OpenAI has a moderation endpoint for detecting "bad" content that acts on text (https://platform.openai.com/docs/guides/moderation), curious why they aren't using it.