I know that in disciplines such as math there are very often have books with less than 50 copies in circulation. The LLM companies buy maybe 2 of a given book which are subsequently destroyed with no backup data. You're now at 48 in circulation, that is a significant reduction. Rinse and repeat a couple of million times and this starts to have a major influence on knowledge.
What's perverse about it is that the data from the books is stored extremely badly within the models. So not only are you destroying sources, you are also necessarily falsifying them.
No, even this is a very bad idea. Ideally, no books are destroyed. A close relative of mine works as an archivist. They have tons of examples of how digitising and then destroying can create extra burdens. The data on hard drives or flash devices is too fragile.