This doesn't surprise me. If a role comes with different language patterns and priorities, I'd expect the model to reflect some of those associations as well.

I have heard that attempts to insult some LLMs make LLM provider to silently replace the model with a stupidier one. How do you think, is it true?