The new AI models will seek novelty[0], pieces of facts not recorded before. It will directly hone in on newness not how it was delivered. Writing styles could be another vector in which novelty could be detected but that's a different matter.

[0] https://people.idsia.ch/~juergen/sice2009.pdf

But the ability to detect, say, content generated by pre-2033 chatbots will open up a whole decade of previously contaminated content.

Which you can use to train a better model to detect AI content.

I wonder whether some models fit this - well, you could certainly call it adversarial, and obviously generative - framework?

right, you're describing a gan-like system/positional arms race. which is how adversarial tech always develops. doesn't mean we won't get better AI out of it.