Whether something is derivative or not is one of the key questions used to determine whether the free use of someone else’s copyrighted work is fair, as in fair use.

I think training an AI model is not fair use. It's either derivative work and needs a license or it's not derivative work and can be used without a license. In both cases it's not fair use (in the legal sense of "fair use").

I'm not sure if you're making an argument about what the law currently says or what it should say. In my opinion the law should be updated to clarify if you need a license to use copyrighted material as training data.

The amount that artists would be paid would be determined by negotiation between the artist (the rights holder) and the entity using their work

Sure, my point is such an agreement will never be made. It's a good deal for AI companies to use the data for free, but if they can't do that, they will not be interested.

Either way, I think there is no way for artists to win this. It's completely possible to train large image generators without copyrighted material. These datasets are so large that paying artists per image will never be feasible.