The training models do contain a heavily compressed version of the art used as training materials. This is easy to prove; if you prompted Midjourney to create a picture of a green eyed Afgani woman, it spat out something clearly close enough to the National Geographic photographic to constitute copyright infringement.…