They need to be publicly available and differentiable to be able to use as an opponent for a GAN.

Pretty much nothing except other neural networks are differentiable, unless you put effort into designing them to be.

It's feasible to design a reinforcement learning based network that uses output from a non-differentiable deepfake detector as a component of the loss function.

And can you imagine the impact of false positives? Prosecutor has the perp on video committing a heinous offense but it's still insufficient evidence because of a software bug.

We can dream of juries that wise... haven't seen one yet though have we?

Which means we would have to rely on an "Authority" who cannot be independently verified, because we do not know the secret sauce they are using.