Here's how I see the moral breakdown for aiart.
Should the AI companies have trained on massive amounts of copyrighted work without permission?
No.
Is that the fault of new media and experimental artists?
No. Unless they're actively engaged in fine tuning using the same or similar techniques for data gathering. The argument for fine tuning being copyright infringement is a lot stronger than for foundation models being copyright infringement.
At least in isolation, taking no other issues into consideration, and assuming a simple prompt only process.
Do the foundation models create things, typically, as part of known standard use cases (without using techniques to break moderation systems or induce bleed) that would fall under traditional understandings of "derivative work?"
Also no. This has been studied multiple times, and as recently as 2024.
Does this same reasoning apply to AI music?
Possibly. Nobody's studied it in a productive way, yet. The datasets are theoretically just big enough to make the same kind of generalization possible. But there are several things that could break that kind of generalizing, like duplication.
Duplication in the dataset is trivial when you've got five billion pieces of source data. When you're looking at under a million unique pieces of data, on the other hand... it's a lot stickier.
Do music files have different properties than image files, and could this change the metrics one way or the other?
Yes. Absolutely.
But nobody knows yet.
Working on it.