This is an abuse of machine learning. Amy algorithm can be tweaked to achieve the intended result, even if you limit the dataset to buildings before the Eiffel tower had lights. It always comes back to a human.
If that's possible, is there really anything that is non-obvious and copyright-able? If it can't be done, the manipulated photo cannot be held accountable for copyright violations (but it certainly would look like it ought to). The fact that it can come down to something any human can optimize for only makes the law surrounding the copyright of a public display on a building even more absurd.