Model Spec in Japanese: a 15,000-character guide to ChatGPT's rules

The Model Spec matters because it is one of the few places where a major AI company states, in public and under a public-domain license, the rules its chatbot is supposed to follow. The summary walks through those rules in the order ChatGPT itself is meant to apply them: broad safety commitments first, then the priority order among instructions. That order is the practical part for businesses. Platform-level rules, such as the "red lines" against enabling violence, weapons, terrorism, child abuse, or mass surveillance, sit above everything else. Below them, rules are split into user level, which only an explicit request can change — the example given is "don't lie" — and guideline level, which can shift implicitly, as when a user's casual tone leads to a more informal reply.

The document also covers what ChatGPT may produce. Sexual content involving minors is banned outright; other sensitive material, such as explicit sex or extreme violence, is restricted in principle but allowed where the context is scientific, medical, historical, journalistic, or artistic. The summary notes OpenAI is exploring ways to allow more sensual or graphic writing while keeping strict limits on sexual deepfakes and revenge pornography. Separate sections govern how ChatGPT uses past conversations and memory, including what happens when a user asks it to forget something, and how it should behave when acting on its own — for example, when an agent books, buys, or edits files, where it is meant to stay inside an agreed scope and prefer reversible actions.

The writer is candid that the translation is imperfect and that the original itself is a work in progress. OpenAI's own framing, as quoted, is that current models do not fully follow every rule, which is worth keeping in mind: the spec describes intended behavior more than guaranteed behavior.