War Against the Em Dash

Let’s start with an experiment. Open a book, let's say War and Peace, to the very first page and start reading:

“Well, Prince, so Genoa and Lucca are now just family estates of the Buonapartes. But I warn you, if you don’t tell me that this means war, if you still try to defend the infamies and horrors perpetrated by that Antichrist—I really believe he is Antichrist—I will have nothing more to do with you and you are no longer my friend, no longer my ‘faithful slave,’ as you call yourself! But how do you do? I see I have frightened you—sit down and tell me all the news.”

What do we see? That's right. An em dash. Does this mean War and Peace was written by ChatGPT? Some are wondering.

These days, the em dash has become a dead giveaway for generative AI. But the em dash has been a fixture in writing forever, so how did we get here? How did it become the scapegoat and bellwether for AI slop, the smoking gun that makes people roll their eyes and say, "ugh, more ChatGPT garbage?"

Back to School

In the spirit of "know thy enemy," let’s take a quick sidebar to learn what an em dash even is, and also its cousin, the en dash. Yes, they are different things.

According to our friends at Grammarly, the em dash is used to:

• Interject or abruptly change the subject. I was going to fill out our sprint plan—but did you just delete the production database?
• Offset an appositive that contains commas. The product launch—originally scheduled for March, then April, and now May—has been delayed again.
• Offset parenthetical information. The deploy—after three failed attempts and one minor outage—finally succeeded.
• Replace a colon for emphasis. There was only one thing left to do—panic.
• Connect a list and another clause in the same sentence when the list comes first. Bugs, regressions, and broken builds—none of them surprised me anymore.

And its shorter cousin, the en dash, is used to:

• Replace the word “to” in ranges and scores. Check out lines 114–138.
• Create complex compound adjectives. New York–based startup.
• Represent conflict or connection between two nouns. Client–server architecture.

The Em Dash Didn't Deserve This

The em dash is a perfectly legitimate piece of punctuation. But the damage is done. The em dash is dying and ChatGPT is holding the knife.

Once the domain of professional writers and puncutation enthusiasts (you know who you are), the em dash was commonplace in high-quality writing: books, news articles, and academic literature. That is the kind of writing that LLMs trained on, Real, intentional, writing. Writing that was proofread and professionally copy-edited.

The conflict, of course, comes when we use LLMs for different kinds of writing. Emails. Slack messages. Internal company documentation and memos. Things that didn't used to have em dashes. In those contexts, the em dash comes off as disingenuous. It feels stiff and low-effort, because no one is copy-editing a Slack message like it is a New York Times feature. It's like, girl, I know you used ChatGPT to write this feature announcement, you're not fooling anybody.

In the world of software development, where efficiency is king and anything over 1,000 words is considered unreadable, the em dash had naturally fallen out of favor even before the days of AI. It also does not help that typing one on a standard Mac keyboard is a small production in itself (it is shift + option + dash, by the way). Now that this mysterious punctuation has started to appear with an increased frequency inside tech-focused companies, people are naturally giving it the side-eye.

The word against the em dash is out, and there are multiple requests on OpenAI's forums requesting that ChatGPT stop using em dashes.

Detecting AI Content is Hard

The em dash isn't a lone gunman here. Detecting AI-generated writing is not as simple as spotting an em dash and calling it a day. AI detection tools are notoriously unreliable. The tools that do exist rely on sophisticated statistical analysis of perplexity (how predictable the next word is) and burstiness (how much variance there is in the writing style of a document), and even so, as models get more advanced, this style of detection can become even less reliable.

This is such a difficult problem space that some AI companies, including Google, are building tools like SynthID, a watermarking system that subtly guides tokenization choices to create a one-way watermark. The watermark is imperceptible to humans but can be detected by the source system.

Conclusion

The irony is not lost on the many lovers of the em dash that a punctuation mark intended to add a human touch, to better reflect how people think and speak, is now associated with AI-generated content. Maybe too many of us did not know what we had. Maybe if we had been using the em dash all along, it would not stand out so much now that LLMs are freely sprinkling them everywhere. But it may be too late for that.

In the war against low-effort AI slop, the em dash has become collateral damage.