The critic idea seems interesting but doesn't really get to the question of whether/to what degree the GAN is just interpolating the training data. It seems more of a useful tool for diagnosing GANs.