Some people have always written with the long dash: to break off a thought, add an aside or land the last word. Since ChatGPT arrived, they have found that habit read as a confession. This is EdCitation's view, and we hold it firmly: the em dash is not proof that a person used AI. Nor is the word "delve", a sentence without a slip, or a neat list of three. People had every one of these habits first, and the models learned them from us. The large studies that find more of them since 2022 describe millions of documents, not one person's essay.
What can be settled about a paper is narrower and more useful: whether its sources exist. EdCitation's free Verify references settles that from the publisher's record. Check AI, part of Pro, shows a writer how a detector reads each passage, and it is an estimate, never proof.
Is the em dash a sign that AI wrote something?
No, not on its own and not about any one text. The em dash (—) is the long dash that sets off part of a sentence more sharply than a comma. Some language models use it heavily, and whole bodies of writing now contain more of it. Neither fact turns one dash into evidence about one writer.
What the counts show
The most careful measurement is a pre-registered study of 69,632 medRxiv preprints. Czuma (2026) found that the share of Discussion sections containing at least one em dash rose from 4.23% before ChatGPT's release to 11.58% after it, reaching 20.3% in 2025.
The author calls the mark "not a per-paper detector" of model use, and the study's own data show why. In an exploratory look at 2025, the 57 preprints that declared no AI use contained an em dash about as often as the rest (24.6% against about 20%), and most that did declare AI use contained none.
The rate depends on the model
Freeburg (2026) had twelve models write essays. GPT-4.1 produced 10.62 em dashes per 1,000 words, Meta's two Llama models none, and OpenAI's newer GPT-5.4 1.43. Eight published human essays averaged 3.23 but ranged from 0.33 to 17.12, higher than any model reached in any condition of the study. Both authors' papers are preprints.
The belief moves too. In a 2025 study, expert readers described AI text as grammatically perfect and sparing with dashes, and one called an article human for its varied dashes and brackets (Russell et al., 2025). The same year, Inside Higher Ed ran an opinion piece on academics reading an unspaced dash as a sign of AI (Mellors, 2025).
Who used the em dash before AI?
Novelists and poets, at least since the 1750s, and the style guides still teach it.
In novels and poems
Laurence Sterne's Tristram Shandy, published in nine volumes from 1759 to 1767 (Laurence Sterne Trust, n.d.), is built on dashes. We counted them on 23 September 2026 in Project Gutenberg's transcription (Sterne, 1759–1767/1997): 9,277 separate dashes in about 181,000 words, roughly 51 per 1,000. The same count gives about 11 per 1,000 in Jane Eyre (Brontë, 1847/1998). A transcription does not keep every printer's choice and our count is not Freeburg's method, so the comparison is rough; the highest model rate in that study was 14.03.
Emily Dickinson used dashes throughout her manuscripts, and her first editors took many out. The Emily Dickinson Museum (n.d.) records that Mabel Loomis Todd and Thomas Wentworth Higginson regularised her punctuation for the first volume of her poems in 1890, and that Thomas H. Johnson went back to the manuscripts for his 1955 edition. Harper (2019), on the Chicago Manual of Style's blog, shows the 1890 printing of "Because I could not stop for Death" with a semicolon where she had a dash, and transcribes a line of her manuscript in Chicago's form: Since then—'tis Centuries—and yet. Her handwritten dashes, Harper notes, look more like spaced hyphens; the unspaced em dash was the printed convention of her time.
In the style guides
APA's punctuation page, created in 2019, already warned that many writers overuse dashes (American Psychological Association, 2019). Chicago's editors describe the em dash as setting off part of a sentence much as parentheses do, with no space either side, and point out that Microsoft Word turns two typed hyphens into one by default (Chicago Manual of Style, 2024). The unspaced dash some readers now call machine-made is Chicago's convention for books.
What do studies say about other AI writing tells?
They find real shifts across very large collections of text, and the largest say plainly that their methods cannot pick out one writer.
| Supposed tell | What was measured | What it cannot show |
|---|---|---|
| "Delve" and other style words | Kobak et al. (2025): across 15.1 million PubMed abstracts, "delves" appeared at 28 times its expected 2024 frequency; at least 13.5% of 2024 abstracts were processed with a language model | The authors say it cannot identify individual abstracts |
| "Commendable", "meticulous", "intricate" | Liang et al. (2024): in ICLR 2024 peer reviews these became 9.8, 34.7 and 11.2 times more likely in a sentence; 6.5% to 16.9% of review text at four AI conferences may have been substantially modified by a model | An estimate for the whole collection, not for one review |
| Formal transitions | Kobak et al. (2025): their ten common marker words include "additionally" and "notably", and also "across" and "within" | Ordinary words count only in bulk: nobody is suspect for writing "within" |
| Perfect grammar, a neutral tone | Russell et al. (2025): readers who seldom used AI for writing did about as well as chance, taking neutral tone and unusual words as signs of a machine | Careful writers and permitted grammar tools produce clean prose too |
| Lists of three | Expert readers in the same study named lists of three as a clue | The triple is far older than software |
| A reference that does not exist | Whether it exists is a look-up in the publisher's record | Not a guess: Verify references checks each one, free |
People now speak like the models
Yakura et al. (2026) analysed 737,083 hours of conversation from 824,634 podcast episodes: words ChatGPT favours, among them "delve" and "meticulous", rose abruptly in unscripted speech after its release, and in an experiment with 496 people, a short chat with a chatbot led participants to adopt its words. A person who writes "delve" may simply have picked it up, and Kobak et al. (2025) say their method cannot separate a model's word from a borrowed one.
Lists of three and clean sentences
Lincoln ended his address at Gettysburg on 19 November 1863 with government "of the people, by the people, for the people" (National Museum of American History, n.d.). No language model taught him the triple. Clean grammar is what editing produces, yet detectors tend to read smooth, predictable prose as machine-made, as How AI detectors work shows.
Can a trained reader spot AI writing by its style?
Sometimes, as a panel, and not reliably alone. Russell et al. (2025) asked people to judge 300 non-fiction articles. Five who often used AI for writing, voting together, misjudged only 1 of the 300; four who seldom did, tested on the first 60, did about as well as a coin toss and called roughly half the human articles AI. When the experts wrongly accused a human article, 31% of their explanations pointed to vocabulary, typically a human writer's "delve" or "crucial".
That is the disagreement to name. Some researchers see trained human judgement as a better detector than software; Washington State University (n.d.) reads it the other way. Noting that even experienced readers wrongly flag human work at 4% or more, it endorses no AI detection tool and states that suspicion of AI use is not enough for a finding. Our view is closer to the university's: five practised readers voting on 300 articles are not one marker's impression of one essay.
What do universities say about a detector score or a hunch?
That it can start a question and cannot settle it, though they differ on how much weight a marker's own judgement may carry.
The Office of the Independent Adjudicator for Higher Education (2025), which reviews student complaints in England and Wales, puts the burden of proof on the university. Its casework note asks universities to consider whether an assumption of AI use could be bias against the way a student writes, naming students writing in a second language and disabled students, and to ask students for their notes and drafts. The University of Portsmouth (n.d.) does not currently allow AI detection tools and wants assessment built on credible evidence of authorship.
Where they differ
Royal Holloway, University of London (n.d.) leans the other way: its regulations treat identifying offences such as commissioning as expert academic judgement, and say an offence can be found from the student's work alone. Its guidance still asks students for drafts, search history and copies of sources, and wants evidence that is clear, attributable to the student and dated.
What real cases rest on
Munoz et al. (2026) coded 1,855 pieces of evidence in 1,162 AI-related misconduct cases at one regional Australian university from 2023 to 2025. Detector outputs were rated uniformly weak, and "AI-typical" writing patterns and changes of style weak to moderate, because they are not unique to AI use. In 58 cases (5.0%), the allegation went ahead on the assertion alone. Fabricated references rose from 10.4% of the evidence in 2023 to 30.0% in 2025; they can be checked against databases, though the authors rate them moderate on their own, since a student can invent a reference without AI.
What can you show if you are accused of using AI?
The record of how you wrote, which no punctuation mark can contradict. In order:
- Ask, in writing, what the concern rests on: a score, the style, particular passages, the references. If the answer is the dashes, it is fair to ask what else there is.
- Leave the submitted file as it is, and export the version history from Google Docs or Word.
- Gather drafts, outlines and notes. Royal Holloway's test suits any university: clear, yours and dated.
- Show the sources you read: annotated copies, library loans, saved searches. Run the paper through Verify references first, free and with no account, so nothing in the list surprises you.
- Bring older writing. If you have used the long dash for years, an essay from before 2022 shows it better than any argument.
- Offer to talk the work through: the argument, and why you chose each source.
What to do if an AI detector wrongly flags your writing sets out the full process, and AI detectors and non-native English writers covers suspicion that is really about a second language.
Should you stop using the em dash?
No, not to escape suspicion. Use it less if your style guide or your teacher asks, and for that reason. A writer who drops the dash to satisfy a hunch will next be asked to drop "notably", then the list of three, then the clean sentence, and every item on that list is something good writers do. Passing honest work through a rewriting tool is worse still, as our guide to AI humanizers explains. Write the way you write, and keep your drafts.
How can EdCitation help if your writing is questioned?
By checking what can be checked. For the one question in an AI dispute with a factual answer, whether each reference exists, EdCitation's Verify references is the best tool there is, because it never composes a reference: it searches for each one in the publisher's record, which no chatbot can promise to do. The cases Munoz et al. (2026) studied show why that matters: by 2025 fabricated references made up 30.0% of the evidence, and they could be checked against a record. Paste a list or upload the paper, free and with no account. The result for each entry is one of three: verified, "check this", or not found. Retracted papers carry a flag, "could not check" is never passed off as "not found", and with the whole paper uploaded, each in-text citation is paired with its entry and each entry with the text.
Check AI comes with Pro, $8 a month including 240 credits. It sorts the passages of your text into AI-written, AI-assisted and human, with a score beside each, in a report no one else sees. Before it runs you are told how many credits it will take, and a check that fails takes none. It is an estimate: another detector, your university's included, may read the same text differently, and no score is proof of anything. Pricing sets out Pro, and Max at $24 a month, which adds Theoretics QA and the Library.
Cite a source builds references from the record: given the DOI of the Kobak et al. (2025) paper on 23 September 2026, it returned the APA 7 entry from Crossref's record, volume, issue and article number included, with no retraction notice. When a sentence needs support, Find sources takes the claim as the search, across about 300 million published works.
Quick questions
Does using em dashes mean I used ChatGPT?
No. Writers have used the long dash since at least the 1750s, Microsoft Word makes one from two hyphens by default, and a pre-registered study of 69,632 preprints says the mark cannot tell whether any single paper was written with AI.
Is "delve" an AI word?
It is a word language models favour: Kobak et al. (2025) found "delves" at 28 times its expected frequency in 2024 abstracts. It is also ordinary English that people now say more often, so one "delve" proves nothing.
Can a teacher tell AI writing from its style?
Not reliably alone. In Russell et al. (2025), readers who seldom used AI did about as well as chance, and expert readers' false accusations often rested on human writing that happened to contain words like "delve".
Should I remove em dashes from my essay before submitting?
Only if your style guide or instructions ask you to. Changing punctuation proves nothing about who wrote the essay; drafts, version history and notes do.
What can I check before I hand in?
Your references, which EdCitation's free Verify references checks one by one against the publisher's record. Check AI, part of Pro, shows how a detector reads each passage, as an estimate only.
References
- American Psychological Association. (2019). Punctuation. APA Style. https://apastyle.apa.org/style-grammar-guidelines/punctuation
- Brontë, C. (1998). Jane Eyre: An autobiography. Project Gutenberg. (Original work published 1847) https://www.gutenberg.org/ebooks/1260
- The Chicago Manual of Style. (2024, January 23). Hyphens and dashes: A refresher. CMOS Shop Talk. https://cmosshoptalk.com/2024/01/23/hyphens-and-dashes-a-refresher/
- Czuma, P. (2026). Em-ergence of the em-dash: A population-level rise in em-dash frequency in medRxiv preprints at the dawn of the large-language-model era [Preprint]. arXiv. https://arxiv.org/abs/2606.29540
- Emily Dickinson Museum. (n.d.). The posthumous discovery of Dickinson's poems. https://www.emilydickinsonmuseum.org/emily-dickinson/poetry/the-poet-at-work/the-posthumous-discovery-of-dickinsons-poems/
- Freeburg, E. M. (2026). The last fingerprint: How Markdown training shapes LLM prose [Preprint]. arXiv. https://arxiv.org/abs/2603.27006
- Harper, R. (2019, October 8). A dash of poetic license. CMOS Shop Talk. https://cmosshoptalk.com/2019/10/08/a-dash-of-poetic-license/
- Kobak, D., González-Márquez, R., Horvát, E.-Á., & Lause, J. (2025). Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Science Advances, 11(27), Article eadt3813. https://doi.org/10.1126/sciadv.adt3813
- Laurence Sterne Trust. (n.d.). The life and opinions of Tristram Shandy, gentleman. https://www.laurencesternetrust.org.uk/collection-highlights-categories/the-life-and-opinions-of-tristram-shandy-gentleman/
- Liang, W., Izzo, Z., Zhang, Y., Lepp, H., Cao, H., Zhao, X., Chen, L., Ye, H., Liu, S., Huang, Z., McFarland, D. A., & Zou, J. Y. (2024). Monitoring AI-modified content at scale: A case study on the impact of ChatGPT on AI conference peer reviews. In Proceedings of the 41st International Conference on Machine Learning (Vol. 235, pp. 29575–29620). PMLR. https://proceedings.mlr.press/v235/liang24b.html
- Mellors, J. (2025, June 20). The em dash is not the problem [Opinion]. Inside Higher Ed. https://www.insidehighered.com/opinion/views/2025/06/20/whats-em-dashai-anxieties-opinion
- Munoz, A., Hinchcliff, M., Langfield, C., & Rogerson, A. (2026). How strong is the evidence in generative AI-related academic misconduct allegations? A mixed-methods analysis. International Journal for Educational Integrity, 22(1), Article 26. https://doi.org/10.1007/s40979-026-00235-9
- National Museum of American History. (n.d.). The White House copy of the Gettysburg Address. Smithsonian Institution. https://amhistory.si.edu/documentsgallery/exhibitions/gettysburg_address_2.html
- Office of the Independent Adjudicator for Higher Education. (2025). Casework note: Complaints relating to AI and academic misconduct. https://www.oiahe.org.uk/resources-and-publications/learning-from-our-casework/ai-and-academic-misconduct/casework-note-complaints-relating-to-ai-and-academic-misconduct/
- Royal Holloway, University of London. (n.d.). Student guidance: Artificial intelligence (AI) allegations [PDF]. https://intranet.royalholloway.ac.uk/students/assets/docs/pdf/academic-investigations/2425/student-guidance-for-ai-allegations.pdf
- Russell, J., Karpinska, M., & Iyyer, M. (2025). People who frequently use ChatGPT for writing tasks are accurate and robust detectors of AI-generated text. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 5342–5373). Association for Computational Linguistics. https://doi.org/10.18653/v1/2025.acl-long.267
- Sterne, L. (1997). The life and opinions of Tristram Shandy, gentleman. Project Gutenberg. (Original work published 1759–1767) https://www.gutenberg.org/ebooks/1079
- University of Portsmouth. (n.d.). AI guidance for staff. https://cadi.port.ac.uk/ai-guidance-staff
- Washington State University, Office of the Provost. (n.d.). Detecting and reporting misconduct related to generative AI. https://provost.wsu.edu/policies/artificial_intelligence/detecting-and-reporting-misconduct/
- Yakura, H., Lopez-Lopez, E., Brinkmann, L., de la Serna, I., Kirfel, L., Gupta, P., Soraperra, I., Eisenmann, T. F., Wulff, D. U., & Rahwan, I. (2026). Empirical evidence of large language model's influence on human spoken communication [Preprint]. arXiv. https://arxiv.org/abs/2409.01754