As long as the moderation follows their rules, and it is always as transparent as shown in this example, I don’t see an issue with this.

My only concern is that LLMs are very good at recognising biases in questions and are more likely to confirm them than push back. So the LLM might pay too much attention to small/possible infringements. But this depends heavily on the model, the prompt, and the reader.