It's my understanding that, with LLMs, being polite can net better results because they're trained off human interactions where being polite is better.

It's not that they care at all, it's just how language works.

I've not seen a study of the phenomenon or even tested it myself, but it seems plausible.