I'm not an English native speaker and my first language has only about 2 million speakers with very limited training corpuses. GPT-3.5 is already shockingly good at it, although the outputs do seem a bit less factually accurate and less practically useful than in English. It's really odd that it gets the language right in all aspects except the content.

The great thing about niche languages is that it's not cost effective to scammers to learn it for their purposes. That all changes now I suppose.