OpenAI has updated its Model Spec, the public document that describes how it wants its GPT models to behave. The new version is dated October 29.
What changed (confirmed). Per the changelog, the Spec is now organized around five values: faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care. In plain terms: do what people actually mean, avoid harm, tell the truth, respect people’s own choices, and be kind. A new section covers using context, memory and personalization in proportion while respecting privacy. Other additions cover helping users understand and choose for themselves, plus following social norms.
The changelog also says OpenAI removed most of its detailed style guidance. The rulebook now calls the model “GPT” instead of “the assistant.” The full October 29 text still lists rules such as “Don’t be sycophantic” and “Assume an objective point of view,” so flattery is still off the table.
An oversight rule. The update adds explicit guidance on supporting human oversight and correction. In the new Spec, “Support human oversight and correction” sits at the top “Root” level, alongside rules like following instructions and having no other objectives.
What OpenAI admits. The Spec says OpenAI is training its models to follow these principles, but that its production models do not yet fully reflect it.
Our take. A values-first rulebook is easier for normal people to read than a pile of edge cases. But a document is a promise, not proof. Dropping style rules leaves more room for drift, and OpenAI concedes today’s models fall short. Watch whether real behavior changes, not just the text.
What’s real here: a clearer public rulebook, with the proof still to come.
Sources
- OpenAI: An updated Model Spec for more capable and beneficial AI
- OpenAI Model Spec, October 29, 2026 version
- GitHub: OpenAI Model Spec changelog
By Willa Ravenwire · The AI Bandwagon