I mean none of the major LLM/DL developers would scrape the entire web (even against robots.txt and paywalls) just to increase the amount of data they could unprofitably train with, because that would expose them to copyright litigation... Oh, wait yeah they did.

Frankly, the idea that no app or company is doing this or selling the data would require an enormous conspiracy of secret suppression. The argument that one can't easily hash compressed sound fragments in a way that matches for copyrighted music or generic advertising terms and send that data back linked to a location/ID is sort of laughable.