The arms race was first automated in 2014, when the GAN paper was published. That's what GANs do: half the network generates garbage, the other half identifies if it was garbage or real. https://arxiv.org/abs/1406.2661
It's only through careful configuration that the loss of the discriminator and of the generator are balanced to be able to improve gradually together.
It's much easier for one loss to explode while the other goes to zero, breaking the training of the GAN.
So the GAN training scheme is not a setting that holds for the real-life cat and mouse game of generated content detection. It's almost trivial to break a trained model whose weights you have.