Assuming it is really using AI to generate tracks, that AI still needs to be trained on existing music, and that music has to come from human contributors.
The way typical stock music, image, and video libraries work, is people upload contributions and earn a royalty whenever someone downloads one of their works.
In theory, bolting an AI engine onto such a system simply enables more variations in output, which is fine, AS LONG AS the original human contributors still get compensated when their works are used to generate the AI output.
In effect, it's essentially no different than existing royalty-free libraries, except that the AI engine is able to stretch the library's source content through recombining. As long as they're transparent about their use of source materials and pay the relevant contributors, this is pretty much business as usual.