OpenAI rewrites its rules for GPT around five core values

OpenAI has rewritten the rulebook for how its AI models should behave. The new version reorganizes its guidance around five underlying values, the company says.

The company announced the change Thursday in a blog post on its updated Model Spec. OpenAI calls the Spec “our public description of how we want our models to behave.”

What changed

The five values are “faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care,” the post says. OpenAI calls faithfulness to human intent “the foundation of this approach.”

The post spells out the other four. Minimizing harm means “respecting firm safety boundaries.” Honesty and transparency mean being truthful about what the model knows and does. Facilitating human agency means “strengthening people’s ability to understand, choose, and act for themselves.” Care includes “following relevant social and professional norms.”

The update adds a subsection on context. It explains how conversations, memories and other evidence should shape the model’s read of what people want. OpenAI also added defaults on supporting human agency and following social norms. It expanded guidance on controlling and communicating side effects.

OpenAI is “retiring most of the Spec’s detailed style guidance.” Style should adapt to what people prefer, the company says. Some stronger defaults stay, such as approaching questions objectively. The model “should not infer that preference merely from what it thinks they want to hear,” the post says.

The Spec now calls the model “GPT” rather than “the assistant.”

Who wrote it and who it covers

OpenAI wrote it. The post names no individual author. It lays out a chain of command: “OpenAI writes the Spec, sets boundaries, and trains models to respect them; developers direct applications within those boundaries; and users direct the model’s work subject to those constraints.”

The Spec covers OpenAI’s own models. Most of it takes “the form of defaults that users and developers can override,” the post says. Firm boundaries remain where OpenAI says they are needed to protect people and society.

What it leaves out

The Spec is a statement of goals, not a description of today’s products. “The Model Spec describes the behavior we are working toward; our models do not yet fully reflect it,” OpenAI writes.

The five values “are not intended to capture the full range of human values our models should respect,” the company adds.

Our read

A shift from rules to values gives models more room for judgment. That cuts both ways. Judgment can handle cases no rule anticipated. But it is harder to audit. A list of rules can be checked line by line. A value like “care” cannot be checked so easily.

The ordering matters too. Faithfulness to human intent sits first, and the chain of command puts OpenAI at the top. Developers and users work within the boundaries OpenAI sets. That is a candid description of who holds power over the product. It also means the company is both the author and the main judge of the standard.

The style changes will draw the most scrutiny. OpenAI is dropping most detailed style rules while keeping objectivity as a default. It also says the model should not infer that people want a particular perspective merely from what it thinks they want to hear. The Spec now gives users and researchers a written standard to test against.

OpenAI deserves credit for publishing its goals and admitting its models fall short of them. The open question is evidence. The company promises training results later. Until then, the Spec is a promise, not a measurement.

What’s next

OpenAI says it is exploring new training methods to help its models learn the core values. “We’ll share results and lessons from this training work in future updates,” the company says. It is also inviting public feedback on the values, trade-offs and language in the Spec.

Sources: OpenAI blog post, Oct. 29, 2026; OpenAI Model Spec.