There's basically an inexhaustible supply of human-created tracks that can be accessed for next to nothing
You train an AI on that, in order to create something that combines all of the best parts that you want. If anything, I think AI music is the natural progression of innate human desire to leverage and "stand on the shoulders of giants" to create something bigger from smaller pieces.
Which is of course nonsense because LLM is from definition unable to bring in something new. It's not standing on shoulders of giants, it's just making endless copies of them.
That is trivially untrue, even if we ignore the misnomer of trying to use a language model for non-linguistic audio file outputs. I can assure you there was no reference material of say Sam Altman getting arrested when he is getting caught stealing GPUs from a shelf of a BestBuy. (One of the uses of SORA.)