Why AI Loves the Em Dash, and How to Spot It

Open almost any ChatGPT or Claude output and start counting em dashes. You'll hit two or three in the first paragraph alone. The em dash has become one of the most reliable signals that a piece of text wasn't written by a human, not because the mark itself is wrong, but because AI uses it compulsively, in the same grammatical slots, at a frequency no human writer ever would.

This guide explains why it happens, how to find it in your own drafts, and what to replace it with.

Why AI reaches for the em dash so often

Language models learn from enormous quantities of human-written text, including journalism, essays, and blog posts. In skilled writing, em dashes appear for emphasis and for interruption. They're a stylistic tool.

The problem is that AI models learn the association between em dashes and sophisticated prose without learning the restraint that makes sophisticated prose actually work. The result is a model that treats the em dash as a general-purpose connective, plugging it in wherever a clause needs to attach to another clause.

There's also a structural reason. AI outputs tend toward long, accumulative sentences. The model generates a main clause, then needs to add a qualification or elaboration, and the em dash is a convenient hinge. Rather than stopping and starting a new sentence, the model keeps adding. Every clause that could have been its own sentence becomes a tail.

The specific grammatical slot AI favors

Watch for this construction: main clause, em dash, then a brief summary or editorial note.

AI models gravitate toward using em dashes to tack on what amounts to a commentary on what was just said. "The process is straightforward" becomes "The process is straightforward — or at least, it should be." That trailing qualifier, separated by an em dash, is an AI fingerprint. Human writers do this occasionally. AI does it constantly.

How to spot the pattern in your drafts

Read your draft and highlight every em dash. Then ask three questions about each one.

First: could this be a comma? If yes, the em dash is probably unnecessary. Second: does this em dash attach a clause that could be its own sentence? If yes, break it. Third: is there more than one em dash in this paragraph? If yes, at least one needs to go.

For a piece running around 1,000 words, two em dashes total is generous. Three is the absolute ceiling before it starts reading like a bot. Anything above that should trigger a full revision of your connective tissue.

You can also search your document for the em dash character (—) and count. Most word processors and text editors support find-and-replace, so the audit takes about thirty seconds.

Before and after: fixing the most common pattern

Here's the kind of sentence that pops up constantly in AI-generated copy.

Before (AI pattern): "Content marketing is a long game — one that rewards consistency and patience — and most brands underestimate how much time it actually takes to see results."

That sentence has two em dashes and a nested parenthetical. It's trying to sound thoughtful but reads breathless.

After (revised): "Content marketing is a long game, and most brands underestimate how much time it takes to see results. The reward for consistency and patience is real, but it rarely shows up in the first six months."

The revision breaks the idea across two sentences, cuts the dashes entirely, and adds a concrete claim ("first six months") that the original was hedging around. It now sounds like a person who has an opinion, not a model summarizing multiple possible opinions at once.

The patterns that travel with em-dash overuse

The em dash rarely shows up alone. When you see it overused, you'll usually find a few other markers nearby.

Long, stacked sentences. AI writing tends to keep building a single sentence rather than finishing it and starting another. The em dash is often the hinge point where a second sentence should have started.

Qualifier clauses. "Which is to say," "in other words," "or more precisely" — these often appear around em dashes as the model hedges or rephrases what it just said. One rephrasing in a piece is fine. Four is a tell.

Noun-heavy phrasing. "The creation of content" instead of "creating content," "the achievement of results" instead of "achieving results." AI writing loves abstract nouns, and the em dash often separates two noun-heavy clauses that have lost their verbs.

For a fuller list of what to look for, see our breakdown of 18 signs a piece of text was written by AI. The em dash is one data point; in combination with the patterns there, it becomes near-certain identification.

What to replace em dashes with

The em dash's replacements depend on what it was doing.

When it introduced a trailing comment: use a period and a new sentence. "The strategy worked — eventually" becomes "The strategy worked. Eventually." The pause and the new sentence weight the word "eventually" more than the dash ever did.

When it separated a parenthetical: use parentheses or commas. "The report — published in March — showed strong results" becomes "The report, published in March, showed strong results." Commas are quieter and less performative.

When it linked a cause to an effect or a premise to a conclusion: use "because," "so," "which," or a colon. "The meeting ran long — nobody had prepared" becomes "The meeting ran long because nobody had prepared." The conjunction is clearer.

When nothing else fits: keep the dash. It's a valid piece of punctuation. The issue is frequency, not existence.

This is also where a good editing pass matters more than any rule. If you want a systematic approach, the free humanizer prompt at /humanizer-prompt includes specific instructions for reducing em-dash density as part of a broader revision sweep.

Why detectors flag it and why that matters

Some AI detectors do track punctuation patterns, including em-dash frequency, as a feature input. Whether that affects a particular detector's output depends on the specific tool, and results vary widely across platforms. We're not in a position to speak to any single detector's methodology.

What's relevant is simpler: human editors and readers notice this pattern too, often without being able to name it. Something feels off. The writing seems overworked or slightly robotic. That reaction usually traces back to one of the patterns described here. Em-dash density is high on the list.

For more on the vocabulary patterns that travel alongside punctuation tells, the guide on words that instantly signal AI-generated text covers the lexical side of the same problem.

---

FAQ

How many em dashes are too many?

For a typical 500-1,500 word article, anything above three is a warning sign. Two is comfortable. One is invisible. The problem isn't the individual dash; it's the pattern of reaching for it in every paragraph.

Does ChatGPT actually use em dashes more than other models?

ChatGPT and Claude both show high em-dash frequency in their default outputs, which is why "chatgpt em dash" gets searched so often. The specific rate varies by model version and prompt, but the tendency is consistent across most current large language models because they share similar training data distributions.

If I just do a find-replace on em dashes, is that enough?

No. Replacing every em dash with a comma or a period without reading for context will break sentences in new ways. The fix has to be sentence-level: read what the dash was doing, then choose the punctuation that serves the actual meaning. A replace-all is a starting point for finding the dashes, not for fixing them.

Can I use AI to fix AI em-dash overuse?

Yes, with a specific prompt. A vague instruction like "make this sound more human" rarely reduces punctuation frequency on its own. You need to say something direct: "Replace em dashes with commas, periods, or appropriate conjunctions. Do not add new em dashes. Aim for no more than two total in the entire piece." The humanizer prompt includes this kind of explicit constraint, which is why it tends to work better than generic rewriting prompts.

Is overusing em dashes always a sign of AI writing?

No. Some human writers (especially journalists working in a certain style) use em dashes liberally. The tell isn't any single dash; it's the combination of high frequency, consistent placement in the same grammatical slot, and the presence of other AI patterns alongside it. One or two dashes in a 1,500-word piece means nothing. Fourteen dashes alongside AI vocabulary and stacked qualifiers means something.

If you want to understand how this fits into the broader set of AI writing patterns, the piece on AI's favorite sentence structure, the rule of three, covers another pattern that shows up alongside em-dash overuse in most AI-generated copy.