Not AI moderation? I'd assume the more scalable option is having an LLM parse messages when reported.