The scan and destroy method is what a judge allowed to do in order to have a copy of the book in the training dataset. With the physical copy destroyed there's still only 1 copy in circulation. Once "inside" an LLM I don't know if anyone decided unequivocally that it's copyright infringement or not, and if that counts as a second copy.

There's no technical reason why an LLM couldn't reproduce verbatim some of the training material. It's sort of a lossy statistical compression engine. Enough of the info will survive to the output in the original form. With the amount of data and the commercial nature it's hard to argue fair-use. But nobody tested this in court. I'm not even sure the US wants to ever test this. Why even attempt something that has a non-0 chance to sabotage your most promising industry/bubble in ages?