The robots have come for your em dashes

Did you know em dashes are a sign of AI writing?

Since when?

Well, that’s what all the LinkedIn influencers say! Humans don’t use them in writing, right? They use hyphens.

Of course we use em dashes—they’re basic punctuation marks! We use em dashes as a pause for emphasis—like I did in the first sentence in this paragraph—or to set off a quasi-parenthetical phrase, like I just did in this sentence. A hyphen is used to form compound words, or to break words by syllable at the end of lines. They’re not the same. Nobody better tell the influencers about en dashes; they’ll spontaneously combust.

So where’d this idea come from?

You got me, but I’d guess it stems from the nature of AI-generated text. It always shows a certain level of polish: no spelling or grammatical mistakes, complete sentences, a clear organization and structure. LLMs get this right in their “first drafts” because the majority of text they were trained on is itself polished—and it turns out that this level of polish covers the correct use of em dashes.

Is there a better way to detect AI-generated text than looking for em dashes, then, smart guy?

All generative AI is, by literal definition, statistically median, and this isn’t something that can be overcome by creative prompt engineering. Flair and personality in human writing comes from individualistic, quirky choices born of a writer’s likes, dislikes, and life experiences. LLMs don’t have any of those. So LLM-generated text tends to read like pleasantly anodyne business memos, or book reports by a precocious ninth-grader. So if the text you’re playing “robot or not” with is a business memo, a high school paper, or a blog post by the kind of LinkedIn influencer who thinks em dashes are a sign of AI writing, it might be a challenge. In other circumstances, it shouldn’t be.

But it apparently is!

The problem—and, dear reader, there is no gentle way to put this—a lot of ostensibly smart people neither write particularly well, nor read enough to get a sense of what good writing is like. Most of the writing they deal with—memos, emails, text messages, Amazon and Yelp reviews—is slapdash, unpolished. So if they come across anything in those contexts that looks refined yet relatively generic, they suspect ChatGPT. Apparently, we’ve so come to expect a lack of writing style from humans—and I mean style in the sense of Strunk & White, not in the sense of actual style—that we’re suspicious of it when we see it.

If you don’t want your writing to be mistaken for AI, what should you do?

Use “fuck” more often.