That is a little like saying every photograph is a copy of the thing. That is just factually incorrect. I have many three layer networks that are not the thing they were trained on. As a compression method they can be very lossy and in fact that is often the point.