With the Em Dash, A.I. Embraces a Fading Tradition

There are countless signals you might look for to determine whether a piece of writing was generated by A.I., but earlier this year the world seemed to fixate on one in particular: the em dash. ChatGPT was using it constantly — like so, and even if you begged it not to.

As this observation traveled the internet, a weird consensus congealed: that humans do not use dashes. Posters on tech forums called them a “GPT-ism,” a robotic artifact that “does not match modern day communication.” Someone on an OpenAI forum complained that the dashes made it harder to use ChatGPT for customer service without customers catching on. All sorts of people seemed mystifyingly confident that no flesh-and-bone human had any use for this punctuation, and that any deviant who did would henceforth be mistaken for a computer.

Those deviants were appalled, obviously. I am one; I am, even worse, a former proofreader who could speak at length and with passion about the uses of the narrower en dash. I understand very well that this dash-happy lifestyle is maybe atypical, but I had not expected to see its whole existence questioned. The dash is a time-honored and exceedingly normal tool for constructing sentences! Dickens, Dickinson, Nietzsche, Stephen King novels, this magazine — all strewn with dashes. Part of what makes them popular, in fact, is that they can feel more casually human, more like natural speech, than colons, semicolons and parentheses. Humans do not think or speak in sentences; we think and speak in thoughts, which interrupt and introduce and complicate one another in a neat little dance that creates larger, more complex ideas. (Or, sometimes, doesn’t: The copious dashing in J.D. Salinger dialogue is a great illustration of all the thoughts we leave unfinished.) This is the whole thing punctuation is for.

The best A.I. signal the dash offers isn’t about punctuation; it’s about orthography. ChatGPT sets its dashes in the traditional style of a printed book — a stroke the width of the letter M, with no surrounding spaces. The average computer user does not type like this. The average user may not know the keystrokes that produce this character. (Or its name; some discussions called it a “ChatGPT hyphen.”) The average user just pops in a hyphen (-) or two (--), which some software corrects to that underloved en dash (–). More important, the average user puts spaces around their dashes, as most online publications do — it helps text wrap more neatly between lines.

But the arguments kept revolving around the dash itself. People talked about it as if it were some uncanny eldritch rune that no self-respecting human would even think to deploy. “Nobody uses the em dash in their emails or text messages,” one commenter insisted. “This punctuation is irrelevant to everyday use-cases.”

Oceans of communication that used to be handled by speech are now left to lone individuals typing into the internet.

I am not writing this to defend dashes. I am writing this because I want to suggest that the phrase “everyday use-cases” signals a genuinely epochal shift in our perception of what writing even is.

Consider that, for a good stretch of recent history, most of the written material that people spent time with — the stuff beyond signs and menus — was full-on writing-writing: text that somebody sat down and composed, maybe revised or edited, maybe even had professionally printed. And this kind of communication was different from our daily interaction with our peers: You talked to your peers, mostly. Even after the internet arrived, this basic psychic arrangement persisted.

And now it does not — like, at all. “Emails or text messages,” posts and chats, DMs and comments, DoorDashers telling you the restaurant is out of coleslaw: Oceans of communication that used to be handled by speech are now left to lone individuals typing into the internet. Even if you remain a dedicated reader, you may still end up spending more of your time dealing in on-the-fly typings, because that has become the everyday use-case of writing.

This everyday language is still marvelous stuff — so playfully expressive that it’s even developed an equivalent of the dopey voice we use to mock bad ideas. (It’s tYpInG LiKe tHiS.) But writing-writing is a different thing, isn’t it? At its best, it captures a different register of ideas: less visceral and immediate, maybe, but often more distilled and deliberate, more elegantly engineered, choreographing the dance of thought with more precision and depth and, usually, punctuation.

Large language models are trained on whole mountains of human-generated prose, including far more old printed matter than you or I will ever absorb. We humans ask them to mimic our writing, but we do not always specify — may not even realize — that what we mean by “writing” now includes the practically oral communication we lob through our screens all day. Then we scan the results, find telltale traces of books and magazines, and begin to fixate on those artifacts as faintly robotic. The machines are vacuously reflecting our own traditions back at us. What we may not realize yet is that we are sliding toward new ones.

Illustration by Outlanders Design