If I remember correctly, it is a claim from the WGAN-GP paper. They have a modified way of how data is sampled that have a higher correlation. That is why they don’t have batch normalization. It is in the context of WGAN-GP but not WGAN in general.