I’m aware that sounded like I had a stroke.
All the AI slop on the internet sounds like the same person: confident, helpful, corporate, and allergic to a sentence that runs longer than ten words. It’s a decent writer, but it doesn’t sound like you, and that’s fixable even though I’d bet money you don’t know how to do it yet.
When you hire your average content agency, they hand you the same content they make for everybody else, maybe with a little industry vocabulary sprinkled on top, and the sentences still read like a machine wrote them. You’ll see an overabundance of words like “load-bearing,” the little inverted flip where they tell you this isn’t X, it’s actually Y, and every other tic that announces a language model did the typing instead of a person.
There’s an actual mathematical way to study whether a piece of writing sounds like a particular human. It was built to verify authorship, to prove who wrote something, but you can run it upside down and use it to force your own AI to sound like you instead of like everybody else.
It’s called stylometry, from the Greek words for “pen” and, I presume, “please stop sounding like a robot.”
It’s less exotic than it sounds. Stylometry just measures the fingerprint of how a person writes: how long your sentences run, how often you reach for a contraction, whether you touch a semicolon or never do, which words you lean on, and the tics you don’t have. Forensic linguists use it to guess who wrote an anonymous letter, and I started using it to work out why my own AI drafts kept feeling subtly wrong.
The surprise wasn’t what the AI added. It was what it took away.
I pulled about a dozen essays I’d written years before ChatGPT existed, roughly twelve thousand words of them, and ran the numbers on the lot. My median sentence ran sixteen words, but the spread around it was enormous, with three-word jabs sitting right next to thirty-word rolling sentences that were built almost entirely out of small, common, forgettable words. Contractions everywhere. Semicolons almost never. About one sentence in nine opened with a flat little And, But, or So.
Then I ran the same numbers on the stuff I’d drafted with ChatGPT and tidied up by hand, expecting it to come back longer and more flowery than the real me. The opposite had happened. My sentences had gotten shorter and far more uniform, the variance had collapsed in on itself, and the long rolling sentences had vanished from the page entirely. My writing hadn’t gotten fancier with a robot in the loop, it had gotten metronomic, the same tidy beat landing on every line, and that flatness is the tell even though I’d have sworn the drafts read fine.
There’s one tic I hunt for in everything now. The flip. “That isn’t growth. That’s applause with overhead.” You state a thing, you negate it, and then you hit the reader with the upgrade, and it reads punchy as hell, which is the exact reason almost nobody catches it in the wild. Across years of my own pre-AI writing that two-sentence flip turns up about once. In a single batch of AI-drafted posts I counted it eleven times. Once is a voice. Eleven is a machine doing a confident-guy-on-LinkedIn impression.
So now every piece of AI-assisted copy I send runs through a script before it leaves my desk, where it gets scored against my own fingerprint and flagged anywhere it has drifted away from me. If the median sentence has crept down to seven words, I go back and let something run long. If a semicolon has snuck in, I kill it on sight. If the flip shows up more than a token once, I rewrite each one as a plain statement. The whole pass takes a minute, and it catches the things my eye glides straight over, because my eye has been trained by the internet to read punchy AI copy as good copy. The numbers haven’t.
Here’s what I’d tell anyone trying to get all the way there with clever prompting alone: you can’t, and I’ve tried every flavor of “write like me, keep it casual, short sentences, make it punchy.” What comes back is a slightly louder draft of the same confident stranger. The model has no idea what you sound like, and if we’re being honest, neither do you, at least not in a form you could hand to a machine. A vibe isn’t a spec. “Sixteen-word median, wide variance, zero semicolons, thirty-six contractions per thousand words, one earned reversal per piece” is a spec, and once you hand the machine numbers like that it stops guessing about you and starts hitting the mark.
The odd part is what all this measuring did to me as an editor of my own work. I used to fix sentences by feel, and feel is the exact sense that polished AI copy is built to fool. Now I know which knob is loose before I’ve touched the page. The confident stranger living in your AI drafts has a fingerprint, and so, whether you’ve ever bothered to measure it or not, do you. The only job that matters is making sure the draft going out the door is carrying yours.
Want to know more? I’m happy to talk about how I can make your copy sound like you.
