What an AI-text researcher at FIU says about why AI writes the way it does, and why telling it to write like a human fails.
In May 2026, people accused Nike of using AI over a single em dash in a social media post about the tennis player Jannik Sinner. I don't use em dashes in anything I publish. That's a style choice. Telling AI to write like a human does nothing, because there's no style called human to copy.
The hunt had started earlier. On 9 April 2025, The Washington Post ran a piece by Daniel Wu with the headline "Some think the em dash is a 'ChatGPT hyphen.' Writers disagree."
The dash is a symptom
Mustafa Ocal researches how to spot AI-written text. He's an assistant teaching professor at FIU's Knight Foundation School of Computing and Information Sciences, and in a 2026 FIU News interview he put the cause in a few words. "AI doesn't think. It predicts."
You can run his test yourself. Open iMessage, type "Tomorrow I will go to the" and look at the three words it offers. Ocal gets the same three. So do a lot of his students. That's a machine guessing the next word from everything it has seen, and the big models do the same thing at a far greater scale.
The dash comes from how they're trained. Engineers score the sources a model learns from. In Ocal's example a celebrated novelist gets a 10 out of 10 and a college student's blog gets a 3. The top scorers, novelists and journalists, use em dashes constantly, so the model learns that a dash is what good writing looks like.
Working writers are frustrated by this. They say they've always used em dashes and AI copied them. Ocal says they're right. He has a colleague who keeps using the dash anyway.
Words work the same way. Ocal says he has never heard a person announce they were about to "delve into" something. The models reach for it anyway. So they do for "a myriad of" and a whole family of fancy openers, at rates almost no human writer would.
People ask for the human version anyway
The interviewer asked Ocal for the simplest fix for these AI-isms, and the answer came in two short lines. "You could tell it to write like a human," he said. "I see a lot of people doing that."
I see why it's tempting. It's one line, and it feels like it should work. A model that predicts can only answer it by guessing what human-sounding text most likely looks like, and the likeliest answer is the stock one.
OpenAI recently made its models better at obeying a no-dash instruction, Ocal says. The default hasn't changed much, because the training data hasn't. Forbid the dash and you push the model away from its most polished phrasing, so the writing can come out flatter.
The tell is that nothing happened
Everything above this is Ocal's research. This part is mine.
Read a page that sounds like a machine wrote it and ask what in it happened. A name and a date. A number somebody counted.
What makes writing read as machine-made is that nothing in it happened.
A model can polish a sentence about a fact, but it can't know yours.
You can't fix it afterward. Swap the dashes for commas and a paragraph with nothing in it still has nothing in it.
It goes in before you type
So the work comes first. Put in what a client said when they pushed back, and the number from your last job. Put in the advice you give and the advice you'd refuse to give.
Take this post. Before I wrote a word I had a news article, a date and a researcher's name. That's what holds it up. Without them I'd have a smooth page about AI writing and nothing to say.
You can type all of that into every new chat, and it gets old fast. There's a fix for that. But it's not more prompts.
If you're thinking about paying for a tool that rewrites the finished text, read why those tools don't work first.
The Kit puts it in once
I build this with clients as the AI Context Kit. It's not a template you fill in. It's the brief your AI reads before every chat, and it takes real work to produce. You and I do that work together in a half-day session, from what you already know: who your clients are, how you talk with them, what you sell, and what you'll never say.
It loads into ChatGPT, Claude or Gemini once. When a draft misses, you show it your edits and the Kit keeps the lesson.
Not every business is ready for a Kit. If yours isn't, I'll tell you that on the call.
What makes writing read as machine-made is that nothing in it happened.