Complementing sibling, consider Google Books. This is where the question first arose: if one puts a book through a scanner, non-destructively, then surely they have made a digital copy of the book? There's the related question: if the scanner destroys the book, then that surely means no copy? The bounds of this were tested with the concept of CDL, which courts did rule against in Hachette v. Internet Archive; they said that CDL is clearly copying. But they also said in HathiTrust that digital preservation is transformative. So preserving is possibly fair but copying is probably infringing; in general one can have a private library but they can't copy it out to other people.

Hopefully it makes a bit more sense from that perspective. Copyright's still stupid as fuck, though. Previously, on Awful, I made a prediction:

I do think that the resulting incentive for model-trainers is not what anybody wants, though; Google Books is still settled and Kadrey didn’t get updated, so model-trainers now merely must purchase second-hand books at market price and digitize them, just like Google has been doing for decades. At worst, this is a business opportunity for a sort of large private library which has pre-digitized its content and sells access for the purpose of training models.

Alsup put this as a point in Anthropic's favour in the recent authorial class action, so now it's received wisdom.