>> In particular, the algorithm points out that Norman Rockwell's Shuffleton's Barber Shop painted in 1950 is remarkably similar to Frederic Bazille's Studio 9 Rue de la Condamine painted 80 years before.
Not at all. Apart from both being of (different sized) rooms painted from an approximately similar angle, there really is nothing else that is the same about the two paintings. It would appear that the computer is keying only off of very large features such as a general observation that a large lightsource of a simliar size and location is in both (but in one painting its half a window which is really a secondary subject, and in the other, a doorway to a room with a light in that is the primary subject). If the computer can only make decisions based on such broad generalisations, it really is pretty much useless.