I corrected my way into the tell
I built a tool that detects AI writing. The first thing it did was call me a fake.
Then it got worse. The paragraphs it flagged hardest included ones I remembered writing start to finish, alone, years ago, with no machine anywhere near them. That led to some sleuthing, and it hit me that I have an entire folder of writing sitting on my hard drive from back when the fanciest AI could barely parse the word “the.” First I found my blog archive from 2008: the one place a language model had never been.
The tells were already in there. The balanced “not this, but that” sentences. The rhythm of threes. Even the em-dash — that one, right there, the mark I get accused of most. Except in 2008 I didn’t know the key existed, so I typed a little hyphen with a space on each side, groping for a pause I couldn’t punctuate.
A blog is easy to dismiss, though. That’s just me alone in a CMS with my bad habits intact. So I dug deeper, remembering that I have the original edited manuscripts of dozens of tech books I wrote.
One such book is Head First JavaScript, which was run through O’Reilly’s copyeditors and proofreaders, an apparatus whose entire job is to catch what’s wrong with a sentence. Open chapter one. The Web (yes I capitalized it), I wrote, “doesn’t mean to shut you out, it just doesn’t know better.” It-doesn’t-X-it-just-Y: the highest-signal AI tell of 2026, in print, professionally blessed, shipped before the iPhone had an App Store. A page later, three clauses in a row, because of course.
The editors didn’t miss those. They’re really good. They left them because they were never errors. They were the voice.
Here’s the part that gets me. I only learned what an em-dash was because of the AI panic. I read that it had become a tell, went to find out what one even was, and discovered it was the mark I’d been reaching for my whole writing career. So I learned the keystroke. Writing it correctly made my prose read more like a robot, not less. I corrected my way into the tell. Thanks AI!
The machine didn’t teach me any of this. It learned these moves from an enormous pile of people writing this way for centuries, me somewhere in the pile, and now hands them back so uniformly that the originals look like the forgeries.
New essay, with receipts dated 2008:
Read the essay
The Tells Were Mine First
Article 7 of 23 so far in The Automation Line
I built a tool that detects AI writing, and it called me a fake. The honest test wasn't my AI-assisted essays — it was my pre-AI archive. The tells were in there, dated 2008, including the one I'm most accused of.
Read the essay →Becoming Gnarly, by email