Do Em Dashes Really Mean Your Writing Was Made by AI?

Somewhere along the way, the humble em dash—that long horizontal line—got branded a snitch. Spot one in an essay and somebody, somewhere, is sure a robot wrote it. It’s a tidy theory. It’s also wrong, and believing it can push you to wreck perfectly good writing for no real payoff.

Let’s pull the myth apart and look at what AI detectors actually react to, why the dash got such a bad rap, and what you should do if your punctuation lands you in the hot seat.

Key takeaways

• No detector flags a single punctuation mark as proof of AI. Em dashes are not a feature these tools actually measure.
• The em dash earned its reputation because some AI models lean on it hard. But plenty of human writers love it too. Correlation isn’t authorship.
• Real AI tells are statistical and structural: flat sentence rhythm, predictable word choice, stock phrasing. Not one keystroke.
• Stripping every dash out of your writing won’t lower an AI score in any meaningful way, and it can make your prose worse.
• If you’ve been accused over punctuation, show your process and write with genuine variety. Don’t censor your own style.

Where the “em dash means AI” idea came from

The myth has a kernel of truth, which is why it spread. People noticed that ChatGPT and similar models produce a lot of em dashes. Ask one to draft a paragraph and you’ll often get sentences stitched together with that long dash, sometimes two or three in a single passage.

So the logic went: AI uses em dashes a lot, therefore em dashes mean AI. That’s a leap. It’s the same reasoning as “burglars wear gloves, so anyone wearing gloves is a burglar.” Lots of innocent people wear gloves. Lots of careful human writers use em dashes.

Why do the models love them? Language models learn from mountains of published text, and published prose is full of em dashes because editors and writers find them useful. The model picks up that the dash is a safe, frequent way to join ideas, and it reaches for it again and again. That’s a stylistic habit of one set of tools, not a fingerprint that detection software hunts for.

What AI detectors actually measure

Here’s the part that breaks the myth completely: detectors don’t read your punctuation and render a verdict. They run statistics over the whole passage.

The two big ones are perplexity and burstiness. Perplexity is a measure of how predictable your next word is to a language model. Human writing tends to be a little surprising, because we make odd choices, reach for a specific noun, or break our own pattern. AI text tends to be smoother and more predictable, because the model is literally choosing likely words. Burstiness is about variation in sentence length and structure. Humans write a long, winding sentence and then a short one. A lot of raw AI output marches along at a steady, even cadence.

If you want the full breakdown of how those numbers work, we wrote a whole piece on perplexity and burstiness, the statistics behind AI detectors. The short version: an em dash doesn’t change your perplexity in any meaningful way. Swapping a dash for a period barely nudges the math the detector cares about.

A quick reality check

It helps to remember the track record here. OpenAI launched its own AI text classifier in January 2023 and quietly retired it that July, citing low accuracy. A 2023 Stanford study (Liang et al., published in *Patterns*) found that several detectors flagged writing by non-native English speakers as AI far more often than writing by native speakers. These tools struggle with the big, structural questions. The notion that any of them keys off a single dash gives them way more precision than they have.

A mini-scenario: two students, one dash count

Picture two students, Maya and Jordan, turning in the same assignment. This is a thought experiment, not a study, but it shows the logic clearly.

Maya writes by hand, then types it up. Her draft is full of em dashes, because that’s just how she thinks—in asides and interruptions. Her sentences swing from four words to thirty. She names a specific court case, misremembers a date and corrects it mid-paragraph, and uses the word “lopsided” where most people would write “uneven.”

Jordan pastes a prompt into a chatbot and turns in the output untouched. It has barely any dashes. Every sentence runs about the same length. The vocabulary is competent and utterly generic. Nothing is specific to Jordan, because Jordan wasn’t there for any of the thinking.

If you judged by em dash count alone, Maya looks more “AI” than Jordan. That’s the whole myth, upside down. A detector measuring perplexity and burstiness would likely do the opposite: it reads Maya’s surprising words and jagged rhythm as human, and Jordan’s flat, predictable prose as machine-like. The dashes were a red herring the entire time.

The real tells worth knowing

If punctuation isn’t the signal, what is? When you’re checking your own work, these are the patterns that actually read as AI:

• Flat rhythm. Sentence after sentence at the same length. Real writing lurches around.
• Predictable phrasing. Stock openers and connective tissue: “In today’s world,” “It’s important to note,” “Moreover.” Stacking these is a louder signal than any dash.
• Vagueness. No names, no numbers, no specific moments. AI loves to gesture at an idea without committing to a concrete one.
• Relentless balance. Every sentence neatly structured, every list a tidy three, every paragraph the same shape. Humans are messier.
• No point of view. Competent summary with nobody actually behind it.

Notice what’s not on that list. Punctuation. You can write something dash-free and still trip every one of these. You can also write something dash-rich that reads as unmistakably human, because it has a voice.

Why deleting your em dashes backfires

Say you buy the myth and decide to scrub every em dash from your essay. Two things happen, and neither helps you.

First, your AI score doesn’t really move, because you didn’t touch perplexity or burstiness. The detector was never counting dashes. You changed the cosmetic layer and left the structural patterns exactly where they were.

Second, you might make your writing worse. The em dash does real work. It sets off an aside, it lands a punchline, it creates a beat. Replace all of them with commas and your sentences can turn into long, comma-spliced mush. Replace them with periods and you can lose the connections between ideas. You’d be sanding off a tool that good writers use on purpose.

If a passage genuinely reads like a machine wrote it, the fix is structural: vary your sentence lengths, swap generic phrasing for specifics, and let your actual opinion show. Curious how the tooling fits together? You can see how our plans work and, for more on the patterns that matter, read more of our writing on AI detection.

What to do if you’re accused over punctuation

It happens. A teacher or editor sees a few em dashes, maybe runs a detector, and decides you cheated. Here’s how to respond without panicking.

Lead with evidence, not denial. Punctuation is not proof of anything, and you can say that directly and politely. Then offer what actually demonstrates you wrote the thing: earlier drafts, version history in your document, your research notes, and your ability to explain your argument out loud. Someone who wrote the paper can talk about the choices in it. Someone who didn’t usually can’t.

If a detector score is on the table, name its limits calmly. Detectors output probabilities, not proof. They’ve been wrong often enough that OpenAI pulled its own. They’ve shown bias against non-native English writers. A score is one data point, not a confession, and a single punctuation mark isn’t even that.

Frequently asked questions

Are em dashes a sign of AI writing? Not on their own. Detectors measure statistical patterns across a whole passage, like word predictability and sentence-length variety. An em dash is just a dash. Plenty of humans use them constantly and plenty of AI text has none.

Why does ChatGPT use so many em dashes? Models learn from polished published writing, where dashes are common, so they reach for the mark often. It’s a stylistic habit of one kind of tool, not a universal law, and it’s not what detectors score.

Will removing em dashes lower my AI detection score? Almost never in any meaningful way. Detectors weigh word predictability and rhythm far more than punctuation. Swapping dashes for commas leaves those signals intact, so the number barely moves.

My professor said em dashes mean I used AI. What do I do? Shift to evidence. Offer drafts, version history, and your ability to explain the work out loud. Note that detectors give probabilities, not proof, and are known to misfire.

Do human writers actually use em dashes a lot? Yes. Essayists, novelists, and journalists have leaned on them for a long time. Emily Dickinson built poems around them. The dash is one of the most flexible marks in English.

The bottom line

A punctuation mark can’t tell anyone who wrote your work. The em dash is old, useful, and beloved by humans long before any chatbot existed. If your writing ever reads like a machine made it, the cause is rhythm, predictability, and missing specifics, not a long horizontal line.

Want to see what a detector is really reacting to in your draft? Run a passage through a humanizer and watch which sentences light up. The dashes won’t be the problem.