The fastest way to spot AI writing is not one tell. It's the combination of five small habits appearing in the same paragraph, none of which a careful human editor would let stand.

That distinction matters because the individual tells are weaker evidence than most people assume. A 2025 study covered by The Conversation found that readers trying to spot ChatGPT text based on things like em dashes and unusual vocabulary performed only marginally better than chance. Meanwhile a separate 2024 study found that 94 percent of ChatGPT-written undergraduate exam answers went undetected by graders at a British university. So no, you cannot reliably out a machine from a single quirk. What you can do is recognize the pattern, and more importantly, edit it out of your own drafting process, because unedited AI output has a rhythm that readers register as generic even when they can't name why.

Here's what that rhythm is built from.

The em dash overload

AI models genuinely do lean on em dashes harder than people do. An independent frequency analysis of GPT-4.1 output found it uses em dashes at roughly 3.28 times the rate of human-written essays, a pattern researchers trace back to the markdown-heavy text the models were trained on. The habit isn't universal (some models barely use them at all) but when a draft has four or five em dashes in a single section, that's worth a second look. Not because the em dash is bad punctuation, it's a legitimate tool, but because its overuse usually signals nobody went back and asked whether a period, a comma, or a full rewrite of the sentence would have said it better.

The fix: Read the draft aloud. Every em dash that could be a full stop, should be. Keep the ones that earn their place.

"In today's fast-paced world"

This phrase, and its cousins ("In an increasingly digital landscape," "In the ever-evolving world of X"), exists to buy the writer three seconds before they say anything real. It is throat-clearing. It contains zero information a reader didn't already have.

The fix: Delete the first sentence of any draft that starts this way. What's underneath is almost always the real opening.

"It's not just X, it's Y"

The contrastive reframe ("This isn't a product, it's a movement") is one of the most identifiable syntactic templates in AI writing, and researchers studying it have found the pattern shows up because models default to sentence skeletons seen constantly in training data rather than building a new thought. Writers who track this pattern in the wild describe it as a structure that promises a turn in logic without ever proving it: the sentence upgrades the drama of a claim without upgrading the evidence behind it, and once a reader has seen the shape a few dozen times across blogs and LinkedIn posts, it starts triggering irritation instead of persuasion.

The fix: If you write this sentence, finish the thought it promises. If "it's a movement," show one specific thing that makes it a movement. If you can't, cut the sentence.

Listicle rhythm and the missing detail

Ask an AI model for "5 ways to improve X" and it will hand you five headers of near-identical length, each opening with a general claim, followed by a sentence that restates the claim, followed by a generic call to action. No item includes a number only your company would know, a mistake you actually made, or a name of a real client, tool, or date. The structure is fine. The absence of anything specific inside it is the problem.

The fix: For every bullet or subheading, add one fact that could only come from direct experience: an actual result, an actual date, an actual name. If you can't add one, that section probably didn't need to exist.

The checklist

Before you publish anything, run the draft against this:

  • Em dashes: count them. More than two or three in a short piece, check if a period works better.
  • Throat-clearing opener: does the first sentence say anything, or just announce that a topic exists?
  • False contrast: any "it's not just X, it's Y" sentences? Does the piece actually prove the Y?
  • Specificity: does every section contain at least one number, name, or detail nobody else could have written?
  • Point of view: could a reader disagree with anything in this piece, or does it just agree with itself?
  • Read-aloud test: does it sound like a person talking, or like the average of every article on this topic?

FAQ

Does using an em dash automatically mean a text is AI-written? No. Human writers have used em dashes for centuries, and some AI models barely use them at all. Frequency is a weak signal on its own. It only becomes meaningful alongside the other tells above, combined with an absence of specific, checkable detail.

Is it bad to use AI to write a first draft? No. AI is a legitimate drafting tool. The problem isn't the draft, it's publishing it unedited. The fix in every section above takes minutes, and it's the difference between a tool and a crutch.

Can AI detectors reliably catch this stuff? Not consistently. Studies referenced above show human graders and readers only marginally beat chance at spotting AI text from stylistic cues alone. The reliable filter isn't a detector, it's an editor asking whether the piece contains anything only a specific, real person or brand could have written.

If your last few drafts read a little too smooth and a little too generic, run them through this checklist before they go out. The fixes take less time than the first draft did.

Sources: