Show me AI text that manages to climb out of the uncanny valley, and I'll show you AI text that's been edited by a human.

The problem is the well's been poisoned just by the fact that I know this is AI trying to hide AI, so I'm already poised to look at the examples and declare "aha! this is obviously AI!" Moreover, it's not single sentences or phrases that make AI text stick out (though obviously those are the biggest tells), it's the text taken as a whole. When you read the full output example in that repo, it seems obvious to me that it's AI (though again, it could be the poisoned well). This is the uncanny valley I was talking about; something is just off about it.

I agree that it feels off and I wonder what I would have thought if I'd seen the "after" example without knowing it's AI output put through a humanizer. Would I think much about the weird use of the word "honest"? About "that's the Lisbon I kept thinking about, not the castle"? Or how the story feels very impersonal somehow, with the author just mentioning their calves and legs sometimes as the only way of convincing the reader of their humanity?

Also, the 'before' segment didn't contain any mention of custard tarts, football, crowded trams, mixed feelings, etc. The original had a very positive travel agency type of tone, which was replaced with a lot of very odd sounding, imperative phrases that sound like engineering-speak. ('earn the fuss' 'build trips around pastry') I'm not convinced that this thing is actually meeting its design goal of not hallucinating shit.

The readme feels AI generated

I am not claiming to have a perfect AI classifier. That is an unnecessary claim that distracts from the broader point.