All this really does is show how flawed the current concept of copyright is. But at some point, a huge corpus of images owned by other people was assembled to create a derivative work (the training corpus).