Even the basic rules turn out to be very complicated.

Core elements are often cliched, but getting from a cliched chord sequence to a complete piece/song that captures the imagination of listeners is very much harder than it looks.

I've seen an expert systems for generating counterpoint that used a grammar with more than a hundred separate production rules, and still failed to generate interesting bass lines.

Interesting music is extremely non-trivial. Conversely, trivial algos used in trivial ways reliably produce trivial output.