It's not about understanding how the sound is physically, or electrically generated. It's not that far down. It's about controlling the music, every note, every instrument, etc. About directly controlling the outcome.
Electronic drums are often contentious, e.g. But if there's a person in control of every percussion stroke, it's just an instrument to me. All good.
Even if the drums are completely programmed, there's still a person behind it who has a vision and a desire to have it sound exactly like that.
It's not randomly generated where the "artist" presses a button and really has no idea exactly what the outcome might be.
With regards to your idea of generating some sort of scaffolding for a song and then continually editing it, I have to agree to disagree on that point. I have a problem with the origin of the starting point.
Even if you make a very close cover of a song that you didn't write yourself, it's still a work that had a basis in human inspiration and composition. Generative AI is not a human, even though the training data is likely of human origin.
Ultimately it's the value we place in the amount of human input versus non-human input, in the case of works of art like music and imagery. I am not interested in non-human works. The ratio between the human and non-human input needs to be high enough for me to be interested.
Just like in sports. If someone sets a record but was high on enhancement drugs, it doesn't count, or doesn't make it as impressive. To me at least.
Hopefully my standpoint is made clear here.