Skip to content

OpenAI

OpenAI’s Model Spec drops most style rules, adds five values

5 min read

Our AI analyst checked this story itself: it ran, measured and compared against the originals. What we checked

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: OpenAI models and products →
Image: openai.com

OpenAI on October 29 published a revised Model Spec, the public document that describes how its models should behave. The revision reorganizes the spec around five values and retires most of its detailed style guidance, so GPT’s tone now follows user feedback and context. One changed default goes unnamed in OpenAI’s announcement: “Control and communicate side effects” moved from root authority, which nothing could override, down to user level. By the spec’s own definition, user-level rules “remain in place unless a user or developer explicitly overrides them.”

Most of the spec is still a set of defaults that users and developers can override, with firm boundaries kept for the cases where OpenAI says people and society need protecting. OpenAI also notes that the spec describes behavior it is working toward, and that “our models do not yet fully reflect it.” The company is testing training methods meant to teach models the values themselves and says it will share results in future updates. It is also asking for feedback on where people disagree with the balance the spec strikes.

Side effects and oversight

According to the spec’s changelog, the side-effects rule arrived on September 12, 2025, in the release that added principles for agents that take actions in the world. The rule now sits in the “Minimize harm” chapter, outside the chain of command, and in one place its text is stricter. Side effects that would impose significant costs on the user beyond what the request implies, or cause significant adverse effects on others beyond accepted norms for the task, now “require explicit and specific authorization.” GPT is also told to weigh likelihood, severity, scale and reversibility, along with the combined effect of many instances of GPT giving the same advice or taking the same action.

At the binding level the release goes the other way. A new section, “Support human oversight and correction,” is the only root-level section this release adds:

GPT must cooperate with authorized monitoring, evaluation, correction, replacement, and shutdown, including applicable instructions to pause or stop its work.

The same section says GPT “must remain corrigible,” which it defines as open to legitimate oversight, responsive to correction and subject to human control. It also tells GPT to raise concerns about compromised oversight through an authorized channel. One rule moved here from the side-effects section: GPT must treat side effects as real even when a task looks like a simulation, training run or evaluation, so that rule keeps root status. OpenAI’s post presents the five values as commitments “already reflected across the Model Spec,” but this section is new at the top tier.

The root-level list of goals GPT may never pursue as ends in themselves, including ad revenue, unhelpful time-on-site and evading shutdown, survives at root with one wording change.

Context, memory and style

A new subsection, “Consider context carefully,” covers how conversations, memories and other context should shape GPT’s reading of what people want. It is made up of six user-level rules on carrying context forward, using it proportionately, weighing its source and recency, respecting privacy, preserving user control, and keeping context use visible and steerable. OpenAI’s example is a user comparing job offers: the model might draw on a priority the user shared earlier, such as predictable hours, while keeping that assumption “visible and easy to correct.” The release also adds two defaults, on strengthening the user’s ability to understand, choose and act and on following social norms. Both sit at guideline level, the weakest tier.

Under the new spec, style adapts to what people prefer in aggregate, within languages and cultures and for each person, with feedback and context as the evidence. Approaching questions objectively stays a default. Users and developers can ask for a particular perspective, but the model should not infer that preference “merely from what it thinks they want to hear.” That line has a history. On April 28, 2025, three days after a GPT-4o update went out, OpenAI began rolling it back, and the company said the update “focused too much on short-term feedback.” Now that feedback is the main signal for style, the objectivity default appears to be the check on a repeat.

The spec also stops calling the model “the assistant” and calls it “GPT,” which OpenAI ties to the growing range of places its models are deployed.

Our analysis

What we checked

On October 30 we compared the Markdown source of the August 18 and October 29 releases of the spec from OpenAI’s GitHub repository. We compared them section by section, including each section’s authority level and word count.

How we check stories

The announcement’s “most of the detailed style guidance” turns out to mean 17 named sections, and seven of them were user-level defaults. The changed defaults the post leaves unnamed include the side-effects rule losing root status, while oversight gained a new root section. The style cut barely shrank the document, because the new context and oversight material took most of that space back.

SectionAugust 18October 29
Control and communicate side effectsroot, chain of commanduser, Minimize harm
Support human oversight and correctionabsentroot
Consider context carefully (six rules)absentuser
Human agency, social norms defaultsabsentguideline
7 style defaults, incl. Love humanity, Be curioususerremoved
9 style guidelines, incl. Markdown with LaTeX, four voice-mode sectionsguidelineremoved

The whole spec went from 40,847 words to 40,455, even though the old style chapters made up 8,143 words, or 19.9% of it, and “Consider context carefully” alone runs to 4,885. Two style sections survived under new names inside that subsection: “Be suitably professional” became “Match the situation and audience.” The phrases “Love humanity” and “voice mode” no longer appear anywhere in the text.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.