My guess is that their racism filter is too coarse and so it treats "jokes about imams" like "jokes about Black people" as opposed to "jokes about nuns". I find it hard to believe that OpenAI would tune their model to avoid offending extremists.