OpenAI revises Model Spec to guide AI judgment beyond literal instructions

OpenAI updated its public rules for model behavior on October 29, 2026, asking AI to understand what people intend—not simply follow their words literally. The revised Model Spec organizes guidance around five values and adds direction on context, oversight and agency. It describes the behavior OpenAI wants, not a standard its models already fully meet.

The catalyst, in OpenAI’s telling, is the growing consequence of models’ actions. As systems become more capable, the company says, they need judgment for situations their instructions do not explicitly cover. That judgment is supposed to remain faithful to human intentions and oversight, rather than become permission for a model to choose its own purposes.

Five values, with a hierarchy behind them

The revision brings together commitments that OpenAI says were already reflected throughout the Spec. The five values explain the reasons behind behavioral rules and provide a structure for handling conflicts between them. They are not meant to capture every human value a model should respect.

• Faithfulness to human intent: understand and help pursue people’s purposes within the rules governing the model.
• Harm minimization: respect firm safety boundaries and account for foreseeable consequences.
• Honesty and transparency: represent facts, uncertainty, capabilities and actions truthfully.
• Human agency: help people understand, choose and act without hidden steering or manipulation.
• Care: attend to the circumstances of people affected by the model’s work.

Those values operate within a chain of command—a hierarchy in which higher-authority instructions take precedence. OpenAI sets boundaries, developers direct applications within them, and users direct the work subject to those constraints. The dated Model Spec distinguishes firm protections from defaults that can be overridden, rather than treating every behavioral preference as equally binding.

Some defaults require an explicit override. The document places honesty, objectivity, avoiding flattery and controlling side effects in that category. It warns that letting a model quietly relax these principles based on what it guesses a user wants could mislead that user or steer them without their knowledge.

Context should inform, not quietly decide

A new context subsection explains how conversations, memories and other evidence should inform a model’s understanding of someone’s intentions. Its principles include using context proportionately, considering its source and recency, respecting privacy and preserving user control. OpenAI’s example is a job-offer comparison: a model might use a previously shared preference for predictable hours, but should make that assumption visible and easy to correct.

The update also adds defaults for supporting human agency and following relevant social norms. It expands guidance on controlling and communicating side effects. OpenAI describes these changes as ways to consider effects on others, strengthen users’ capabilities and produce dependable work while respecting their goals and instructions.

At the same time, OpenAI retired most detailed style guidance. Style should adapt to individual preferences, languages and cultures, using feedback and context as evidence. A stronger default toward objective answers remains: users and developers can request a particular perspective, but the model should not infer that request merely from what it thinks they want to hear.

The Model Spec describes the behavior we are working toward; our models do not yet fully reflect it.

OpenAI, October 29 Model Spec announcement

Training results are still to come

Alongside the revision, OpenAI says it has been exploring training methods to help models learn these values and apply the Spec’s principles in unfamiliar situations. The stated aim is to teach why the principles matter, not only which behavior to produce. Results and lessons from that work are promised in future updates.

OpenAI is also inviting feedback on the values, the balance between competing priorities and the document’s specific language. It asks readers to identify disagreements, omissions and unclear passages as it continues revising both the Spec and its models.