It seems plausible that someone could make a decent discriminator for distinguishing human text from AI text. I wouldn't immediately discount the idea.