The Em Dash and AI Writing: What the Mark Proves and What It Does Not

No punctuation mark has been accused of more than the em dash. The em dash AI argument goes like this: chatbots use the long dash constantly, so a draft full of them must be machine-written. Part of that is true. Studies have measured both the habit in model output and its rise in published writing. The conclusion does not follow. This page explains what the mark is for according to the style manual, why it became a perceived tell, what the numbers show, and how to mark dashes in a draft without accusing anyone of anything.

What the em dash is for

A dash is a choice. The question is whether anyone made it.

The Chicago Manual of Style's answer on hyphens and dashes describes three marks of increasing length: the hyphen, the en dash and the em dash, with the work each does roughly matching its length. The hyphen joins words that act as one idea (toll-free call). The en dash connects things related by distance, such as a range of pages or months.

Chicago gives the em dash several jobs. It lets an extra thought break into a sentence, much as parentheses do. Three em dashes in a row stand in for a repeated author's name in a bibliography. One or two can show interrupted speech. And it can serve as a kind of bullet before list items. The same answer adds that using the mark to break into a sentence is a matter of taste, judged by the writer's and reader's ear. The example sentences in Chicago's answer set the dash tight against the words on either side, with no spaces.

Why the em dash became a perceived tell

Wikipedia's field guide to AI writing describes the pattern more precisely. Model output was much more disposed to use em dashes than nonprofessional human text of the same genre, and used them where people would more likely put commas, parentheses or colons, often in a formulaic, punched-up sales style. Its dashes also tended to be surrounded by spaces, against common typographic guidance. The guide calls the sign most useful in combination with other indicators, not by itself, and notes it was far more common on discussion pages than in article text.

That guide now files dash overuse under historical indicators, dated 2022 to September 2026. It reports that some AI companies tried to suppress the mark in newer chatbots, and that the latest models are more and more often found to avoid it entirely. A sign that a vendor can switch off is a sign about one model at one time. It cannot carry an accusation.

Em dash AI detection: what the numbers say

Two recent papers measured the dash directly, one in model output and one in published writing.

Freeburg's study of twelve models from five providers (five of the largest model developers) asked them to write prose, then asked them again while telling them to avoid markdown formatting. Headers, bullets and bold text disappeared or nearly disappeared. Em dashes persisted in most models. Rates ranged from 0.0 per 1,000 words for Meta's Llama models, which produced none at all, to 9.1 per 1,000 words for GPT-4.1 under the no-markdown instruction. Even an explicit ban on em dashes failed to remove them in some models.

Czuma's pre-registered study of medRxiv preprints took first versions posted from 2020 to 2025 with a Discussion section of at least 500 characters, 69,632 papers in all, and asked whether the Discussion contained at least one em dash. The share rose from 4.23% before ChatGPT's release on 30 November 2022 to 11.58% afterward, an increase of 7.35 percentage points. The rise was gradual: near 4% through 2023, 8.0% in 2024 and 20.3% in 2025. The author is explicit that the dash is a population-level indicator, not a per-paper detector, and that the design cannot establish causality.

What a dash proves, and what it does not

Put the pieces together and the honest reading is narrow. A draft with many dashes was written by someone, or something, that likes dashes. People have used the mark for centuries, and style manuals treat it as a matter of taste. So a dash count can say that a draft resembles a certain kind of output. It cannot say who wrote it.

• It does not prove machine authorship. Professional writers and editors use the mark, and a writer who learned to type it decades ago did not stop in 2022.
• Its absence does not prove human authorship. Models that avoid the dash, or a quick find-and-replace, leave no dashes at all.
• Placement says more than presence. A dash used where a comma would do, wrapped in spaces, setting up a punchy reveal, matches the pattern the Wikipedia guide describes. A dash marking a real interruption is ordinary punctuation.
• Context says most. A dash next to stock vocabulary and a list of three is part of a cluster. Alone, it is a style choice.

How to mark dashes in a draft

1. Find every dash

   Copy one long dash from the draft into your editor's find box and step through the hits.

2. Name the job each one does

   In the margin, write what the dash is doing: aside, interruption, list, range, or none. A dash with no job is the one to mark.

3. Check the spacing and the house style

   Chicago's examples close the dash up; many newspapers put spaces around it. Mark only dashes that break the style the piece is supposed to follow.

4. Suggest the plain mark

   For a dash doing a comma's work, write comma. For an aside, parentheses. For a reveal, colon or a full stop. Standard proof marks for these changes are on the chart of proofreading marks for AI drafts.

5. Read the dash with its neighbours

   If the same sentence carries stock words or a list of three, mark the whole cluster as described in the guide to highlighting AI-sounding sentences.

A worked example: two dashes, two verdicts

Both sentences below were written for this page as examples. Each long dash is written out as [dash] so you can see where it falls.

Example A uses the dash for a real break in thought: the sentence turns on it, and a comma would flatten the turn. Label it aside or interruption, leave it alone. Example B uses the dash to set up a reveal inside a not just ... it's construction, with vague praise on the far side. Label it none and mark the whole sentence, because the dash is the least of its problems. A rewrite might read: The new schedule moves all meetings to Tuesday mornings.

Pitfalls when reading dashes as evidence

• Ignoring house style. Spaced dashes are normal in many publications. Check the style the writer was asked to follow before you mark spacing.
• Making it an accusation. A dash is never enough to raise a concern about authorship. The process for that is in the guide to fair markup for suspected AI writing.

Tips for writers who like the dash

• Keep the ones that earn it. A real interruption or a sharp aside is what the mark is for.
• Edit the whole sentence. If you remove a dash and the sentence still sounds like a slogan, the problem was never the punctuation. The AI writing guides cover the rest of the pass.

The em dash is a useful, old and slightly showy mark that some AI models overuse and others never touch. Studies confirm the rise across thousands of documents, and the same studies say the mark cannot identify a writer. When you mark a draft, give every dash a job label, change the ones with no job, and read each one with its neighbours. A dash inside a cluster of stock phrases is part of a pattern worth fixing. A dash on its own is punctuation.