Great point! I think with captchas, some humans might be able to identify obviously made up facts and very poor sentence structure to tag a text as low quality. But I don't think you'd be able to give a reliable assessment on whether a text was AI generated or not. Especially if it's a random group of people (unfamiliar with modern content generation) filling out the captchas.
Captchas can be made adversarial. Not sure if it's a good idea to try that with text since we don't know how humans react. Maybe that's what the Phenomenon is about?
