Five values AI should uphold
In the October 29, 2026 edition of its Model Spec, OpenAI reorganized the content around five values: fidelity to human intent, minimizing harm, honesty and transparency, human agency, and care. Most detailed style guidance was removed.
Key facts
| Edition | October 29, 2026 |
|---|---|
| Document | OpenAI Model Spec |
| Five values | Fidelity to intent, minimizing harm, honesty and transparency, agency, care |
| Authority order | Root, System, Developer, User, Guideline |
| License | CC0 1.0 |
What happened
The Model Spec is a public document describing how OpenAI models should behave. The October 29, 2026 edition reorganizes the whole document around five values. This edition carries a notice that a newer edition exists, so check the document's front page for the current version.
The five values
- Fidelity to human intent: understand what users and developers are trying to do, even when the goal is unclear.
- Minimizing harm: keep safety boundaries and weigh the consequences of actions in advance.
- Honesty and transparency: report facts, uncertainty, capabilities and actions taken as they are.
- Human agency: help people judge for themselves, without hidden nudging or manipulation.
- Care: consider the people affected by the work.
Whose instructions come first
The document sets an order of authority for instructions. Instructions inside web pages, files and tool results have no authority by default and are treated only as information. So the principle is not to follow a command hidden in a page.
What changed
These points come from the public changelog.
- Restructured around the five values, with a revised overview and authority order.
- Changed the document's subject from "the assistant" to "GPT".
- Added principles on supporting the ability to understand, choose and act, and on following social norms.
- Added guidance on human oversight and correction.
- Added context items (memory, personalization, privacy, user control).
- Removed most detailed style guidance.
Why it is worth reading
Items such as "does not adopt OpenAI's revenue or engagement time as its own goal" and "cooperates with human oversight and shutdown" are published in the document. Our reading: when a model answers oddly, you can use it as a reference for whether that was the intended behavior. Still, the document states goals; it is not a guarantee that the model always behaves that way.
How to use it
- Starting with each section's compliant and violating examples is the fastest way in.
- Developers can use the authority order to design a clear separation between system instructions and user input.
- The document is released under CC0, so it can be freely quoted and reused.
Sources
Questions
What is the Model Spec?
A public document describing how OpenAI models should behave.
What are the five values?
Fidelity to human intent, minimizing harm, honesty and transparency, human agency, and care.
Does the model follow instructions inside web pages?
Tool results and web content have no authority by default. Instructions inside them are treated as information.
