They use, unwittingly, proxies for "AI" stylometry, with no specific or explicable features they can point to and say "Look! That is evidence that the text is written by AI", as if there were such features in the first place. The best they can do is validate the very lazy one-shot patterns that humans should be able to highlight in the first place.

Anything beyond that - even telling your AI to go use a "humanizer" skill in your prompt - creates a text distribution that is functionally indistinguishable from human generated writing. Throw in the fact that people who use AI will be influenced by the writing, often in positive, beneficial ways, and the software becomes a vicious, punitive tool. It's worse than waving a dowsing rod or pendulum, because the software comes with the implication of legitimacy and fairness.