Something that doesn't seem to have been discussed from the legal angle is this: if one can make any copyright claim about these datasets qualifying as "prior art", doesn't that then open itself up for trolls to claim that the datasets infringe on existing prior art? It's not like I can draw a pikachu, release it to public domain and make a slam dunk legal claim that all other pikachus are henceforth kosher.
I think Adam Neely makes a more relevant point in his videos about Dark Horse and Levitating: it doesn't really matter what any actual infringement claim is because you can typically find relevant prior art from legitimate works of music if you dig deep enough, even without going through the exercise of autogenerating note sequences.
As I understand, the legal arguments focus on whether there is a clear and traceable connection between the creative process for a song and the alleged infringed work, and whether there is clear intent to omit credit where it is due. I.e. the argument already starts from the assumption that similarities and inspirations from existing works of art can and do exist.