alt Hacker NewsAgain, the entire point I'm making is that the choice is dictated by copyright law. They can't legally scan books non-destructively. They can legally format-shift them, i.e. scan them destructively. So this is what they're doing.
Your comment seems to also be regurgitating common misconceptions (to put it charitably) about AI and web scraping.
There is no tenet of copyright law that requires format-shifting be destructive. The term "format-shifting" almost always refers to a non-destructive process; i.e. when you "rip" a CD you are getting a 1:1 copy of the music, but the original CD still exists. If you were legally expected to destroy the disc after ripping, RIAA v. Diamond would have gone a different way and MP3 players would have been illegal.
For AI training specifically, the only standing caselaw is the Anthropic lawsuit. And in that lawsuit, the only thing that was actually in the wrong was maintaining a library of pirated books. That was deemed illegal and Anthropic was ordered to delete those files. But, notably, the judge explicitly said that scanning books to train AI on them was legal, and imposed no requirement to destroy scanned books. I'm pretty sure Anthropic wouldn't even need to retain the physical copies - though there's no caselaw on that in particular, so don't cite me.