If it were memorizing training data, it would due poorly on a Wasserstein critic trained on a validation set drawn from the same data distribution.