There is uncertainty as to whether the underlying work itself.... If it is a digital copy of a book, it would be a literary work regardless of what stage. Think of natural language processing, for example. You would want to teach a system how to understand a sentence and context and meaning and sentiment. You would probably want to give it a lot of different books and a lot of different texts to figure that out. That would all be part of the training process, so when you fed it something in real time, and not in the training stage, it could figure out the context of that sentence and the meaning of that sentence.
In the training process, it's actually unclear, under the current act, whether that activity would be an infringement of copyright or not. There's uncertainty there, so that often prohibits using works that a company might want to use for that training process. They're not releasing those works, and they're not a substitute for the digital text, but they're using them to learn how those sentences are structured in that process.