What is an em dash, and why do people link it to AI?
An em dash is the long dash, about the width of a capital M, used to set off an aside, an afterthought or a break in a sentence. People link it to AI because chatbot text is said to use it a lot, but counting dashes cannot show who or what wrote a page.
By Jack Stovell · Published
Short answer: the em dash is ordinary punctuation that writers used long before chatbots existed. In UK brand writing, its rate barely moved after ChatGPT arrived: 2.43 per 1,000 words before and 2.50 after, across 37 brands that each had at least 10 pages on both sides of the line in the same channel (SGR-011).
What an em dash is (and how it differs from a hyphen and an en dash)
Three marks look alike and do different jobs.
- The hyphen is the short mark on your keyboard. It joins words, as in "well-known" or "sign-up".
- The en dash (Unicode U+2013) is about the width of a capital N. It marks ranges, such as a span of years, and with a space either side it can mark a pause in a sentence.
- The em dash (Unicode U+2014) is about the width of a capital M. It sets off an aside, adds an afterthought or breaks a sentence off.
American style usually sets the em dash with no spaces. Many British publishers do the same job with a spaced en dash instead, which is why our counter looks for both.
You won't find either dash in this post. The site's own style guide prefers colons and full stops, because this is a site that counts dashes.
Why the em dash became a supposed AI tell
Once chatbots were in everyday use, readers noticed that their answers often reached for long dashes. The mark picked up nicknames such as the "ChatGPT dash", and for some readers a dash in a post or an email became a reason to suspect a machine.
The counter-view is simple. The em dash is long-standing punctuation. Novelists, journalists and copywriters used it long before any chatbot, and many still do. Someone who has never opened a chatbot might use two in a paragraph.
Brand writing makes a useful test, because company blogs and press pages hold years of text from before ChatGPT. SGR-011 compared pages published from January 2019 to November 2022 with pages from January 2024 to October 2026, leaving 2023 out as a transition year. Each brand had to have at least 10 pages before and 10 after in the same channel. Across the 37 brands that qualified, em dashes went from 2.43 to 2.50 per 1,000 words, a change of plus 0.07 with a 95% interval of minus 0.36 to 0.49. The study's own wording: "Despite the 'ChatGPT dash' reputation, brand writing used em dashes at the same rate after ChatGPT as before."
Two details are worth knowing. In an exploratory split, press releases gained dashes (plus 0.52 per 1,000 words, interval 0.11 to 0.91, 22 brands), while blog posts did not (minus 0.15, interval minus 0.67 to 0.37, 27 brands). And live copies of old pages carried fewer em dashes than their archived versions, so some old pages have been edited since, which the study says may understate any rise.
How we count em dashes per 1,000 words
The measure is em_dash_per1000, shown on the site as "Em dashes per 1,000 words". For each page, the counter:
- counts every em dash character, with or without spaces around it
- adds every en dash that has a space on both sides
- divides the total by the page's word count and multiplies by 1,000
Pages under 150 words are left out. A rate lets a 400-word press release and a 2,000-word blog post be compared fairly. Each brand's pages are averaged first, so a brand with hundreds of pages does not dominate its sector (SGR-010).
What it does not count:
- hyphens, including a spaced hyphen typed as a stand-in for a dash
- en dashes with no spaces, such as those in ranges of years or pages
- double hyphens, which some people type when they have no dash key
The net is narrow on purpose. A writer who types a spaced hyphen wherever they mean a dash will score zero, however dash-heavy their style. The number counts two characters in the stored text, and that is all it counts.
What the sector averages look like, with page counts
From SGR-010, blog posts and press releases together, with brands weighted equally:
| Sector | Em dashes per 1,000 words | Brands | Pages | |---|---|---|---| | Software | 3.0 | 16 | 3,422 | | Energy | 3.0 | 11 | 1,949 | | Telecoms | 2.4 | 15 | 1,816 | | Banks | 2.3 | 18 | 2,669 | | Insurance | 2.2 | 10 | 2,049 | | Food and drink | 2.0 | 8 | 751 | | Travel | 1.9 | 8 | 459 | | Supermarkets | 1.9 | 10 | 620 | | Fashion | 1.8 | 10 | 221 | | Charities | 1.6 | 10 | 1,994 |
Software and energy sit at the top at 3.0. Charities sit at the bottom at 1.6, despite 1,994 pages behind that figure. The whole spread is 1.4 dashes per 1,000 words: about one extra dash in a typical long page.
Read the middle of the table loosely. Several sectors are within 0.1 of each other, and the study notes that its sector figures "rest on 8 to 18 brands each and carry no intervals". Fashion is the thinnest, with 221 pages across 10 brands, about 22 a brand, so treat its 1.8 as a sketch.
The sector pages on this site set every brand side by side with its own page count. Brands with fewer than 30 pages are greyed out as context, and brands with fewer than 10 show no measures at all. Around 20 brands are added each week, so the live pages can drift from the study figures above.
Why a count cannot tell you who wrote a page
A high em dash rate does not show that a machine wrote a page, and a low one does not show that a person did. Dashes, like the site's word lists, are style markers that people used long before ChatGPT. Among the things SGR-010 says it does not show: "That voice measurement, ScriptGrain's or anyone's, can tell whether a text was written by AI."
Think about how a brand page gets written. One person drafts it, another edits it, a style guide shapes it and a template formats it. Any of them can add or remove a dash. One editor who likes dashes can mark a hundred pages, and one who dislikes them can clear the lot. Pages also change after publication, as the gap between live and archived copies in SGR-011 shows.
So when someone says a page has lots of em dashes and must be AI, the fair reply is that the dash cannot settle the question. A tidy rate per 1,000 words looks like evidence, but it only tells you how many dashes are on the page.
If you want fewer dashes in your own writing
If you would just rather use fewer, the swaps are simple:
- A colon for "here is what I mean": the explanation follows it.
- A full stop for the swerve. Start a new sentence instead.
- Commas for a short aside, and brackets for a longer one (like this).
- Find and replace. Search your draft for the em dash character and for a space, en dash, space, then decide case by case.
None of this makes writing more or less human. It only changes what a counter finds. To see how individual brands compare, start from the brands index, where each page shows its own em dash rate with its page count.
Questions people ask
What is an em dash?
It is the longest of the common dashes, about the width of a capital M (Unicode U+2014). Writers use it to set off an aside, add an afterthought or break a sentence off.
Why are em dashes associated with AI?
Chatbot answers often use long dashes, so some readers began treating the mark as a sign of machine writing, and it picked up nicknames like the "ChatGPT dash". The em dash is long-standing punctuation, though, and in UK brand writing its rate did not rise beyond chance after ChatGPT: 2.43 per 1,000 words before and 2.50 after, across 37 brands with at least 10 pages on each side (SGR-011).
How do em dashes look, and how are they different from hyphens and en dashes?
The hyphen is the short mark on your keyboard and joins words, as in "well-known". The en dash is about the width of a capital N and marks ranges or, with spaces either side, a pause. The em dash is the longest, about the width of a capital M, and interrupts a sentence.
Does using an em dash mean a page was written by AI?
No. Punctuation counts are style markers that people used long before ChatGPT. A count can describe how a page reads. It cannot show who wrote the page or whether a tool was used.
How do you count em dashes per 1,000 words?
Count the em dash characters and any en dashes with a space either side, divide by the page's word count and multiply by 1,000. That is the em_dash_per1000 measure used on this site. It does not count hyphens, unspaced en dashes or double hyphens.