Well, from the pure GAN theory perspective, it is theoretically possible for a perfect generator to exist which generates samples that perfectly match the 'source data distribution' (the only question is how do we obtain one or close to it), and thus a perfect discriminator can't exist - a fake image could be realistic and indistinguishable in all aspects.

And it does seem that the human-visible weaknesses of many currently published deepfake examples are those that would be hard for the current discriminator architectures to detect, not those which would be harder to generate than any other part of the fake.