Is there a case that actually says this? Why would whether something is fair use depend on that? For that matter, how would they even show that a given AI model was trained on something from a recursive crawler rather than the same articles added to the training data after being downloaded by hand?