Training neural nets is a novel transformative process - it only takes gradients for each training example, it doesn't save any image, just take the gradients a

Training neural nets is a novel transformative process - it only takes gradients for each training example, it doesn't save any image, just take the gradients and sums them up overlapping the influence of all the examples, mixing up everything in a way we can't tell which feature comes from which training example. More like inverting the dataset into abstract principles and then generating top-down novel examples. It's like learning, but differs in that humans can't possibly see so much data. It's not like copying, but it can generate copies of training examples as an edge case. It feels like stealing, but you can't assign who is the victim, maybe it draws a bit from multiple sources of inspiration. It's very creative and inspires all sorts of idea juxtapositions, but some people consider it's only parroting known styles. In the end I think we will want to use the new toys so we will allow training on anything. Maybe a "Do-not-use-for-training-neural-nets" flag will be observed by the conscientious labs.