Evidence review7 min read
Do em dashes mean AI? What the evidence actually says
Em dashes are the most quoted AI tell and the weakest single one. Where the claim came from, what it can and cannot prove, and when a dash really is a signal.
Updated September 25, 2026
How this was tested
The density figures are measured, not remembered: we counted the dash characters in the 13 edited posts on this blog (the same engine that powers the counter tool) and in a paragraph written in full AI-house style. Reproduce both with npm run measure:tells in the site’s public repository.
Numbers last verified:September 25, 2026
The short version
- A single em dash proves nothing. People who read a lot, write professionally, or use a Mac keyboard shortcut have been using them for decades.
- The real signal is density and consistency: several em dashes per paragraph, spaced the same way, in a text that has no other irregularities.
- No mainstream AI detector scores em dashes as a feature. Detection tools look at token probabilities, not punctuation counts.
- Overuse is still worth fixing — for readability, and because readers now notice.
Somewhere around 2023 the em dash stopped being punctuation and became an accusation. A single — in a paragraph is now enough for someone to reply "this is ChatGPT" and mean it. That is a remarkable amount of weight to put on one character that has been in print since the eighteenth century.
So does the em dash actually mean AI? Short answer: no, not on its own. Longer answer: a specific pattern of em dash use does correlate with model output, and understanding the difference is the whole point of this article.
Where the claim came from
The belief spread from the place most beliefs about AI spread from: visible, repeated examples. In 2023 and 2024, large numbers of people used ChatGPT for the first time to write emails, posts and essays, and shipped the output unedited. Those drafts had a recognisable fingerprint — and the em dash was the most visible part of it, because it is a character most people cannot type.
That last detail matters. On a Windows keyboard there is no simple em dash key. You need Alt + 0151, or Word's autocorrect, or a character map. On a Mac it is Shift + Option + -. Models emit the character instantly and correctly. So readers learned to associate a typographic character that requires effort with text that had obviously had no effort applied to it.
The reasoning was backwards, and it still worked
People inferred "AI" from "character most humans do not type". That inference was sound in 2023, when the only people producing em dashes in bulk were models and publishers. It stopped being sound as soon as everyone learned the shortcut.
What the evidence actually supports
Split the claim in two and the picture becomes clear. Claim A: em dashes appear more often in AI text than in human text. This is broadly supported — models do over-produce them, for reasons covered in why ChatGPT uses em dashes so much.
Claim B: an em dash in a text means the text is AI. This is not supported, and it cannot be, because the character is not exclusive to models. A finding that "AI uses X more" tells you nothing about the probability that a given X came from AI unless you also know how often humans use X.
On the second question the base rates are not friendly to the accusation. The em dash is standard in published prose: most American style guides require it for parenthetical breaks, and it has been standard for well over a century. Anyone who writes for a living, was taught by a publisher, or writes in an environment with smart punctuation will produce em dashes without thinking about it.
Why the signal is weaker than it looks
Not a signal
One or two em dashes in a long email, used where a comma would also work, from a writer who also uses semicolons and a mix of short and long sentences.
Worth a second look
Every third sentence carries a spaced em dash, every paragraph is the same length, no fragment, no aside, and the same three adjectives appear throughout.
The distinction is between a feature and a distribution. One dash is a feature, and features overlap constantly between human and machine writing. What separates them is the distribution: how often, how evenly, and how much else varies. Uniformity is the honest signal, and the dash is only its most visible symptom.
This is also why the accusation fails so often in public. People screenshot a single dash, and the person who wrote the text points out — correctly — that they have used em dashes since 2009 and can show receipts.
The density numbers, measured on this site
Argument needs numbers, so here are ours. We counted every dash character in two places: the 13 edited, human-reviewed posts on this blog, and a paragraph written deliberately in AI-house style. The overlap is the finding.
| Text | Em+en dashes per 1,000 words | What it shows |
|---|---|---|
| This blog’s 13 edited posts | Average 6.7, range 0–24 | Careful human editing lands here |
| — the punctuation post alone | 24.2 | A post about dashes is the densest text on the site |
| — the detector-accuracy post | 0.0 | The same editorial standard, zero dashes |
| AI-house-style paragraph | 11.5 | Machine-set rhythm with the dash as a crutch (the sample is small: 1 dash, 87 words) |
Read the range, not the average. If density alone proved anything, the most machine-looking text on this site would be the one explaining punctuation tells — and the zero-dash posts would be the only clean ones. Density is evidence about editing effort, nothing more. That is also why the pragmatic threshold below is generous: under about three per thousand words, the dash tells you nothing at all.
When an em dash really is worth acting on
Even with the caveat above, there are situations where the dash deserves attention. They are about consequence and readability, not detection:
| Situation | What to do |
|---|---|
| Academic submission with an AI-use policy | Check the style guide first. APA 7 permits em dashes but advises restraint; some departments discourage them entirely. See em dashes in academic writing. |
| Client work where the draft was AI-assisted | Remove them anyway. The client is paying for writing that does not look templated. |
| You are being accused | Do not argue about the dash. Produce your drafts, your edit history and your research trail: what to do if you are accused. |
| You use them out of habit, not intent | Trim, do not eliminate. Two per page reads deliberate; ten reads like a tic. |
The reliable test takes ten seconds
Count the dashes and the words in the same passage. Under about three per thousand words, the dash tells you nothing. Over six, the text is doing something unusual in either case — model-written or dash-happy — and it is worth an edit.
How to check your own text
You do not need a detector. You need a count and a bit of judgement:
- Count the dashes. The em dash counter reports every dash character in a paste, including the horizontal bar and two-em dashes that hide in copied text.
- Divide by words and multiply by 1,000. That is your density.
- Read one paragraph aloud. If the dashes are doing the work commas and full stops should do, replace them.
- Fix it in one pass with the AI dash remover, which also catches typed
--and---.
Clear the dashes in one pass
Paste the text, remove every em dash, en dash and typed hyphen, and keep the wording exactly as it is.
Open the AI dash removerFrequently asked questions
Do em dashes mean a text was written by AI?
No. The em dash is standard punctuation that predates language models by centuries, and many people type it routinely. What correlates with model output is high em dash density combined with uniform sentence and paragraph lengths, not the presence of the character on its own.
Why does everyone say em dashes are an AI giveaway?
Because in 2023 and 2024 many people published unedited model output, which does use a lot of em dashes, and because the character is hard to type on a Windows keyboard. Readers learned a shortcut that was accurate for that moment and has been weakening ever since.
Should I stop using em dashes because of AI?
You should use them deliberately rather than constantly, which was good advice before AI existed. Chicago warns that dashes "tend to be overused". A couple per page is fine; a dash in every paragraph is worth editing whether or not a model was involved.
Is there an em dash detector?
Not a meaningful one. You can count dash characters exactly, which the dash counters on this site do, but no tool can turn that count into an authorship claim. Treat any "em dash detected = AI" result as marketing.
Do AI detectors use em dashes as a feature?
Publicly documented detectors describe per-token probability scoring rather than punctuation heuristics. Some products mention formatting as a weak secondary signal, which is why editing a draft can move a score without any of the substance changing. See how accurate AI detectors are.
I use em dashes constantly and I am not using AI. What now?
Keep using them, and be ready with an example if the subject comes up. For anything you submit somewhere that runs a detector, note your drafting process in advance. If the accusation has already landed, the steps at accused of AI writing are what actually resolve it.
Keep reading
Dashes5 min read
Why does ChatGPT use em dashes so much?
Training data, token economics and preference tuning all push the same way. Here is the mechanical reason ChatGPT reaches for the em dash, and why Claude does it too.
Updated September 25, 2026
AI writing5 min read
Signs of AI writing: the complete list of tells
Seventeen structural tells, the vocabulary and punctuation patterns, plus the "giveaways" that are actually myths. Ordered by how much each one really proves.
Updated September 25, 2026
Detection5 min read
How accurate are AI detectors?
Vendors claim 98% accuracy and under 1% false positives. Independent testing tells a different story, especially for non-native English writers.
Updated September 25, 2026