Why does ChatGPT use so many em dashes?
The best-supported explanation is training data: large language models learned to write from published, professionally edited prose and from Markdown-formatted web text — both of which use the em dash (U+2014) far more often than everyday typists do. Two 2026 studies put numbers on the habit.
10.62 vs 3.23
Em dashes per 1,000 words
GPT-4.1 output averaged 10.62 em dashes per 1,000 words against a 3.23 human baseline, across roughly 240,000 generated words. The authors argue Markdown-heavy training data is the cause. Preprint: arXiv:2603.27006.
4.23% → 11.58%
Papers containing an em dash
In a preregistered study of 69,632 medRxiv preprints, the share of Discussion sections containing an em dash rose from 4.23% before ChatGPT to 11.58% after — +7.35 percentage points, odds ratio 2.96. Preprint: arXiv:2606.29540.
Neither paper is peer-reviewed yet, but they agree with what everyone has noticed since 2023 — and with OpenAI itself. On November 13, 2025, Sam Altman posted:
"Small-but-happy win: If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it's supposed to do!"
— Sam Altman, post on X, November 13, 2025
That "finally" is doing a lot of work: it implicitly concedes both that the em-dash habit is real and that, for two years, telling ChatGPT to stop often didn't help. Em dashes are one of several typographic tells — curly quotes and the ChatGPT watermark characters are the others — that ride along when you paste AI output.
How to remove em dashes from AI text
The fastest fix: paste the text into the cleaner at the top of this page. Any preset with typography normalization on — AI Clean (the default, so this works without changing anything), Safe Clean, or Plain ASCII — converts every em dash (U+2014) and en dash (U+2013) to a plain hyphen and reports how many it changed, along with curly quotes, ellipses, and any hidden characters found in the same pass. Only Keep Typography leaves dashes as they are, which is the preset to pick when you want the hidden characters gone but the punctuation intact.
Prefer to do it by hand? Three other routes:
- Microsoft Word: press Ctrl+H (Find and Replace), enter ^+ in the Find box — Word's built-in code for the em dash (^= finds the en dash) — and your replacement in the Replace box. Replace All.
- Google Docs: Docs has no character codes in Find and replace, so copy an em dash from your text, paste it into the Find field (Ctrl+H), and replace all. Repeat for the en dash if needed.
- At the source (prompt-side): add "Do not use em dashes" to ChatGPT's custom instructions under Settings → Personalization. Per Altman's post above, since November 2025 the model actually honors it. A one-off "rewrite that without em dashes" also works within a conversation.
If your style guide wants em dashes and you only need the invisible stuff gone, use the Keep Typography preset instead — the quotes and dashes normalizer page explains exactly what each mode converts.
Are em dashes proof of AI writing?
No. The em dash is legitimate punctuation that professional writers, editors, and publishers use heavily, and no punctuation mark can prove a text was machine-generated.
Even Wikipedia's "Signs of AI writing" guide — written by the editors who clean AI text out of Wikipedia — lists em-dash overuse and curly quotes as things to watch for while explicitly cautioning that they are observations, not proof of AI authorship. The statistics above are population-level: they show the share of documents containing em dashes more than doubled (4.23% to 11.58%, a 2.7x rise), not that any individual dash is synthetic. Note the base rate — 4.23% of medRxiv Discussion sections used em dashes before ChatGPT existed. Those authors didn't stop being human in 2023.
This matters because the false-accusation risk is real. Plenty of people learned to love the em dash from Emily Dickinson, not GPT-4, and writers have been publicly accused of using AI on the strength of their punctuation alone. If you're assessing authorship, look at provenance — drafts, edit history, whether the author can discuss the content — not glyphs. And if you're a writer tired of defending your dashes, normalizing them is a two-second job with the AI text cleaner. The flip side is equally true: stripping em dashes does not make AI text "human." Detectors model word-choice patterns, not punctuation in isolation, so this is a style fix — never a detector bypass.
What should you replace an em dash with?
Match the replacement to the job the dash was doing: a comma for a light pause, a colon before an explanation, parentheses for an aside, a period when the sentence deserves to be two.
- Comma — light pause or afterthought. "It works — mostly." becomes "It works, mostly."
- Colon — introducing an explanation or payoff. "One thing mattered — speed." becomes "One thing mattered: speed."
- Parentheses — a true aside. "The results — all three trials — were positive." becomes "The results (all three trials) were positive."
- Period — two complete thoughts. "The deadline slipped — nobody was surprised." becomes "The deadline slipped. Nobody was surprised."
The cleaner's automatic substitution is a plain hyphen because it's the only replacement that never changes meaning; treat it as a safe default, then hand-polish the sentences that deserve it. For the rest of the AI-formatting cleanup — quotes, spacing, markdown residue — the full AI text cleaning guide covers every rule.