↓ Skip to main content
Ben Piper

I'm the Reason AI Loves Em Dashes

·949 words·5 mins ✨ AI-Assisted
Career 2026 AI LLMs Writing Claude Anthropic
FTC Disclosure: As an Amazon Associate, I earn from qualifying purchases. Some links on this site are affiliate links.
Ben Piper
Author
Ben Piper
Wiley bestselling author — 100k+ copies, AWS Solutions Architect Associate (SAA) & Cloud Practitioner (CLF) bestsellers, 7+ books. 45 Pluralsight courses (4.7-star, 3,003 ratings). 10+ yrs 100% remote, solo CCNP ENCOR.

Every “how to spot AI writing” listicle on the internet has the same complaint: too many em dashes. Writers treat it like a smoking gun—proof that a human didn’t touch the text. What none of them ask is where that habit came from. Models don’t invent punctuation preferences out of thin air. They learn them from something. And it turns out at least part of that something is me.

I’ve been writing with em dashes since long before anyone was training large language models on anything. My Cisco networking book from 2015 is full of them. So are the AWS study guides Sybex published in 2019. This isn’t a new tic I picked up to sound clever. It’s just how I write, and it’s been in print for over a decade.

The training data connection
#

Anthropic has faced public lawsuits and disclosures over the use of published books in training data for its Claude models. Some of those disclosures have included titles from technical and educational publishers—the same kind of publisher that put out my books. I’m not going to pretend I know the exact weighting my sentences got in a multi-billion-parameter model. Nobody outside Anthropic actually knows that. But the mechanism is real, documented, and not remotely speculative: books get scraped, books get trained on, and the stylistic fingerprints of the humans who wrote them get folded into the statistical soup that becomes the model’s “voice.”

So when millions of people started noticing that Claude (and plenty of other models) leans hard on the em dash, I read the complaints with a certain quiet pride. Somewhere in a dataset, a sentence I wrote about VLANs or IAM policies got vectorized right alongside a thousand novelists and a few million Reddit posts. The em dash didn’t come from nowhere. It came from somewhere. And there’s a real chance “somewhere” includes me.

Why the em dash gets blamed for being an AI tell
#

Here’s what bugs me about the “AI overuses em dashes” complaint: it treats the em dash itself as the problem, when the actual problem is repetition without variation. A model that drops an em dash into every third sentence isn’t wrong to use the punctuation mark—it’s wrong to use it as a crutch instead of a choice.

The em dash is one of the most useful marks in the English language. It does something a comma can’t: it interrupts a sentence with force, sets off a phrase that needs emphasis, and lets you insert a thought without derailing the grammar around it. A comma is polite. An em dash is decisive. When you want to slow a reader down and make them notice something, the em dash does that better than almost any alternative.

The failure isn’t the mark. It’s the monotony. If every single sentence in a paragraph uses the same structural trick, the writing starts to sound like it’s following a template rather than making decisions. That’s true whether a human does it or a model does it. The fix isn’t banning em dashes. It’s using them the way any punctuation should be used: because the sentence needs it, not because the last ten sentences did too.

Style is data, whether you meant it to be or not
#

There’s a broader point buried in this that’s worth taking seriously, even if the framing is a joke. Individual writing style doesn’t stay individual once it’s published. If your book sells enough copies to end up in a training corpus, your sentence rhythms, your punctuation habits, your favorite transitional phrases all become raw material for a statistical model that will go on to write emails, essays, and code comments for people who have never heard of you and never will.

Nobody signed up for this when book contracts got negotiated. It’s not something authors were warned about, and it’s part of why the lawsuits exist in the first place. But it does mean something practical for anyone who writes for a living now: your voice on the page isn’t just yours anymore. It’s an input. Somewhere downstream, a model absorbed a fraction of a percent of your habits and is now reproducing them at a scale you’ll never personally reach.

That’s a strange thing to sit with. It’s also, honestly, a little bit flattering.

So what do you do with this?
#

Nothing, really. This isn’t a call to action or a warning about protecting your intellectual property, though that’s a real conversation happening elsewhere. It’s just a reminder that writing style isn’t as personal or as contained as people assume.

If you’ve noticed AI-generated text leaning on em dashes and found it annoying, you now have a slightly more interesting theory than “the model is broken.” Some fraction of that habit traces back to real writers who used the mark deliberately, correctly, and well before it became a punchline. I happen to be one of them, and I’m not giving up the em dash just because a chatbot picked up the habit too.

If you found this post useful, you might also be interested in:

Featured image by Milad Fakurian on Unsplash

Recommended Reading#