> The problem is that neural network models, and their outputs, are not copies of the original works. They are a set of probabilities (weights) that are trained

> The problem is that neural network models, and their outputs, are not copies of the original works. They are a set of probabilities (weights) that are trained based on thousands or even millions of data points. True, but in order to train these models you have to copy and save images that are for the most part copyrighted. So while the models are not copies of images, they are definitely derivative works.