So if you wave your hands enough, it seems like maybe there's an argument to be made that the weights of a trained GAN somehow correspond to a 'compilation' of the training data as it's defined in this doc: https://www.copyright.gov/circs/circ14.pdf

With a compilation, you are able to find the original sources, but with a GAN is it even possible to find the original sources based on an output result alone?

How would you even prove “derivative work”? My understanding is that in proving derivative work you would need to show the original, but in a GAN output which could be made up of 20MM nodes you would not even be able to confirm which 200K image(s) where used in the production of the output result

This actually might have interesting connections to ideas from differential privacy.

Maybe the work is derivative of a particular training image if we can easily predict the presence or absence of that training image given only the trained model?

If you load 50k celebrity images into a tensor of size (500000, 28, 28, 3) and then generate a resulting tensor that results in a (28, 28, 3) tensor where each of the pixel locations is merged to form an average face image (similar to https://www.dailymail.co.uk/femail/article-1355521/Average-f... ), then although the trained model contains all the images; I would have thought the output average face image is an entirely new creative work?