That's not even the real problem. It's a problem, yes, but not the real problem. The problem is that before they could train the model on the book, they had to copy the book from somewhere. Is it ok to make illegal pirated copies of a copyrighted book to train your model? I think that's the issue we are dealing with here.
Whether it is ok to create a derivative work or not is beside the point.