@Snarf Zagyg
Thank you for your detailed input.
My take- this cuts heavily against AI Firms, given that it's a commercial use. Mitigated by the argument that it's transformative (they are not simply cutting and pasting, but using it to train).
Would it help, on the other hand, open-source models that are distributed non commercially (Mistral or Stable Diffusion come to mind), though they might not care about US law)?
3. Amount and substantiality of the portion used in relation to the copyrighted work as a whole.
This is arguably the worst factor for AI Firms, because, they are using the "whole" work. That said, there are two things to consider; while difficult, it is not impossible to find fair use of an entire work. In addition, I think that AI Firms would argue that they aren't "using" the whole work, given that the outputs are constrained to not regurgitate the whole thing.
Actually, it doesn't contain any part of the copyrighted works. The model itself is a mathematical result, so they could claim that their is no similarity between the originals (the books) and the result.
4. Effect of the use upon the potential market for or value of the copyrighted work.
Here, we are looking at the damage to the market for the original work done by the alleged unlicensed product. I think that the AI Firms have a decent case in terms Books3 on this factor ... I am far less certain about how this would work for some of the image generating AIs, which I believe can mimic the styles and outputs of certain artists, and arguable would affect the existing or future market for the copyrighted works. Honestly, you could probably make the same argument for Books3 at a certain point, to the extent that the corpus contains enough of a writer's output to mimic the writer and damage the market for their books. So ...
I think the argument that it can mimic the style of certain artists is giving too much credit to AI generative ability at this point (I am not speaking for the far future, like June or July 2024). It has been noted that people used artists names in the prompt to modify the output, but it's far from being able to replicate a style. I've tried with several, if you add Hopper you get a chance to find a diner in your aerial view of the orcish capital, if you type Van Gogh you'll get a weird sky that's far from being Van Gogh's. "by Rutkowsky" might get you a generic fantasy style, but it wouldn't be different than prompting for "D&D style" or "conan style".