OpenAI has rewritten the Model Spec, the public rulebook describing how its models should behave. Instead of long lists of rules, the new document is built around five core values, with "fidelity to human intent" placed at the top of the hierarchy. In the same document, the company makes a striking admission: its models cannot yet fully follow this specification.

What happened?

The updated Model Spec adds new guidance on context and human agency, strips out most of the style rules, and reorganizes the model's naming. Rather than rigid "do this, don't do that" instructions, the document now offers a priority order that explains which value takes precedence when situations conflict. Fidelity to human intent comes first; the other values are positioned beneath it.

Why does it matter?

On the surface this change looks like a documentation cleanup, but it actually reflects a shift in alignment philosophy. A rule-based guide tries to define in advance every situation a model might encounter; in practice this is impossible, because language and context are infinitely varied. A value-based guide, by contrast, gives the model a compass and leaves the interpretation to the model. With this transition, OpenAI is saying it has chosen to set a general direction rather than micromanage behavior.

The truly critical point here is that the company openly admits its models cannot fully implement the specification. This is an indirect acknowledgment that the alignment problem remains unsolved. Abstracting rules into values also makes evaluation harder: the question "Did it produce that forbidden sentence?" is measurable, but "Did it stay faithful to the user's true intent?" is open to interpretation. OpenAI is moving from a framework that is easy to test to one that is harder to test but more realistic.

The removal of most style rules completes the same picture. Early chat assistants had an ideal of the "neutral, personality-free helper." Loosening the style constraints shows a move away from that ideal and toward giving the model a wider space for expression. This freedom can enrich the user experience, but it also makes it less clear when the model will choose which tone.

Freedom, or a blurring of responsibility?

There are two sides to the coin. The optimistic reading is this: fixed rules, in unexpected situations, produce results that are either overly cautious (unnecessary refusals) or blindly obedient. A value hierarchy gives the model room to weigh the situation and can produce more accurate responses.

The cautious reading is this: leaving interpretation to the model also blurs responsibility. If a model, in the name of "fidelity to human intent," assumes something the user never explicitly said, is the source of the error the rule or the model's interpretation? Abstract values can make accountability harder when something goes wrong. The more open to interpretation the specification is, the harder it becomes to give a clear answer to the question "Was the spec followed?"

ApproachAdvantageRisk
Strict rule listMeasurable, auditableRigid or blindly obedient in unexpected situations
Value hierarchyFlexible, context-sensitiveInterpretive ambiguity, weaker accountability

What's next?

There are a few things to watch. The first is how OpenAI will measure the rate at which models comply with the spec, and whether it will disclose this publicly. The weakest link in a value-based guide is verifying compliance. The second is how competitors respond: companies like Anthropic already champion value-based, principle-driven alignment approaches. OpenAI's shift looks like part of a broader industry move from rule lists toward principle sets.

The third is that placing "fidelity to human intent" at the top raises the question of what happens when intents conflict. When a user's intent clashes with platform policy, which one wins? The more clearly the document's hierarchy answers this question, the greater its practical value will be.

Ultimately, this update is not a marketing refresh; it is an open admission that governing AI behavior remains an unsolved engineering and philosophy problem. OpenAI is moving to a more realistic framework, but whether that framework works will be shown not by the rules themselves, but by the measurable behavior of the models.