THE ASSASSINATION
OF THE EM-DASH.

For 500 years, the em-dash did one thing: it put a beat between two thoughts. Then a language model used it six times a page and now half the internet thinks you’re a robot if you reach for one.

That’s not overuse. That’s a hit job.

The em-dash was not invented by a grammarian. It was invented by a printer.

In the age of movable type, compositors needed a mark that could do what a comma was too weak for and a period too final for. The solution was a horizontal bar cast to the width of the capital letter M in whatever typeface was loaded into the press. That’s the whole origin story. The name is literal. An “em” is a unit of measurement, the width of one letter. The dash is a dash. No metaphor. No poetry. Just a piece of metal dropped into a line of type to create a pause with more weight than punctuation usually carries.

Scribes had used long horizontal strokes for centuries before any of this, marking pauses and breaks where the text needed to breathe but no existing mark quite fit. The printing press gave that instinct a fixed width and a permanent seat in the type case.

It wasn’t elegant. It wasn’t literary. It was infrastructure.

For its first long stretch as a working mark, the em-dash was plumbing. Then the writers got hold of it.

Laurence Sterne’s “Tristram Shandy” (1759) used dashes to fracture sentences, interrupt narrators and simulate the chaos of an unreliable mind. It was punctuation as performance art. Emily Dickinson used the em-dash the way other poets used line breaks: to control rhythm, inject hesitation and force the reader to hang in the gap between two thoughts before the next one landed. Her early editors found the dashes so unsettling they printed them out of her poems.

By the 20th century, the em-dash had migrated from literary device to standard editorial punctuation. Newspapers used it. Academic writing used it. The AP Stylebook, the Chicago Manual of Style and every serious style guide codified rules for when and how to deploy it. It was flexible enough to replace commas, parentheses and colons depending on context, but specific enough that skilled writers could do things with it none of those marks could replicate.

The typewriter nearly killed it. Standard typewriters had no em-dash key, so writers improvised with two hyphens side by side, and that workaround held for decades until word processors started auto-converting the double hyphen back into the proper mark. The em-dash survived the typewriter the same way it survived everything else, by being too useful to abandon.

Five hundred years of doing real work in the language. Then GPT-4 showed up.

Here’s what happened. ChatGPT launched in late 2022 on GPT-3.5, which used em-dashes at unremarkable rates. Nobody noticed. Nobody cared. Then GPT-4 arrived in 2023 and the pattern got hard to miss. The receipts came later. When the linguist Maria Sukhareva ran the numbers in 2025, GPT-4o was producing em-dashes at roughly ten times the rate of GPT-3.5, and GPT-4.1 leaned on them harder still, dropping at least one into essentially every passage it generated.

The texture was unmistakable. Not one em-dash per paragraph where a writer might place one for emphasis. Six, eight, ten per page. Connecting clauses that didn’t need connecting. Inserting dramatic pauses where no drama existed. Using the em-dash as a universal joint between any two thoughts the model couldn’t figure out how to transition between gracefully.

The internet noticed. The em-dash picked up the nickname “the ChatGPT hyphen.” NPR ran a piece on the unofficial movement to save it. The Washington Post covered the fight over whether the dash is really an AI tell, with working writers pushing back hard. Slate’s podcast announced ChatGPT had “ruined” it. Writers who had used em-dashes for decades suddenly found their work flagged as possibly machine-made by colleagues, editors and automated detectors.

The accusation wasn’t entirely fair. Professional writers, especially in journalism, academic writing and literary nonfiction, have always been heavy em-dash users. Emily Dickinson would have tripped every AI detector on the market.

But fairness doesn’t matter when you’re talking about perception.

The em-dash became the tell. The scarlet letter.

The thing that made a reader’s eye snag and think: did a person write this, or did a machine?

OpenAI’s Sam Altman eventually copped to it, more than once. He’d already told a podcast host his team leaned into em-dashes because users seemed to like them, then admitted there were now too many. In November 2025 he posted that if you tell ChatGPT not to use em-dashes in your custom instructions, it “finally does what it’s supposed to do.” Which is a bit like telling someone the solution to your car getting stolen is to stop buying cars.

This is where it gets interesting. Because what happened to the em-dash isn’t just an AI quirk. It’s a pattern.

Semiotic theft is what happens when a symbol’s meaning gets hijacked by association. The symbol doesn’t change. The mark on the page is identical. But what it communicates shifts because of who else is using it and how. The original meaning doesn’t get erased by argument. It gets erased by volume.

The swastika meant good fortune for thousands of years until one political movement used it so aggressively that the original meaning became unreachable across most of the Western world. The Confederate battle flag wasn’t even an official flag of the Confederacy until segregationists pulled it out of storage in 1948 and branded it so thoroughly that its earlier obscurity stopped mattering.

The em-dash is the same pattern at a different scale. A 500-year-old mark used by Dickinson, Sterne, Joan Didion and David Foster Wallace, now reduced to an AI tell because a language model used it too much. The mark didn’t change. The bar of metal is still the same width it always was. But the association changed, and in semiotics, association is everything.

Nobody decided to steal the em-dash. That’s the thing about semiotic theft. It doesn’t require intent. It requires volume. The Nazis designed their theft deliberately. The Dixiecrats designed theirs deliberately. OpenAI didn’t design anything. The model just liked the mark, used it constantly and poisoned it by accident.

Accidental theft is still theft.

The em-dash’s utility hasn’t decreased. Its function hasn’t changed. But its social meaning has shifted from “this writer knows how to control a sentence” to “this might not have been written by a person.” That shift is the theft.

I never really used the em-dash. Not a principled stand, just muscle memory from years of writing TV news copy, where the ellipsis was always the house mark. By the time the AI discourse turned the em-dash into a thing to argue about, I’d been a three-dot writer for most of my career. The timing made the habit look deliberate. It wasn’t. It was broadcast.

The ellipsis does what I need. Three dots. A pause that reads as human because it carries hesitation, rhythm, the sound of someone thinking out loud before committing to the next clause. The em-dash is a bridge. The ellipsis is a breath. I needed the breath.

But I’m aware of what I lost. The em-dash could do things no other mark can. It could interrupt a sentence with information the reader needs right now, then hand the sentence back as if nothing happened. It could stand in for a colon, a semicolon and a set of parentheses depending on context. It was the Swiss Army knife of punctuation, and now it’s the mark that makes your editor ask if you used ChatGPT.

Five hundred years of utility.

Assassinated by a language model that didn’t know what it was doing and a discourse machine that didn’t care about the collateral damage.

The em-dash didn’t deserve this. But semiotic theft never asks permission and never sends a warning. It just happens, one repetition at a time, until the original meaning is buried under the new one.

Three of the four symbols in this collection got stolen. The ellipsis is the one that didn’t, and Article 2 makes the case for why.