What's the "sorta, kinda" as you understand it? Has any judge, or law, created a requirement to destroy copies?
From what I understand, depending on the purpose, making digital copies may be fair use if it's "transformative". In a decision involving Google's book search feature, the court's decision highlighted that the amount of text displayed in the product was limited and not a "significant market substitute" [2]. The AI firms aren't trying to show you a digital copy of the books, and in a bunch of contexts already, if their product regurgitates too much verbatim text, that's going to be a liability (e.g. educational uses and plagiarism). I've also tried prompting LLMs to finish a paragraph from inside a novel from a well-known author, and they have not been able to get the next sentence typically. You can't easily and with confidence just extract back a book. I have not tried the experiment of seeing how many tokens you need to spend before a books is reproduced and what the cost is relative to the original work -- and you would already need the digital text to recognize when you had succeeded.
What's transformative? Previously the courts have said that scanning to offer a search product was transformative, given some limitations on how much of the matching book is shown to the user. I think there's a decent argument that training an AI model is at least as transformative as search indexing.
https://en.wikipedia.org/wiki/Authors_Guild,_Inc._v._HathiTr...
https://en.wikipedia.org/wiki/Authors_Guild,_Inc._v._Google,...