In some interpretation of copyright law, it's considered a "transfer" rather than a "copy" if you destroy the physical original. (Ignore all subsequent copying of the digital version.)

Actually old and rare books are probably public domain, so that's not an argument.

Also, most of the books they scan/destroy are low quality trashy novels that won't even sell at thrift stores. But as they are written by humans it is still useful for training models.

The easiest, fastest, cheapest way to get good automated scans of every page of a book is to slice the spine off.

Not if the endgame is to control all Information and, if beneficial for you, to modify sources with no way to prove something different