OpenAI Model Spec

What it is

The Model Spec is OpenAI's public document describing how it wants its models to behave. It covers how a model should rank instructions from different sources (the "chain of command"), what it must never produce, how to handle risky or sensitive requests, how to be honest about uncertainty, why it should not flatter users, and how to choose length, tone and format. The version current in October 2026 is dated 18 August 2026, and the document is released under the CC0 public domain dedication.

Why it matters for AI trainer work

Most AI trainer tasks ask you to judge or write responses against principles like these: is this answer helpful, honest and safe; does it follow the user's instructions; is it the right length; does it refuse when it should and help when it should? The Model Spec is a long, real example of such principles, and it is full of side-by-side examples showing a good and a bad response with the reason. Reading it trains the exact judgement rating rubrics test.

How to use it for practice

1. Read the overview and the chain of command section first, so you understand how conflicting instructions are ranked.
2. Pick one principle at a time, such as "Express uncertainty" or "Don't be sycophantic".
3. For each worked example, hide the verdict, decide which response is better and why, then compare.
4. Write your own short example for that principle: a prompt, a weak answer and a strong answer, with one sentence of justification.

A caution

Every company and every project has its own guidelines, and they can differ from this one. On a paid project, the project's guideline always wins. Use the Model Spec to practise reasoning about responses, not as a rulebook for other companies' work.

Best for: people preparing for AI trainer, rating or safety evaluation work who want to practise judging answers against written principles.