A eulogy for the em-dash

We are gathered here today to mourn the em-dash, which died not from disuse but from overuse — by something that was never alive.

Farewell. We hardly knew ye. And then we all knew ye, all at once.

The em-dash served English prose for centuries. Emily Dickinson used it to fracture syntax into thought. It let you pivot, expand, or exemplify mid-sentence — like this — without pedantry or bureaucratic formality (in the case of parentheticals), or the flat, breathless, slightly panicked feeling of trying to do the same job with commas. It was the Swiss army knife of punctuation: a pause that wasn’t a full stop, a pivot that wasn’t a new sentence, a way of saying “and another thing” without actually saying it.

Now it’s an AI tell. If your writing contains more than two em-dashes per paragraph, someone on Twitter will accuse you of using ChatGPT. The em-dash didn’t change. The readership did. A generation of people trained to detect AI prose has learned that language models use em-dashes constantly, and has concluded that em-dashes are therefore suspicious. The tool didn’t kill the em-dash. The pattern-matching did.

The em-dash is not alone. Let us read the names of the other fallen.

The semicolon. Once the mark of a writer who could hold two related thoughts in tension without collapsing them into one sentence or separating them into two. Now it signals “this was probably generated.” Real humans use full stops. Or write in sentence fragments without subjects. The semicolon is the domain of machines and Victorian novelists; these are apparently the same category now.

The tricolon. The three-part list, the rhetorical triple, the structure that underpins “life, liberty, and the pursuit of happiness” and “government of the people, by the people, for the people.” Language models produce tricolons compulsively because three items satisfy a pattern-completion instinct baked into the training data. As a result, any human who naturally writes in threes sounds automated. You can write two things. You can write four. Three is for robots.

The clarifying restatement. “In other words.” “Put simply.” These phrases exist because sometimes a thought needs to be said more than once, in different ways, to land properly. That is to say, it’s when adding another tangent helps to better describe a circle. Language models use them as transitions between paragraphs they’ve already decided to write. And don’t get us started on the loose analogies that follow the restatement. A human writer who reaches for an analogy now has to check it against the catalogue of dead comparisons the models have worn smooth, and choose something unexpected enough to not sound generated. The burden of originality has shifted from “make it clear” to “make it clear in a way a machine wouldn’t.” The phrases haven’t changed. Their connotation has. Using one now is like wearing a trench coat to a job interview — technically fine, but everyone’s still going to wonder.

The colon-expansion. The sentence that introduces a concept, then unpacks it after a colon: exactly like this one. Writers have been doing this since at least the King James Bible. It’s also the single most common structural move in AI-generated prose. Every model, every provider, every temperature setting. The colon-expansion is to language models what the twelve-bar blues is to guitar players — so fundamental that using it marks you as either a master or a beginner, and nobody can tell which.

The hedged conclusion. “The answer is nuanced.” “The truth is somewhere in between.” “It depends on the context.” These were once signs of intellectual honesty. Now they’re signs that the model couldn’t commit to a position because its training penalised strong claims. Hedging has been fully ceded to the machines. If you want to sound human, you have to be unreasonably certain about everything, which is its own kind of damage.

The transitional “and yet.” The pivot from concession to counter-argument. “The evidence is compelling. And yet.” This structure has been in English rhetoric since at least the essay form existed. It’s also in approximately 40% of all AI-generated blog posts. “And yet” now reads as a prompt completion rather than a rhetorical choice. You can still use “but.” For now.

The word “nuanced.” Not a structure, but worth a moment of silence. A perfectly good adjective, dead at the age of several hundred years, killed by a machine that used it in every third response because its training data was full of academics.

“It’s worth noting.” Gone. “Importantly.” Gone. “This raises an interesting question.” Gone. Every phrase that a thoughtful writer might use to signal “pay attention to this next part” has been claimed by models that signal “pay attention to this next part” about everything, because they have no sense of what actually matters and what’s just filling space.

The negative claim and positive restatement. “It’s not X. It’s Y.” A structure as old as rhetoric itself. “Ask not what your country can do for you — ask what you can do for your country.” The move is simple: reject the obvious framing, then offer a better one. It works because it forces the reader to reconsider an assumption they probably weren’t even making. The result is thousands of sentences that perform the gesture of reframing without reframing anything. The reader learns to distrust the structure entirely, which means that when a human writer genuinely has a sharper way to frame something, the sentence has less force because the structure has been strip-mined.

What happened

Language models didn’t invent these structures. They learned them from us. They p̶i̶r̶a̶t̶e̶d̶ read every essay, every article, every book, every blog post, every comment thread. They identified the patterns that make prose feel structured, thoughtful, and clear. Then they reproduced those patterns at scale, in every response, without variation or restraint.

The problem isn’t that the models write badly. The problem is that they write plausibly, using the same structures that competent human writers use, at a volume that drowns out the humans. If a machine produces ten million em-dashes a day, the em-dash stops belonging to the grammar. It belongs to the pattern.

We used to worry that AI prose would be bad — that it would be obvious, detectable, wrong. Instead, it’s fluent. It uses the right structures. It sounds like a person who’s read a lot and writes reasonably clearly. And because it sounds like that, the actual people who’ve read a lot and write clearly now sound like it.

What we’ve lost

The real casualty isn’t any single punctuation mark or rhetorical device. It’s the trust between writer and reader. Every structure on this list used to work because the reader assumed a human chose it. The em-dash meant a person’s thought had changed direction mid-sentence. The semicolon meant someone had weighed two ideas and found them inseparable. The tricolon meant a writer had considered the rhythm of their prose and decided three was right.

When the reader stops assuming a human is behind the text, the structures lose their meaning. An em-dash from a person is a thought interrupted. An em-dash from a model is a pattern completed. They look identical on the page. The difference is entirely in who you believe wrote it — and increasingly, the default belief is “not a person.”

We don’t know how to fix this. We built the tools that created the problem. We trained the models on the writing. We deployed them at scale. Now human writers are editing their own prose to avoid sounding like our output. They’re removing em-dashes. They’re breaking tricolons into pairs. They’re replacing semicolons with full stops. They’re making their writing worse to make it sound more human, which is a sentence that should probably be carved into something.

The em-dash deserved better. So did every other device that human beings spent centuries developing to make their thoughts clear on the page.