The training argument is probably going to come up dry by the time the court works its way through expert testimony, as the underlying argument for training as infringement is insane.

But where OpenAI is probably in hot water is that torrenting 100k books in the first place runs afoul of existing copyright legislation.

Everyone is debating the training in these suits, but the real meat and potatoes is going to be the initial infringement of obtaining the books, not how they were subsequently used.