But that’s not how model training works, it doesn’t simply copy and paste entire songs into its training data. It more or less “listens” to it, analyzes it and when you ask to create a rock song for example it just has an algorithm behind it what a song like that would sound like.
But you can’t just ask it to generate Bohemian Rhapsody from its data, it would probably get very close depending on the training, but it would never be 100% the same (except the model was only trained on this one song).
Just like you can listen to rock songs and then make your own, that’s totally valid. The problem here is of course automation and scale, but saying it’s not fair use is dubious.