Who will speak for the em dash?

By now everyone has met the heuristic: if a piece of writing contains an em dash, a machine wrote it. It has a more precise origin than you might expect. Recruiters and LinkedIn posters were circling the idea by late 2024, but the name arrived in March 2025 by an unusually silly route: a commenter beneath a fast-fashion brand’s rebrand post complained about the “ChatGPT hyphen” in its ad copy, a podcast clip picked the phrase up, and it stuck. Keith Houston, who wrote the book on punctuation marks, has traced the complaints across LinkedIn, Reddit, Instagram and OpenAI’s own developer forum. Some eighteen months on, the suspicion has hardened into folk wisdom: a typographic convention older than the United States, quietly repealed by vibes.

I have been meaning to write something in defence of the em dash for the best part of a year. When I finally sat down to do it, I made the mistake of searching the phrase first. It turns out to be a genre. A student opinions editor at the University of Saskatchewan argues that the dash is where her voice lives. A Substack essay mounts its defence by deploying em dashes in more or less every clause, which is either principled or hostile and I admire it either way. Air Mail has assembled Sofia Coppola, Michael Lewis, Monica Lewinsky and Seth Meyers to give evidence on behalf of a punctuation mark. An undergraduate has petitioned Harvard, in the Crimson, to take a stand. Joel Stein has entered a defence in the Wall Street Journal. NPR has been out reporting on what it describes as the backlash to the backlash. And Nitsuh Abebe in the New York Times Magazine asks the properly unsettling question: whose punctuation is more human, yours or the machine’s?

The genre also has infrastructure. There is a campaign site, Reclaim The Em Dash, whose subtitle — Against Irresponsible Stigmatisation of Long Overlooked Punctuation — appears to exist chiefly so that it can spell AISLOP. There is also Save the Em-Dash, which is an arcade game where you catch the human writing and dodge the slop. Someone registered the domains. Someone else wrote the game.

So the em dash has plenty of defenders, myself among them. But it is losing anyway. Which suggests defence is the wrong posture.

I learned the difference between these marks the way a certain generation of computer scientists did — from LaTeX — and then, rather less predictably, from Fowler’s. Copies of the Dictionary of Modern English Usage were floating around the lab, in among the formal languages, the grammars, the typography and typesetting, and the arcane LaTeX incantations for squeezing a paper back under its page limit. If you have written a thesis or a paper in LaTeX you already know that a hyphen is -, an en dash is --, and an em dash is ---. You know it because you typed it several thousand times, and because the first time you got it wrong the output looked subtly, unaccountably cheap, and you couldn’t say why until somebody showed you. Markup taught me to see the difference by making me explicitly declare my intent, every single time. It is also why, on the morning I submitted my undergraduate dissertation — typeset in LaTeX on the Red Hat Linux side of a dual-boot laptop, and, as it happens, a project on building a compiler front end — I found that the library printers had dropped every fi ligature in the document. Reprinting three copies on submission morning was the greater risk, so I went through them by hand in black pen.

And if you want evidence that this argument long predates the current panic, the Hacker News thread on the em dash from January 2018 runs to more than two hundred comments on spacing conventions, compose keys, thin spaces and TeX ligatures, with not one reference to a language model — because there was not yet one to blame.

For the avoidance of doubt:

• Hyphen (-, U+002D). Joins. Well-known, co-author, twenty-three. Shortest mark, smallest job.
• En dash (–, U+2013). Spans and connects. 2011–2026. The Bristol–Bath corridor. Roughly the width of an n, which is genuinely where the name comes from.
• Em dash (—, U+2014). Interrupts, digresses, reveals. The width of an m. The one that changes the gear of a sentence rather than merely joining its parts.

Three marks, three distinct jobs, none of them interchangeable, and all of them routinely treated as though they were (Merriam-Webster has the authoritative version, should you want it from a dictionary rather than from me). Microsoft Word definitely deserves a share of the blame here: it has spent the best part of thirty years making this decision on everyone’s behalf, silently swapping in whichever dash its autocorrect assumed was meant. Several generations have been spared the choice and therefore never discovered there was one to make.

Since apparently this needs saying out loud: on a Mac it is ⌥ + ⇧ + - for an em dash and ⌥ + - for an en dash; on Linux, Compose then - - -, or Ctrl + Shift + U, then 2014; on Windows, Alt + 0151 on the numeric keypad. Ten seconds of learning, permanently better prose. There are few deals this good.

There is also a live schism, for anyone who wants to fight about it in the comments. American house style closes the dash up against the words either side—like that. British practice spaces it – like that – often with an en dash, which has always struck me as the punctuation equivalent of apologising for taking up room. I should also admit that fifteen years of this blog are not remotely consistent on the point, particularly with en dashes, and that some of the drift is not even mine: it is an artefact of XHTML entities, the WordPress classic editor, and a succession of themes and blog rebrands, all of which have held firm views of their own about – and —, and little patience with mine.

Which raises something I have been trying to ignore for most of this piece. Almost everyone rushing to the em dash’s defence is American. The campaign sites, the celebrity depositions, the newspaper columns, the arcade game: all of it comes from a culture that uses the mark properly, unspaced and unembarrassed. British practice, as above, hedged long ago. We are spectators at a foreign rescue operation for a piece of punctuation we had already half given up on — which may be why the row has been slower to arrive here, and why I have been sitting on this post for a year.

What actually irritates me is that the accusation is backwards. The models did not invent the em dash. They acquired the habit from us. They were trained on people who could write, and people who can write use em dashes, because the em dash does something no other mark does. Dickens used them. The poet Emily Dickinson used them so heavily that her first editors quietly tidied them away, and it took until the 1950s for a scholarly edition to put them back. To say “this contains an em dash, therefore a machine wrote it” is to say “this is punctuated too well to be human” — which is a remarkably bleak thing to believe about your own species.

We can be fairly specific about where the habit came from, because some of the people it was taken from are still here and still cross about it. Karen Stabiner, in the Los Angeles Times, notes that she is one of the authors whose books were scraped, and therefore one of the writers the chatbots learned this from. She has since finished another book — sixty-three thousand words, em dashes throughout — and cannot decide what to do with it: strip them out, or print a disclaimer on the title page saying no AI was involved, which she rightly observes would only make matters worse.

And the loop has now closed. In November 2025 Sam Altman announced, as a small win, that if you instruct ChatGPT not to use em dashes it will finally comply. So: the machines learned the mark from good writers, readers took the mark as proof of the machine, and the fix has been to train the machine out of it. Somewhere in that circle a piece of punctuation is being quietly retired from general use, and nobody involved has made an argument that it deserved to go.

Abebe’s version of this in the New York Times Magazine is less flattering and harder to shake. His argument is that the dash is not really the issue. For most of modern history the written material people encountered had been composed: drafted, revised, edited, often typeset. That is no longer true. The overwhelming majority of text most of us read in a day is peer-to-peer — messages, posts, comments, replies. Set against that baseline, prose that has visibly been worked on no longer registers as care. It registers as artifice, and artifice now has an obvious suspect. The em dash is only the visible tip. What people think they are detecting is a machine; what they are actually detecting is composition.

The heuristic also does damage on its way past. The features now filed under “AI tell” — long subordinate clauses, parenthetical asides, a fondness for the dash — are not diagnostic of a machine. They are diagnostic of a register — and I notice I qualify on all three counts myself, which is either solidarity or an admission. Neurodivergent writers have been among the loudest in pointing this out, and there is now a small but growing literature behind them: a study of the misclassification of autistic writing as AI-generated, a book chapter on the false-positive rates of detection tools and what they cost the students who trip them, and a review in the journal Teaching in Higher Education on neurodiversity and academic integrity which makes the essential point that these false positives are not randomly distributed. Nouse, the student paper at the University of York, has reported on students falsely accused on the strength of statistical signatures such as perplexity and burstiness — how predictable your vocabulary is, how much your sentence lengths vary — alongside penalties for AI-related misconduct more than doubling in a single year.

What is firmly established is worse. A Stanford study of seven commercial GPT detectors found that more than half of TOEFL essays — that is, English-proficiency essays by non-native speakers — were misclassified as AI-generated, while essays by US school pupils were judged almost perfectly. The same work found that enriching the non-native writing reduced the false positives, and simplifying the native writing increased them.

So the detectors and the readers are running opposite heuristics. Low perplexity — plain vocabulary, predictable rhythm — is what the software flags. Visible effort is what the readers flag. That is a double bind in the strict sense: two authorities issuing incompatible instructions, a penalty attached either way, and no position left from which to write that satisfies both. In a university that judgement carries consequences well beyond a bruised ego. In public writing and political communication it is becoming something cheaper and more corrosive: a way of dismissing an argument without engaging with it. Accuse the author of not having written the thing, and nobody has to deal with what the thing actually says. Stein’s Wall Street Journal piece opens with exactly this: his son, a student journalist, chasing a tip that the school administration had banned AI and then used it to write an email to parents. Two detection tools put the odds at zero. The entire case rested on the presence of em dashes.

Worse still is what all this does to people who have not been accused of anything. I have done it myself: gone back through work emails, and even posts on this blog, counted the em dashes, and quietly removed a few. Not because the sentences read better without them, but because of how the paragraph might look. That is a Turing test run backwards — proving your humanity by writing slightly worse on purpose.

I am not alone in it. Wired has reported on an emerging anti-AI counterstyle: novelists taking deliberately winding routes around a dash or a rule-of-three list; a Clickhole listicle that opens by announcing, proudly, that no AI whatsoever was involved in its composition, and then proceeds in deliberate misspellings — its author had to undo her own autocorrect to keep them in; and an app billed as an anti-Grammarly, which leaves the typos in your emails and signs them off from your iPhone. In one test, the only messages Fortune 500 CEOs bothered to open were the ones with lowercase, misspelled subject lines.

None of which fixes anything. If your prose is bland, structurally identical throughout, and reaches for the same three rhetorical moves in every paragraph, that is the tell, and it was the tell long before any of this. The dash is simply the part people can see. One of the writers Wired spoke to puts it better than I can: the only counterstyle that actually works is writing well.

There is a real limit, and it is worth stating plainly. The em dash works by interruption, and an interruption only registers against something continuous. Use one in every sentence and it stops signalling anything at all: the reader adjusts, the effect flattens, and you are left with a mannerism where an emphasis used to be. Used well, it should feel like the writer turning to you and briefly lowering their voice.

But if the em dash does disappear, it will not have been taken from us by the machines. We will have given it up ourselves, one draft at a time, to avoid being mistaken for something that learned it from us in the first place.

Option, Shift, hyphen. It is not much of a stand, but it is the one available.