Depressin’.
The real AI giveaway is, of course, something that you can’t use another AI to spot: the absolute lack of substance in all those effortless paragraphs of Perfectly Cromulent English.
But that would mean actually reading what people write, and that’s too much work for internet vigilantes looking for a quick power trip.
As a teenager (mana-teen?), I was just discovering em dashes and liked them a bit too much. I had a thing for big words and wrote convoluted sentences that, in hindsight, are kind of embarrassing.
Sixteen-year-old me would have been incensed if someone had thought a machine wrote my work.
Sometimes “AI voice” is actually “autistic homeschooler who spends more time around books than around other teens” voice.
Hell, you can tell just from the URL. Substack is Medium, but for newsletters, and with more Nazis.
Thanks for sharing. I was already skeptical of “one simple trick“ style approaches to detecting synthetic text, but this is an angle on the whole thing that I hadn’t considered before.
I ran a few of my old university essays through an AI checker out of curiosity to see if I would have survived university if I did it now. My first-year ones came back as definitely human, but the ones I wrote for post-grad were “likely AI”.
It’s very clear that most models were trained on academic databases. If you’re writing academically on the internet, you probably need to chill out. 
?? I don’t follow, sorry.
Thank you, geez. Like, I’m sure we’re not the first two people in the world to notice, but it sure has been hard to come across someone mentioning it. I think this kind of thing is the real essential tell, things like “does it not feel empty” or obviously disingenuous pretense at human experience.
See also the broader structural habits demonstrated in the part near the end that shows only the first four words of each sentence: chatgpt-voice.surge.sh