The problem I find is it can take a year or longer before you learn what is idiomatic and what isn't and it's not always obvious. You can ask colleges of course but it can slow down development drastically if you need to ask often. Most times, you don't have the luxury of practicing the language for a year before you ship code so you can use AI to get a general sense of what is best practice in that language. Worst case if AI gets it wrong, it will get caught in review.

As for understanding code bases, I'm working in code bases with millions of lines of code. Before AI I would usually ask someone to point me in the general vicinity or spend 30 minutes trying to find the right ballpark to look at for a particular thing. With AI, it can usually find it in 5 minutes or less. I still learn where the code lives if I need to reference it again in the future so I don't really see much of a downside here other than that I get to save a chunk of time from having to pore through hundreds of source files.