# Em-dash density Part of [The AI Tells Index](https://feedsquad.com/ai-tells). A tell signals low effort. It does not identify an author. Skilled writers produce every shape listed here, some of them daily, and automated detectors misread those writers for it at rates measured above 60 percent on non-native English prose. Nothing in this index proves that a machine wrote anything. Read an entry as one piece of evidence to weigh against the false-positive notes printed beside it. ## Facts - Id: em-dash-density - Category: Formatting (formatting) - Subcategory: punctuation - Also known as: dash addiction, WP:AIDASH, em dashes per thousand words - Status: Contested. Credible people dispute that this signals anything at all. - Severity: medium - Evidence grade: corroborated (named independently by multiple credible secondary sources) - Languages: en - Added: 2026-08-15 - Updated: 2026-08-15 - Page: https://feedsquad.com/ai-tells/em-dash-density ## Description Em dashes arrive at a rate well above what the genre and the surface usually carry. The count is the unit here, not the mark. Spacing is a separate entry and this one measures density only. ## Why it reads as machine-written Model output uses the mark more often than non-professional human writing of the same genre, and puts it where a comma or a full stop would do the same work. The population evidence is solid and narrow. Czuma measured em dash presence in medRxiv Discussion sections rising from 4.23 percent of documents before ChatGPT to 11.58 percent after, odds ratio 2.96 with a confidence interval of 2.77 to 3.17, reaching 20.3 percent by 2025. That author states the measure is not a per-paper detector. A count on one document tells a reviewer where to look and nothing about who typed it. ## Detection Type: statistical (a measured metric against a threshold with a stated basis) Metric: em-dashes-per-1000-words Threshold: 10 (fires when above) Threshold basis: > Ten em dashes per 1,000 words is a review trigger, not a finding, and four things about its provenance ship with it. Freeburg is a single-author preprint whose human baseline is eight essays. The abstract runs from 0.0 for Llama to 9.1 for GPT-4.1 under suppression, so the trigger sits above the top of the abstract-verified model range. The unconstrained model top figure and the human mean are body-level table values this project could not confirm at page level, and a Crossref search for the same measurement in peer-reviewed venues returned nothing, which is how a preprint comes to stand alone here. The false-positive rate is unknown and no baseline exists for marketing or social copy. Czuma supplies the verified population shift, 4.23 percent to 11.58 percent of medRxiv Discussion sections with an odds ratio of 2.96, and states it is not a per-paper detector. ## Examples Constructed specimens. Written for this index. Never quoted from anyone's posts. ### Specimen 1 Before, exhibiting the tell: > The rollout went well—better than we planned—and support tickets dropped. Our onboarding flow—short and honest about limits—did most of the work. The lesson—start smaller—holds for any team. After, repaired: > The rollout went well and support tickets dropped by about a third in the first fortnight. The onboarding flow is four screens long, and screen two says that imports over 50MB will fail. That sentence killed most of our old tickets. Note: Figures in this repair are invented for the specimen. ## False positives Who legitimately writes this way. Essayists and literary writers have leaned on the mark for two centuries, and the published human range of 0.33 to 17.12 per 1,000 words contains every model mean anyone has measured. Two of the twelve models Freeburg tested emit none at all, so a low count says nothing either. Typographers, anyone drafting in a tool that auto-converts a double hyphen, and writers trained on Chicago style all produce high counts by preference. A reviewer who wants to use this number needs the writer's own earlier work as the baseline, because no baseline exists for marketing or social copy anywhere in the literature. ## Model attribution Reported per model, and the ordering moves. Freeburg's twelve-model table puts a Claude model and GPT-4.1 near the top and Llama at zero. The Economist reported in July 2026 that among contemporary models only Claude exceeded professional writers and that ChatGPT used fewer; this project read that claim second-hand through Wikipedia and labels it reported. OpenAI shipped em dash suppression in GPT-5.1 in November 2025, so any per-family claim needs a date attached to it. ## Platform notes - linkedin: The claim that LinkedIn penalises em dashes has no first-party source. No LinkedIn policy or newsroom post this project read mentions punctuation at all. - wikipedia: Shortcut WP:AIDASH. The guide says the sign is most useful alongside other indicators and that it turns up far more on discussion pages than in article text. ## Sources 1. Czuma, Em-ergence of the em-dash: a population-level rise in em-dash frequency in medRxiv preprints, measured on Discussion sections (arXiv:2606.29540) https://arxiv.org/abs/2606.29540 (tier: primary-doc; accessed 2026-08-14) 2. Freeburg, em dash rates across twelve models, single-author preprint (arXiv:2603.27006) https://arxiv.org/abs/2603.27006 (tier: primary-doc; accessed 2026-08-14) 3. The Economist, how to spot AI writing (30 Jul 2026) https://www.economist.com/culture/2026/07/30/how-to-spot-ai-writing (tier: press; accessed 2026-08-14) 4. Wikipedia: Signs of AI writing https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing (tier: community; accessed 2026-08-14) 5. Washington Post interactive, how to detect ChatGPT em dashes https://www.washingtonpost.com/technology/interactive/2025/how-detect-chatgpt-em-dash/ (tier: press; accessed 2026-08-14) ## Status history Ids are permanent. A retired tell keeps its id and its page. - 2026-08-15, Contested: Opened as contested. The population shift is measured and holds up: Czuma reports em dash presence in medRxiv Discussion sections rising from 4.23 to 11.58 percent with an odds ratio of 2.96, and its author says the measure does not work per paper. Against that, the published human range of 0.33 to 17.12 per 1,000 words contains every model mean measured so far, two of twelve models emit none at all, and OpenAI shipped suppression in GPT-5.1 in November 2025. A per-document verdict is not available from any of it. ## License CC BY 4.0. https://creativecommons.org/licenses/by/4.0/ Attribution: The AI Tells Index, feedsquad.com/ai-tells Reuse the data, including commercially. Keep the attribution line. --- Dataset version 1.0.0. Schema version 1. Part of [FeedSquad](https://feedsquad.com). Built by [Herman Foundry](https://hermanfoundry.com) from Levi, Finnish Lapland.