Is It a Hard Argument for Fair Use?

Most of the AI training data sets, they're copying in a notional way. It's not like Napster, let's say 20 years ago, when music was being copied and then distributed out. The copying is there to train on and then they're producing something totally different. So it raises a question about whether this is something that's just comfortably within fair use as we've understood it or maybe requires an elaboration of the idea of fair use for new technology.