A set of rules telling developers which kinds of text, images, and code are not allowed when using OpenAI models.
Developers creating content-creation workflows read this policy to ensure compliance before launching new features.
01What it is and how it works
OpenAI publishes a public policy that classifies content into allowed, restricted, and disallowed categories. The policy is enforced by automated filters that scan prompts and outputs for keywords, patterns, and context. When a request matches a disallowed category, the API returns an error and logs the incident. The policy is updated regularly, and developers are expected to check the latest version before launching new features.
It is a list of things you cannot ask an OpenAI model to create or share.
02What to do about it
Review the policy page this week and map each disallowed category to your current use cases. Add a checklist to your content‑creation workflow that flags any prompt containing hate, self‑harm, or illicit activity. If you find a risky prompt, rewrite it to stay within the allowed scope or use OpenAI’s moderation endpoint to pre‑screen user input. Document the changes in your team’s style guide so future copywriters have a reference.
03How it is measured or noticed
OpenAI provides a moderation API that returns a confidence score for each policy category. When the score exceeds a threshold, the request is blocked. In the OpenAI dashboard you can see a log of blocked calls, the category that triggered the block, and timestamps. Monitoring these logs lets you spot trends—like a sudden rise in political persuasion attempts—and adjust your prompts accordingly.
04Common mistakes
- Assuming “political content” is always allowed; some political persuasion queries are restricted.
- Relying only on human review and skipping the moderation endpoint; automated filters catch patterns that humans miss.
- Copy‑pasting user‑generated text without checking it against the policy; user input can contain disallowed material.
05Limits
The policy does not apply to internal testing environments that are not exposed to end users, but any public‑facing product must comply. It is often confused with OpenAI’s Terms of Service, which covers billing and account usage; the Content Policy is solely about the nature of the generated content.
06Worked example
"Disallowed content includes instructions or facilitation of violent wrongdoing, self‑harm, or illegal activities."
Frequently asked questions
How is the OpenAI Content Policy different from a typical terms‑of‑service agreement?
It is not a general terms‑of‑service; it specifically classifies content into allowed, restricted, and disallowed categories for model usage. The policy provides concrete moderation rules and a moderation API, whereas a TOS usually covers broader legal responsibilities.
Should we adopt the OpenAI Content Policy for our internal prototype that will later become public?
It depends on your launch plan; if the prototype may become public‑facing, you should align it with the policy now. Internal testing that never reaches end users is exempt, but planning ahead avoids costly re‑engineering.
How do we actually check if a piece of user‑generated text violates the policy?
You send the text to OpenAI’s moderation endpoint, which returns a confidence score for each category. You then compare the scores against your own thresholds to decide whether to block or flag the content.
Does the moderation API still work for new policy updates?
Usually it does, because the API reflects the latest policy version automatically. However, you should monitor OpenAI’s changelog for any breaking changes that might affect your thresholds.
What happens if our product serves disallowed content without catching it?
It can lead to suspension of API access and potential legal liability, and you’ll likely see error responses from the API. You would notice the issue when OpenAI sends compliance warnings or when users report prohibited material.
How long after we start using the moderation API can we see reliable confidence scores?
It is immediate; the API returns scores in real time for each request. You can begin measuring compliance as soon as the first call is made, though you may need a short period to tune thresholds.
Asked out loud
spoken, not typedThe same term in the words somebody uses speaking to an assistant rather than typing into a box — written from the situation, which is why each one carries the situation it came from.
Yes, you can use OpenAI’s moderation API which returns a confidence score for each policy category in real time. Just send the user’s message to the endpoint and act on the result before the response is sent.
Usually you integrate the moderation call right after the model generates a reply and before you play it back. If the score exceeds your threshold, you replace the reply with a safe fallback.
It depends on whether the copy includes disallowed content such as hate speech, illicit instructions, or personal data. Run the text through the moderation API; if any category returns a high confidence, rewrite that portion to remove the offending material.