Forget the em dash: what actually gives away AI-written text

The em dash is no longer proof of anything. The Economist analyzed 55,940 sentences and 1.2 million words, comparing its own writing with responses from ChatGPT, Claude, Gemini and Grok, plus New York Times and Washington Post articles and novels published between 1950 and 2022. The result: among current models, only Claude uses the em dash more than humans do — and ChatGPT uses it far less than any human author in the study.

In other words: the hunt for the "ChatGPT dash," which has become a national sport in marketing meetings, is aiming at the wrong target. Worse: it's condemning people who write well, since the em dash has always been a legitimate tool for anyone comfortable with the written word.

What's left as a signal

The markers that held up in the study are much less convenient to hunt for, because they're about structure, not punctuation:

• Sentences that are all the same length. Human text has irregular rhythm: a long sentence, then a short one. Machine text tends toward constant rhythm.
• Commas that disappear. A subtle economy of punctuation in places where a person would pause.
• The "it's not X, it's Y" construction. That symmetrical contrast that sounds profound and, when repeated, gives away the mold.
• Long words where a short word would do. The models prefer significant, increasingly, implications, along with rare and scientific terms like parameter, methodology, interdependence.
• Nominalization. Turning a verb into a noun — "carry out the implementation" instead of "implement."

Paul Graham, co-founder of Y Combinator, pointed out a cousin of this last item this week: the overly colorful verb. A journalist writes that a proposal "received" a hundred votes; the machine picks a flashier verb to say the same thing. Nobody talks like that in real life.

The study's uncomfortable conclusion is something else, and it matters more than the checklist: with every new version, machine writing gets closer to human writing. Anyone betting on detecting it by form is going to lose that race.

The real problem isn't getting caught. It's sounding like everyone else

For a company, "will people notice I used AI?" is the wrong question. Nobody fires a vendor over punctuation. What actually costs money is something else: when your content is indistinguishable from any competitor's content, it stops giving anyone a reason to choose you.

And that risk didn't arrive with AI — AI just sped it up. Generic text already existed before, written by people, on autopilot. The difference is scale: now you can churn out twenty mediocre pieces a day.

Five things the machine doesn't have

What separates your text from everyone else's isn't style, it's information that only you have:

1. A number you measured yourself. How long it took, how much it cost, what the error rate was. One data point from your own work is worth more than three paragraphs of context.
2. The case with a name, a deadline and a value. "A client in the industrial sector" convinces no one. "A 427-part assembly, converted in 31 seconds" convinces.
3. The mistake you made. Telling what went wrong and how it was fixed is the hardest thing to imitate — and the one that builds the most trust.
4. The opinion that carries risk. Saying what you wouldn't do, who your service isn't for, which trend you think won't last.
5. The operational detail. The thing only someone who actually does the work knows: the step that always runs late, the question the client always asks in the third meeting.

How to use AI without becoming just another one

It's not about abandoning the tool. It's about changing the order of the work:

• You bring the facts, AI organizes them. The opposite — asking it to invent the content and then hunting for the facts — is the shortest path to interchangeable text (and to factual errors).
• Read it out loud before publishing. Wherever you stumble, the sentence isn't yours. This test catches more than any detector.
• Cut a third of it. Almost all the fat in automated text is in warm-up paragraphs that say nothing.
• Swap the long word for the short one. If "use" would do, don't write "utilize."
• Check what the text claims. Every time a number or a date shows up, verify it against the source before publishing. This paragraph exists because this is the step most often skipped.

In one sentence

Stop hunting for em dashes in other people's text and start putting into yours what no machine has access to: what you measured, what you got wrong, and what you think. If your company's content is about teaching for real, it's also worth reading content that teaches sells more than content that advertises — and if you want that production running with a method behind it, that's what we do in digital marketing.

Sources: The Economist, "How to spot AI writing" (August 2026), with coverage by Fast Company; and Paul Graham's post on X on August 19, 2026.