The paradox is analogous to AI poisoning its own training data as more and more AI generated content is released on the Internet. Indeed, I see a hard terminus for both man and machine at some point.
Absolutely right. First gen models are trained on 'virgin' non-LLM output. Subsequent models are tainted by ingesting AI replies. Rinse and repeat, and all you end up with is AI copies of AI replies, and the noise will completely overtake the signal.