Well here's an interesting thought. Let's say you train a generative network on illegal imagery. Is the combination of network and weights illegal to posses? What about images it generates?