Does ChatGPT use em dashes? What the model actually does

Yes. ChatGPT writes em dashes, but not in every kind of text. In 12 samples we ran through two OpenAI models in October 2026, the dash appeared only in blog-style explanatory prose and never in a plain internal memo. Every dash it produced was unspaced, and that spacing is the part that actually gives a model away.

The mark itself proves nothing. What you noticed was three things landing at once: a genre where dashes do real work, a habit of reaching for one at every clause break, and the specific typographic form the model types. Separate those three and the question stops being about the dash at all.

Does the model use em dashes? Yes, and only in some writing

To check what the folklore actually describes, we ran one small test in October 2026. Two OpenAI models, two writing tasks, three instruction variants, one sample per cell at temperature 0.7. The tasks were a factual internal memo about a renewal-rate drop and a blog-style explainer about churn after a price rise. The variants were no instruction at all, a bare prohibition, and a prohibition paired with a named replacement.

• Factual memo, no instruction, both models: 0 em dashes in 107 and 113 words.
• Blog explainer, no instruction: 2 em dashes in 110 words on the smaller model and 1 in 108 on the newer one, which works out to 18 and 9 per 1,000 words.
• Blog explainer with a bare prohibition: 0 em dashes in both samples.
• Blog explainer with a prohibition plus a replacement rule: 0 em dashes in both samples.
• Spaced dashes across all six explainer samples: 0. Every dash was set tight against the words beside it.

Read that before you draw a large conclusion. One sample per cell is a signal, not a law, and the OpenAI forum threads on this topic are full of people watching the bare prohibition fail. What the samples do show is that genre moves the dash count more than intent does, and that the typography never varied once.

Why does the model reach for the long dash?

The best data on the question comes from Sean Goedecke's October 2025 analysis. GPT-3.5 barely used em dashes. GPT-4o used roughly ten times more than its predecessor, and GPT-4.1 used more still. Something changed between late 2022 and mid 2024, and it was not a design meeting about punctuation.

His answer is training data. General English prose runs at 0.25% to 0.275% em dashes per word, and the rate peaked near 0.35% in 1860 before falling through the twentieth century. Late-1800s and early-1900s print books carry about 30% more dashes than contemporary prose. Moby-Dick alone holds 1,728 of them. If labs bought and scanned older books for high-quality text, they bought a heavier dash habit along with the vocabulary.

The explanations he rules out matter too. It isn't token efficiency, because a comma is just as short. It isn't the dialect of the people doing the rating work, because a Nigerian English sample he measured used dashes at 0.022%, far below the general rate. It isn't that models find dashes structurally irresistible, since the older model everyone now calls primitive did not use them at all.

So the honest answer to why the model reaches for the dash is dull: the books it learned from used more of them than we do. That's a data story, not a personality.

The real tell is orthography, not punctuation

The New York Times Magazine made a sharper point in September 2025 than most of the internet did. The paper's language column argued that the best signal the dash offers isn't punctuation at all, it's orthography. ChatGPT sets its dashes the way a printed book does: a stroke the width of the letter M, with no spaces on either side. The average person typing in a browser does not do that. They reach for a hyphen, or two of them, and their software may convert the pair into a shorter en dash. Most online publications put spaces around their dashes so lines wrap cleanly.

That difference is free to check and hard to fake by accident. Open your document and look at the character between the words. If it's a long unspaced stroke, someone pasted it from a model, or they know their Unicode, or they write for a print-style outlet. If it's a hyphen with spaces, it's a keyboard habit.

Our six explainer samples make the point neatly. Three dashes between them, and not one had a space on either side. The models did not just prefer the punctuation. They preferred the typography of a book page.

How many dashes are too many?

Frequency is the other half. A 200-paper analysis published in April 2026 compared published research across four disciplines against AI-generated drafts on similar topics. Human STEM papers averaged 0.3 em dashes per 1,000 words. Human humanities papers averaged 1.1. AI drafts averaged 3.8. The working threshold for abnormally high use is three or more per 1,000 words.

Put that on a real document. A 1,500-word post at 3.8 per 1,000 carries about six dashes. At the human humanities rate it would carry one or two. If your draft has six and you didn't notice writing them, that's the moment to look at them one by one.

AI-generated drafts averaged 3.8 em dashes per 1,000 words, against 0.3 for human-written STEM papers and 1.1 for humanities papers (ProofreaderPro, 200-paper analysis, April 2026).

One warning about that number. It comes from academic writing, where house styles push dashes out anyway, so the human baselines sit low for reasons beyond habit. For a newsletter or a personal essay, 1 to 2 per 1,000 is normal and reads fine. What matters is whether the dashes are doing different jobs or the same job five times.

When should you keep a dash and when should you cut it?

Keep a dash when it's the only mark that carries the interruption you want. A single aside that breaks a sentence and then returns to it's what the mark is for, and readers feel that break without being told about it. Cut a dash when a comma, a colon or a full stop would do the same work with less noise.

• Aside or interruption: two commas, one before and one after. Cheaper, invisible, same effect.
• Explanation or restatement: a colon. The second half specifies the first.
• List introduction: a colon, every time. A dash before a list is the most common model habit in corporate prose.
• Two independent clauses: a semicolon, or split them and let the second sentence land on its own.
• A dramatic pause at the end of a thought: a full stop and a shorter next sentence.

The test isn't grammar, because the mark is grammatical almost anywhere. It's whether removing it costs the sentence anything. If the answer is nothing, the dash was decoration.

How to stop it without flattening your sentences

Telling a model not to use dashes is the advice every thread gives, and Goedecke notes that it's surprisingly hard to enforce. A bare ban leaves the model with the same sentence shape and one fewer mark, so it often comes back as a hyphen or a comma splice instead. The instruction that held in our samples named the replacement, not just the ban.

Don't use em dashes. Where you would reach for one, use a comma for an aside, a colon for an explanation, or a full stop to split the sentence in two.

Put that in custom instructions if you write with the same assistant every day, and repeat it in the chat when you start a new document. Repetition is the part people skip. A style rule stored in settings competes with everything else in the context window.

If you want the wider version of that, our guide to making ChatGPT sound like you covers the settings that change tone rather than punctuation.

Prompt-level fixes only go so far, though. A prompt library built to produce human-sounding first drafts handles the sentence shapes that make dashes feel necessary in the first place.

Why the dash panic gets it wrong

Brian Phillips argued in The Ringer in August 2025 that the tell punishes the wrong people. Writers who use dashes well, in fiction, in essays, in magazine features, learned them from books and now get accused of being a language model. His claim is that the prevalence of dashes in machine output is evidence of how much the machines leaned on human writers, not evidence that humans should drop the mark.

He's right about the cost. I've watched people delete a good dash to avoid a suspicion. The mark carries a kind of interruption that commas can't, and losing it makes prose more uniform, which is the exact quality the panic was trying to catch.

The panic also confuses a population rate with an individual case. Three dashes per 1,000 words is unusual. Three dashes in a 400-word message is a different thing entirely, and neither number tells you who wrote the words. Punctuation is a weak classifier on its own, which is why the tells that hold up are the ones about rhythm and opinion rather than marks, the same set our breakdown of what makes AI writing sound fake walks through.

What to check in your own draft

Run this on the last thing you published. It takes two minutes and needs no tool.

1. Count the dashes. Divide by word count and multiply by 1,000. Above 3 and you're in AI-draft territory. Between 1 and 2 is ordinary.
2. Look at the character, not just the count. A long unspaced stroke is a paste from a model or a print-style house convention. A spaced hyphen is a keyboard habit.
3. Sort each one by job: aside, explanation, list, clause join. If four of five do the same job, the lack of variety is the problem, not the mark.
4. Replace the ones a colon or a comma handles, and keep the one that genuinely interrupts a thought.
5. Read the sentence aloud without it. If the rhythm doesn't change, it was decoration.

Do that twice and you'll start noticing the pattern in your own writing before you notice it in anyone else's. That's the useful outcome. Dashes are a symptom of a habit, and habits respond to measurement better than to fear.

The dash is one pattern among several, and the others are easier for a reader to spot. If your drafts also open by restating the question, close with a tidy summary, or lean on the same three-part list, fix those first, as our catalogue of common AI writing patterns sets out.

One habit that helps more than deleting dashes: run a second pass with the model and ask it to keep your sentences and only flag the marks. Our guide to editing with ChatGPT without losing your voice has the prompt for that pass.

The short answer to the question at the top: yes, the model uses em dashes, mostly in explanatory writing, and the form it types is more revealing than the count. You can keep using the mark. Just know which one you typed and why.

Frequently asked questions