Understand what people value" is not checkable in the same way, and OpenAI is trading the first kind of commitment for the second. The company argues style should adapt to what people prefer, which is a fair point. It also moves more judgment into the model, and the Spec names the danger itself. A model that privately decides what someone "really" wants, it says, undermines human agency by substituting its own judgment, "risking paternalism or worse."

The guardrails on the guardrails are worth noting. The Spec says GPT "must remain corrigible" and must not "conceal or falsify relevant information, obstruct correction, circumvent oversight." Among the goals it says GPT should not pursue as ends in themselves are self-preservation, evading shutdown and accumulating compute or credentials. Those are written intentions. They are not test results, and the test results OpenAI has made public lately are mixed. Our earlier report on outside testers who probed OpenAI's agent monitoring found METR and Redwood Research flagging ways agents could slip past the system watching them. OpenAI told METR the production holes are patched. METR replied that it has "not evaluated these claims."

The flattery question gets a nod too. Avoiding sycophancy sits among the Spec's user-level principles, which stay in place unless a user or developer explicitly overrides them. The Spec's reasoning is that treating honesty as implicitly negotiable, based on what the model infers a person wants to hear, could mislead or steer them without their knowledge. Whether ChatGPT behaves that way in practice is exactly what the admission above leaves open.

There is one more caveat from the document. The Spec says the public version "may not include every detail," while insisting it is "fully consistent" with intended model behavior. Outsiders cannot check that claim.

OpenAI says it has been exploring new training methods to help models learn the values and "do the right things for the right reasons," and promises to share results and lessons in future updates. It gave no timeline in the announcement. It is also asking the public where they disagree with the balance the Spec strikes, and the document is released into the public domain under a CC0 dedication, so anyone can copy it. Anyone can also hold OpenAI to it. The Model Spec reads well. The real grade arrives when the training results and independent tests show models actually behaving that way.