Taste is still what separates AI output from human output regardless of whether it's music or software, IMO. Thinking in modes vs. functional harmony, how you hear a chord in context, which voicing you reach for etc. are things that can hardly be abstracted into strict music theory, let alone pattern-matched.

I imagine you can get pretty close to replicating an artist's sound if you specifically train on their output, but at that point it's the same question of whether you're satisfied with abstracting away the things that make writing music/software interesting to us (with music the answer's a lot more obvious since the point of writing music has always just been... to write music).