The em dash and AI: is it really a sign of ChatGPT?
An em dash on its own is not a sign of AI. Writers used it long before chatbots. But AI models tend to overuse it, and studies show em dashes became more common across whole bodies of writing after ChatGPT arrived. One dash proves nothing; a dash in every paragraph, alongside other AI habits, is what people notice.
The em dash has had a strange couple of years. It went from a punctuation mark most people never thought about to a running joke about chatbots, and some writers who have used it for decades now delete it from their drafts so nobody suspects them. This guide sorts out what's actually known, what's guesswork, and what to do about it in your own writing.
What is an em dash?
The em dash (—) is the longest of the common dashes, traditionally about the width of a capital M. Writers use it to set off an aside, to mark a sudden turn in a sentence, or to add emphasis to what follows. It's different from the hyphen (-), which joins words like "well-known", and the en dash (–), which marks ranges like "pages 10–20".
| Mark | Looks like | Typical job | Example |
|---|---|---|---|
| Hyphen | - | Joins words | a well-known writer |
| En dash | – | Ranges and connections | the 2021–2025 data |
| Em dash | — | Asides, sudden turns, emphasis | The draft was fine—until the last line. |
Style guides disagree on spacing. Many book styles write it closed up (word—word), while AP-style news writing puts a space on each side (word — word). That detail turns out to matter in one of the studies below.
Why does AI love the em dash?
Nobody outside the AI labs knows the full answer, so be wary of anyone who says it's one simple thing. The most detailed explanation so far comes from a 2026 paper that tested twelve models from five providers (Freeburg, arXiv 2603.27006). Its argument, in plain terms:
- Models learn from huge amounts of text formatted in markdown: headings, bullets, bold and so on.
- When a model is asked to write plain prose, it drops the obvious formatting, but the em dash survives as a kind of leftover structure. It's a way to bolt a second idea onto a sentence, much like a bullet point would.
- Fine-tuning then strengthens or weakens the habit, which is why the rate differs so much between model families.
The paper reports em dash rates ranging from none at all in Meta's Llama models to about 9 per 1,000 words for one OpenAI model, even when told to avoid markdown. It also found that an explicit "don't use em dashes" instruction didn't fully stop some models. That's one study with its own hypothesis, so treat it as the best current explanation rather than settled fact.
The em dash also fits the wider pattern of how chatbots write. They favour smooth, balanced sentences with an extra clause added for emphasis, and the dash is the easiest way to add one. We cover the rest of that default voice in why ChatGPT sounds robotic.
Is an em dash a sign of AI writing?
As a population trend, yes; as proof about one text, no. A pre-registered study of nearly 70,000 medRxiv preprints found that the share of Discussion sections containing at least one em dash rose from about 4% before ChatGPT to about 12% afterward, reaching around 20% in 2025 (Czuma, arXiv 2606.29540). A follow-up by the same author found a similar rise in U.S. congressional press releases in 2025.
But the author is explicit that the em dash is a population-level indicator, not a per-paper detector, and that the design can't prove what caused the rise. That's the key point. Across thousands of documents, more dashes suggests more AI-assisted writing. In a single essay or email, a dash tells you almost nothing, because plenty of people have always written that way.
Wikipedia's editors have reached a similar view. Their Signs of AI writing field guide now lists em-dash overuse under historical indicators (2022 to September 2026), noting that newer models are increasingly set up to avoid the dash altogether, so its absence is no evidence of human writing either. Our guide to the signs of AI writing covers the tells that still hold up.
Do AI detectors flag em dashes?
Detectors don't publish their exact features, so nobody outside those companies can say how much weight a dash gets. In general, detectors score how predictable a whole passage is: sentence rhythm, word choice, structure. A single punctuation mark is a tiny part of that. The bigger risk is human readers. Teachers, editors and hiring managers now look for em dashes, and someone who spots three in a short cover letter may start reading with suspicion. Our guide to checking whether your writing sounds like AI explains what detectors actually measure and why their scores disagree.
What does AI em dash use look like?
Here's a typical chatbot paragraph. The dashes aren't wrong, but notice how they all do the same job:
"Good writing isn't about fancy words—it's about clarity. The best writers know their audience—and they write for them. In the end, the goal is simple—communicate clearly."
Three sentences, three dashes, each one setting up a neat little reveal. A person might write the same idea like this:
"Good writing is mostly clear writing. Know who's reading, and write for them. That's it."
The fix wasn't just deleting the dashes. It was dropping the pattern of building every sentence toward a punchline. That pattern, along with words like "crucial" and "landscape", is what makes text read as AI. Our list of AI words to avoid covers the vocabulary side.
Should I stop using em dashes?
No. If the dash is part of your voice, keep it. Giving up a useful mark because chatbots overuse it would make your writing worse, not more human. A few practical rules instead:
- Ration them. One or two on a page reads like a person. One in every paragraph reads like a habit, whoever wrote it.
- Use the mark that fits the job. A comma for a light pause, a full stop for a real break, brackets for a true aside, a colon before an explanation.
- Watch the "it's not X—it's Y" shape. That construction, with or without a dash, is one of the most recognisable chatbot moves.
- Keep your other quirks. Uneven sentence lengths, specific details and real opinions do far more to make writing sound like you than any punctuation choice.
For the broader set of habits, see our pillar guide on how to not sound like AI.
How do I stop ChatGPT from using em dashes?
Ask directly and specifically. Vague requests like "write naturally" won't touch punctuation. Something like this works better, and it's best placed in custom instructions so it applies to every chat:
Then check the result. The 2026 study above found some models keep slipping dashes in even under an explicit ban, so search the output for "—" before you use it. If you want more rules like this one, our ChatGPT prompts to sound human collect them by job.
Is it honest to edit AI text this way?
Removing dashes from a chatbot draft doesn't change who wrote it. If your school, employer or publisher has rules about AI help, follow them; punctuation changes don't get around a policy. The aim of this guide is the reverse case: making sure your own writing sounds like you, and that a dash you've used for years doesn't cause needless suspicion.
Frequently asked questions
Is an em dash a sign of AI?
Not on its own. Em dashes became more common across large bodies of text after ChatGPT, but researchers stress that this is a population trend, not a test for one document.
Why does AI use em dashes so much?
No one knows for sure. One 2026 study argues it's a leftover of markdown-heavy training text, shaped by fine-tuning, which is why rates vary so much between models.
What is an em dash in AI writing?
The long dash (—), usually unspaced, used to bolt a second thought onto a sentence. Heavy use of it has become a popular tell for chatbot text.
Should I stop using em dashes?
No. Keep them if they're part of how you write, just don't put one in every paragraph. Commas, full stops, brackets and colons can share the load.
How do I stop ChatGPT using em dashes?
Tell it not to in custom instructions and name the alternatives. Some models still slip, so search the output for the long dash before you use it.