Comment by tikkun

I expect they're using the moderation filter (https://platform.openai.com/docs/guides/moderation/overview), but calling it in parallel to the generation so that it doesn't add latency to the response.