It sounds like after the model is trained, it also pulls from the training database at generation time. Can someone explain how this isn't a kind of "cheating"

It sounds like after the model is trained, it also pulls from the training database at generation time. Can someone explain how this isn't a kind of "cheating" at generation? Happy to be wrong.