Machine learning models of text are generalizations of what the text represents. A generalization being copyrightable seems like a bad idea, though, I don't think the legal system has really decided. In my opinion, owning a generalization is like stating that Apple should own all color gradients because they used them predominately in their advertising. It seems to cover too much, but, Apple probably does own copyright on final created art pieces that use uncopyrightable gradients to create something.