It'd be interesting to see if you get the same result locally. A lot of these are filters after the LLM-generation. OpenAI does it too, where if it detects a word or name, it prevents the output from getting sent to you (mostly due to a bunch of legal cases where it spits out pure garbage about certain names).