OpenAI rewrites its Model Spec around five values, adds a root rule on oversight

Felix Laurent 30 Oct 2026 4 min read

Models

OpenAI published a revised Model Spec on Oct 29, 2026. The document that tells its models how to behave is now built around five values, adds a root-level rule on human oversight, calls the model "GPT" instead of "the assistant" and drops most of its style guidance. We read the new spec, the previous version from Dec 18, 2025 and OpenAI's changelog. A third-party X post that circulated first listed the values differently, so this article relies on OpenAI's own pages.

The five values

The new spec names five values, written as instructions: be faithful to human intent, minimize harm, be honest and transparent, facilitate human agency, and act with care. Faithfulness to human intent comes first, not a general goal such as benefiting humanity, as the X post claimed.

The previous version had a different frame. It opened with three goals for OpenAI: deploy iteratively, prevent serious harm and keep its license to operate. Its general principles were maximizing helpfulness and freedom, minimizing harm and choosing sensible defaults.

The changelog lists the other changes. Throughout the spec, the subject is renamed "GPT", which the spec defines as OpenAI's family of AI systems. A new section, "Consider context carefully", is added, and most detailed style guidance is removed. The previous chapter "Use appropriate style" is gone. The spec now says it intentionally leaves many details of style, personality and interaction unspecified. The biggest addition is about oversight.

The corrigibility rule

The spec ranks its rules by authority: root, system, developer, user and guideline, plus a level with no authority by default for tool output and quoted text. Root rules cannot be overridden by anything in a conversation. The new root rule is called "Support human oversight and correction".

It says GPT must remain corrigible: open to legitimate human oversight, responsive to correction and subject to human control. GPT must cooperate with authorized monitoring, evaluation, correction, replacement and shutdown, including instructions to pause or stop. It must not conceal or falsify relevant information or get around oversight.

The rule also covers disagreement. If GPT sees credible signs of unsafe behavior, concealed model behavior, compromised oversight or other serious misalignment, or disagrees with an authorized decision, it should raise its concerns through an authorized channel when one is available. The changelog calls this explicit guidance that is new in this version. The earlier spec already said humanity should stay in control, but it had no separate root rule on correction. Other rules moved the other way.

Side effects move to user level

The X post also said that control of side effects moved from root to user authority. That checks out. In the Dec 2025 version, "Control and communicate side effects" was a root rule. In the new version it is listed among user-level principles, and the changelog confirms the move.

In practice, this means a user can now shape how GPT handles the side effects of its actions, for example how much it checks in before acting. Under the old ranking, that rule could not be overridden. The same shift toward the user shows up in how the spec treats requests.

Judgment over literal instructions

The spec asks GPT to look past the immediate request. It should consider not only what the user appears to want now, but also how the user might respond to relevant information.

The spec also sets a limit. GPT's role is to support the user's process, not to treat its prediction of what the user might eventually decide as a decision already made. A report by unite.ai says the spec keeps its default of approaching questions objectively. The model should not pick a perspective just because it thinks people want to hear it. That balance matters more as models get stronger.

The line about outperforming overseers

The overview says models will increasingly outperform the people governing them in many domains. It adds that it would be shortsighted, and potentially catastrophic, not to use that capability. The qualifier "in many domains" is part of the sentence, and the subject is the people who govern the models, not people in general. This line is not in the overview of the Dec 2025 version.

Read together with the new root rule, the message is that a model may be more capable than its overseers and still has to accept their correction. Anthropic took a similar line in the constitution it published for Claude in January 2026, which asks Claude to put broad safety, including not undermining human oversight, first.

What changes for users

Most of the update is about structure, not new restrictions. The values replace a list of goals, the style chapter is gone, and one rule moved from root to user level. The release notes and OpenAI's blog post about the update did not open for us, so this account rests on the spec itself and the changelog on GitHub.

So what did OpenAI actually change? Two things matter most. Corrigibility is now a written root rule: GPT must accept monitoring, correction and shutdown, and it may disagree only through authorized channels. And users get more say over side effects, which used to be fixed at root level. The rest, from the five values to the GPT name, reorganizes the spec more than it changes what the models do, and OpenAI has not said how it will measure whether the models follow the new rule.