I think you're misunderstanding that, I don't expect it in either case, I'm saying you have to judge the output not the input. So even if it trained on a ton of copyrighted artwork, if the output isn't a ripoff of something in the training data, I don't think there should be any copyright issues.
