You've seen it. That long, elegant horizontal line cutting through a sentence like a high-end chef's knife. It's the em dash. If you've spent more than five minutes reading a blog post or an email drafted by an AI in 2026, you've probably seen hundreds of them. They’re everywhere. Large language models (LLMs) love them more than a college freshman loves a thesaurus.
The em dash—technically the width of the letter "M"—is a beautiful piece of punctuation. It creates a pause. It adds emphasis. It can replace commas, parentheses, or colons. But in the hands of an algorithm, it’s become a digital fingerprint. It's the "tell" that gives away the ghost in the machine. While human writers use them for dramatic flair or to manage complex thoughts, AI uses them as a structural crutch. It’s a way to glue disparate ideas together without having to commit to a more difficult sentence structure.
I’ve spent thousands of hours prompting, refining, and dissecting AI outputs. The pattern isn't just a quirk; it's a window into how these models "think" about language and why your favorite chatbot is obsessed with looking sophisticated.
The Training Data Culprit
Why does an AI prefer an em dash over a simple period? Look at the source. Most LLMs were trained on a massive scrap of the internet. This includes academic journals, high-end journalism, and classic literature—places where the em dash is a sign of "elevated" writing.
In those contexts, the em dash is a tool for nuance. But the AI doesn't understand nuance. It understands statistical probability. If it sees that high-quality, authoritative text frequently uses complex punctuation, it assumes that using that punctuation makes its own output "high quality." It’s basically cargo cult writing. The AI builds the runway out of straw and em dashes, hoping the "authority" plane will land.
It also comes down to the way tokens work. In the world of LLMs, words and punctuation are broken into tokens. The em dash is a very efficient token. It allows the model to extend a sentence's "attention" without ending the thought. By using a dash instead of a period, the model keeps the context of the first half of the sentence more tightly linked to the second half in its immediate processing window. It’s a path of least resistance.
The Sophistication Trap
There's a psychological component here too. We’ve been conditioned to think that longer, more complex sentences are smarter. AI companies want their products to sound helpful and intelligent. If a chatbot speaks in short, punchy sentences, it might come across as robotic or blunt.
To avoid this, the models are tuned to be "helpful" and "verbose." The em dash is the perfect tool for verbosity. It allows for the "not only, but also" style of expansion that fills up a page. It's the punctuation equivalent of a "yes, and" in improv.
The problem is that real people don't actually talk like this. If you’re at lunch with a friend and they say, "I went to the store—the one on 5th Street—to buy some milk," it sounds fine. If they do that every three sentences, you'd think they were having a stroke or trying way too hard to sound like a 19th-century novelist. AI doesn't know when to turn it off. It lacks the social awareness to realize it’s being annoying.
Breaking the Structural Monotony
If you want to spot AI writing, look at the rhythm. Human writing has a heartbeat. It’s messy. Sometimes we use fragments. Sometimes we ramble. AI tends to produce sentences that are all roughly the same length, often joined by that signature dash.
Look at this typical AI-style sentence:
"The market is shifting rapidly—presenting new challenges for investors—while also creating opportunities for those who remain agile."
It’s grammatically perfect. It’s also incredibly boring. It’s a "sandwich" sentence. Idea A — Middle Clause — Idea B. When you see this pattern repeated three times in a single paragraph, you aren't reading a human; you're reading a probability engine.
How to Fix Your Own AI Outputs
If you're using AI to help you write—and let’s be honest, most people are—you have to be the editor. You can't let the machine drive the punctuation bus. The em dash should be a rare spice, not the main course.
- The Two-Dash Limit: Try a self-imposed rule. If a 500-word article has more than two em dashes, start cutting.
- The Period Is Your Friend: Take those dashed-off clauses and turn them into their own sentences. It creates a much more energetic pace.
- Vary the Length: Look at your paragraph. If every sentence looks like a long, thin rectangle, break one in half. Make one three words long.
- Use Parentheses: If you really need to insert an aside, use parentheses. They feel more conversational and less "Look at me, I'm a writer."
The goal is to sound like a person who has something to say, not a machine that has to fill a word count.
The Future of the Dash
As models get better, they might learn to mimic the "messiness" of human speech more effectively. We’re already seeing "temperature" settings that allow for more randomness. But for now, the em dash remains the most prominent scar of the AI era. It's a reminder that no matter how smart these systems seem, they're still just mimicking the patterns they found in our old books and websites.
If you want your writing to stand out in a world flooded with AI content, go the other way. Be direct. Be blunt. Use a period where the AI would use a dash. It’s the easiest way to prove there's a human behind the keyboard.
Start by auditing your last three sent emails. Count the dashes. If you find more than five, you're likely letting the algorithm's "voice" bleed into your own. Delete them. Replace them with periods. See how much faster the text moves. Your readers will thank you for not making them wade through a swamp of horizontal lines just to get to your point.
