The Em Dash Tell: Why Overusing Dashes Flags Your Writing as AI and How to Fix It

You’ve probably felt it before you could name it. A paragraph looks fine at a glance, the grammar’s clean, the points are reasonable, and yet something about it reads *off* — like it was assembled rather than written. Lately, one of the loudest culprits is a single piece of punctuation: the em dash.

The dash isn’t the villain here. The frequency is. AI models reach for the em dash so often that a wall of them has quietly become one of the first things sharp readers notice, and one of the first things people point to when they say a piece “sounds like ChatGPT wrote it.”

Key takeaways

• The em dash is good punctuation. The tell is *overuse* — three or four dashes per paragraph, all doing the same job.
• Models lean on the dash because it’s a low-risk, all-purpose connector that fits almost any sentence, and their tuning rewards that smooth, hedge-everything rhythm.
• A real tell is a pattern, not a single mark. One or two dashes in an essay is just normal writing.
• The fix is variety: swap some dashes for periods, commas, colons, or parentheses, and let your sentence lengths actually change.
• Punctuation alone won’t fool a detector. It’s one surface signal among many, so treat a dash audit as part of real revision, not a find-and-replace stunt.

Why does AI use so many em dashes?

Language models don’t “choose” punctuation the way you do. They predict the next likely token based on enormous amounts of text, and the em dash happens to be one of the most flexible tokens available. It can join two clauses, drop in an aside, or build toward a little reveal. It almost never feels grammatically wrong. For a system optimized to produce confident, smooth-sounding prose, that flexibility is gold. The dash is the path of least resistance.

There’s a second reason, and it’s about taste rather than grammar. The models most people use were tuned to sound balanced, measured, and explanatory. That register loves the dash. It lets a sentence add nuance without fully committing to a new thought — pause, qualify, resume. Humans do this too, but we run out of patience for it. A model never does. It’ll happily produce the same rhythm in sentence after sentence, because nothing in its objective tells it that repetition is boring.

If you want the deeper statistical picture of *why* predictable patterns like this get flagged, our breakdown of perplexity and burstiness covers the math. The short version: human writing is lumpy and a little unpredictable, and machine writing tends to be smooth. The dash is one visible symptom of that smoothness.

What the tell actually looks like

Here’s the part people get wrong. They read “em dashes are an AI tell” and panic-delete every dash they’ve ever typed. That’s an overcorrection. The em dash has been a legitimate tool for centuries, and plenty of excellent human writers use it constantly. One dash proves nothing.

The signal is *density and sameness*. Watch what happens when a model writes a paragraph:

> The new policy — which many employees had requested — aims to improve work-life balance. It introduces flexible hours — a long-overdue change — and remote options. Managers will need training, of course — but the benefits are clear — and the rollout begins next quarter.

Three sentences, five dashes, and every single one is doing the identical job: inserting a parenthetical aside. Nothing else is varying. The sentences are all roughly the same length. The rhythm is a metronome. *That’s* the tell. Not the dash itself, but the fact that it’s the only tool in the box.

Now compare a version a person might actually write:

> The new policy aims to improve work-life balance, something employees had been asking about for years. It adds flexible hours and remote options. Managers will need training, sure. But the benefits are obvious, and the rollout starts next quarter.

Same information. One dash’s worth of nuance, redistributed into commas, a period, and a blunt little “sure.” The second version breathes. The first one marches.

How to fix em dash overuse without flattening your voice

The goal is variety, not abstinence. Here’s how to get there.

1. Do a dash count, then cut by half

Drop your draft into any editor and search for the dash character. If a single paragraph has three or more, that’s your starting point. You don’t have to remove all of them — just decide which ones are earning their place and rewrite the rest. The first cut is almost always easy, because at least half were never necessary.

2. Match the replacement to the job

A dash usually stands in for something more specific. Ask what it’s actually doing:

• Joining two related thoughts? A period or a semicolon often reads cleaner. “She finished the report — then she left” becomes “She finished the report, then left.”
• Adding a quick aside? Commas or parentheses do the same thing with different texture. Rotating between them keeps a page from feeling uniform.
• Building to a definition, a list, or a punchline? A colon is frequently the better, more confident choice.

You’re not banning the dash. You’re refusing to let it be your default.

3. Let sentence length actually move

Em dash overuse usually travels with another AI habit: sentences that are all the same medium length. Fixing one helps the other. Write a short sentence. Then write a longer one that winds through a couple of ideas before it lands. Then a fragment. Variety in *length* does more for the human feel of a paragraph than any single punctuation swap, and it naturally crowds out the metronome rhythm that dashes create.

4. Read it out loud

This is the cheapest test there is. When you read aloud, your mouth notices repetition your eyes skip. If you keep hitting the same little pause-aside-resume beat, you’ll hear it. Anywhere your voice stumbles into the same shape twice in a row, break the pattern.

A quick mini-scenario

Say you’re a marketing writer with a blog post drafted by an AI tool. It’s decent. The ideas are right, the structure works, and you could ship it in five minutes. But you scan it and count eleven em dashes across eight paragraphs, most of them inserting throwaway asides.

You don’t rewrite the whole thing. You spend ten minutes doing exactly what’s above: cut the dashes to four, convert two asides to parentheses, turn three into clean sentence breaks, and chop two of the longest sentences in half. The meaning doesn’t change. But the post stops sounding like it was extruded and starts sounding like a person wrote it on a slightly caffeinated afternoon. That’s the whole win, and it’s the kind of edit a reader feels even when they can’t say why.

If you’d rather not do that pass by hand every time, here’s what PaperBleach actually does: it rebalances rhythm and punctuation for you, and then you give it a final human read. For more on the craft side of this, our blog digs into the other tells worth knowing.

A caution about treating this as a hack

It’s tempting to think “no dashes equals undetectable.” It isn’t true, and believing it will burn you.

AI detectors don’t sit there counting punctuation marks. They mostly score statistical properties — how predictable your word choices are, how much your sentence lengths vary, how the whole thing clusters against patterns they’ve seen. Punctuation is a surface signal a *human* reader catches, not the core of what a machine measures. OpenAI even retired its own AI text classifier back in July 2023 because it was too inaccurate to rely on, which tells you how fuzzy this whole space is. And a 2023 Stanford study by Liang and colleagues found that several popular detectors disproportionately flagged writing by non-native English speakers, a reminder that these tools score style, not truth, and they make mistakes.

So fix your dashes because it makes your writing *better*: clearer, more varied, more obviously yours. That improvement happens to reduce one of the most recognizable AI fingerprints, which is a nice bonus. But it’s a craft move, not a cheat code, and it works best as one part of a genuine revision rather than a robotic find-and-replace.

Frequently asked questions

Why does AI use so many em dashes? Models predict the next likely token from patterns in their training data, and the em dash is unusually flexible — it can join clauses, add an aside, or build to a reveal without committing to a more specific structure. The polished, explanatory style these models were tuned to produce also leans on the dash for its smooth rhythm, so they default to it far more than most people do.

Are em dashes themselves a sign of AI writing? No. The dash is normal, useful punctuation that good writers have used forever. One or even two in a long piece proves nothing. What reads as a tell is the density and the sameness: several per paragraph, the same construction repeated, and no other punctuation doing any work.

Will removing em dashes make my writing pass an AI detector? Not by itself. Detectors mostly score predictability and sentence variation, not one punctuation mark. Deleting dashes cleans up a signal a human notices, but if the underlying rhythm is still flat, a detector can still flag it. It’s one layer of revision, not a shortcut.

How many em dashes is too many? There’s no hard rule, but if a reader notices them as a *pattern*, it’s too many. More than one or two per paragraph starts to feel mechanical, especially when each does the same job. Read it aloud. If you keep hitting the same pause-aside-resume beat, vary it.

What should I use instead of an em dash? Match the replacement to the job: a period or semicolon for joining thoughts, commas or parentheses for asides, a colon for building to a list or definition. The point isn’t to retire the dash, it’s to stop making it your only tool.

The bottom line

The em dash didn’t get worse. It just got overused, and overuse is what readers and detectors both notice. Audit your draft, redistribute the dashes into other punctuation, let your sentences change shape, and most of the “this feels like a machine wrote it” quality evaporates on its own.

If you want a faster first pass before your own final edit, try PaperBleach and see how your draft reads once the rhythm’s been rebalanced.