I think if we replaced "AI" with "taking averages over subsets of historical examples", then there'd be no mystery for when "AI" will be good or bad at anything.

Would we expect a discrete melodic structure to be expressible as averages of prior music? No.