It probably doesn't even matter for AI research. I don't know any professional AI researchers, but I bet reading arxiv and reproducing other's work, and publishing your own work is how you go about that. Why would you do coursework when there's papers to read?
_Maybe_ you don't need that much CS per se (debateable, but an argument can be made); however you do need a lot of math coursework.
Also, it's far easier to grasp the basics of AI from a course than to reconstruct "basic body of knowledge" just from reading arxiv. As a matter of practical considerations, I just don't believe anybody can become an expert by reading only the research, without going first through the basic training. It's too damned difficult, too much work (and pointless, too - people did that work for you already and built great courses with the summaries, why not take advantage of that?
