It doesn’t matter. Even the training process makes it pretty much impossible to tell these things apart.

And if we do find a way to distinguish, we’ll immediately incorporate that into the model design in a GAN like manner, and we’ll soon be unable to distinguish again.