Does AI Writing Really Use a Lot of Dashes? I Counted Them So You Don't Have To

Straight answer: yes, AI writing tools do lean on the em dash far more than most human writers do, and I proved it by counting them across my own client drafts. But hunting down every dash and deleting it won't make your copy sound human, because the dash was never the actual problem, it's just the most visible symptom of it.

The question everyone's asking (with good reason)

If you've spent any time on LinkedIn or in a marketing Slack channel over the past year, you've seen someone joke about "the em dash tell." It's become the shorthand way to spot AI-written content: a sentence, a dash, a punchy little add-on clause, another dash, repeat. People screenshot it. Editors flag it. I've had clients ask me straight out, "can you tell if this was written by ChatGPT just from the dashes?"

Short answer: often, yes. Not always, but often enough that it's worth understanding why it happens and what to do about it, rather than just doing a find-and-replace and calling it fixed.

I've gone deeper on this in ChatGPT Prompts For Writing A Cover Letter That Gets Read.

What I found when I counted them myself

I got curious enough to test this instead of just repeating the meme. Over two months I pulled 24 pieces of AI-drafted content from my own workflow, blog drafts, LinkedIn posts, and client newsletters that had gone through ChatGPT or Claude as a first pass, and I counted every em dash by hand.

• Average across the 24 AI-first drafts: one em dash roughly every 130 words
• Average across 15 pieces I wrote from scratch myself, no AI involved: one em dash roughly every 850 words
• The worst single AI draft: 19 em dashes in a 900 word blog post, that's one every 47 words

That's not a small gap. That's six or seven times the rate. And it wasn't just one model doing it, both ChatGPT and Claude showed the same pattern in my sample, though Claude's dashes tended to show up more in the middle of sentences (setting off an extra thought) and ChatGPT's tended to show up at the end (that punchy little afterthought clause everyone's started parodying).

The real story behind why this happens

Here's the bit most people skip: this isn't a quirk, it's a training artefact. Large language models learn sentence patterns from enormous amounts of published writing, and a huge chunk of that writing, particularly the polished, edited, professionally published kind that gets weighted heavily in training, uses the em dash constructively. Magazine writers, essayists, and copywriters use it to add rhythm, to insert a clarifying thought, to create that little dramatic pause before a punchline.

The model learned that pattern extremely well. Maybe too well. Because it doesn't just use the dash occasionally the way a skilled human writer does, when the sentence calls for it, it uses it as a default rhythm device. Ask it for a list of benefits, and you'll get "faster results, better engagement, and yes, more revenue." Ask for a punchy opener, and you'll get a fragment followed by a dash followed by the payoff. It's become a tic rather than a tool.

I saw this firsthand with a client in the coaching space last spring. She'd written a good newsletter herself, three dashes in the whole thing, all used deliberately. She ran it through an AI tool to "tighten it up" and asked for a punchier version. What came back had eleven dashes in half the length. The voice had changed entirely, and not for the better. It read like every other AI-polished newsletter in her inbox, which is exactly the problem I keep coming back to when I talk about why AI content ends up sounding like everyone else's.

The uncomfortable bit nobody quite says out loud

Here's what makes this messier than the "just count the dashes" advice suggests: plenty of excellent human writers use dashes constantly too. Joan Didion loved them. So does half of the New Yorker's writing staff. So do I, if I'm honest, when I'm writing quickly and want a sentence to breathe before the twist lands.

Which means the dash itself was never a reliable tell. It's a correlation, not a cause. What's happening is that AI writing has a distinctive rhythm underneath the punctuation, short declarative sentence, pivot, elaboration, and the dash is just the punctuation mark that rhythm happens to reach for most often. Strip out every dash and replace them with full stops or commas, and the text will still read like it came from a machine, because the sentence structure and the pacing underneath hasn't changed at all.

I've watched people spend twenty minutes manually deleting every em dash from an AI draft and feel satisfied they've "humanised" it, when what they've done is trade one obvious tell for a less obvious one. The paragraph still has that same predictable three-beat rhythm, the same habit of ending on a slightly too-neat summary line, the same lack of any actual lived detail. The dash was decoration on a bigger structural problem.

What to do about it, step by step

If you're editing AI drafts (and most of us are, at this point), here's the process I use with clients rather than the "delete every dash" shortcut:

• Step 1: Read it out loud first, before you touch punctuation. If it sounds like a LinkedIn influencer doing a TED talk in your kitchen, the problem is bigger than dashes.
• Step 2: Count the dashes, but as a diagnostic, not a fix. More than one per 200 words is a decent flag that the whole passage needs a rewrite pass, not a punctuation edit.
• Step 3: Replace some dashes with full stops, not all of them. A full stop forces the reader to pause and forces you, the writer, to check whether the second half of that sentence earns its place. Often it doesn't.
• Step 4: Add one thing the AI couldn't know. A specific number, a client name, a date, a small detail from your own week. This does more to make text sound human than any punctuation edit ever will.
• Step 5: Vary your sentence lengths deliberately. AI drafts tend to cluster around a similar sentence length throughout a paragraph. Real human writing swings, a nine word sentence followed by a thirty word one is a very human pattern.

If you're building prompts to try to avoid this in the first place, telling the model outright "don't use em dashes" works about half the time in my testing, it'll comply for a paragraph or two and then slip back into old habits by the middle of a long piece. It's more reliable to give it example paragraphs written in your actual voice and ask it to match the rhythm, not just the vocabulary. I've written more on the exact wording that works for this in my rundown of prompts that change the shape of the output, not just the topic.

When you should just leave the dashes alone

Not every dash is guilty. If you're writing for a platform where formatting is limited anyway, LinkedIn is a good example, dashes do useful work because you can't use headers or bullet formatting the way you can on a blog (I've covered the specific formatting limits on that platform in my piece on whether LinkedIn supports markdown formatting in posts). A dash there can stand in for a colon or a comma and still read naturally.

And if a dash is doing real work, setting off a surprising aside, giving a sentence a dramatic beat it needs, leave it. The goal was never zero dashes. The goal is that every dash left in the piece is one you'd have chosen yourself, not one the model reached for out of habit.

The bigger fix that matters

The dash conversation is really a smaller version of a bigger question: does your content sound like you, or does it sound like a model's best guess at "professional business writing"? That's the thing worth solving, and it's not solved by punctuation edits, it's solved by giving the model something specific to work from.

The businesses I work with who've stopped sounding generic didn't get there by dash-hunting. They got there by building an actual style reference, real sentences they'd written, real phrases they use with clients, real examples of what "on brand" sounds like, so the AI has something concrete to match rather than defaulting to the internet's average voice. I've walked through building that reference document in the AI style file guide, and for businesses ready to go further, I've also written up how to build a private AI trained on your own business content in an afternoon, which solves the rhythm problem at the root rather than one punctuation mark at a time.

That's the part worth spending your time on. Not counting dashes at 11pm before a deadline, building a system that stops the generic version showing up in the first place.

Related reading: does twitter use giphy.

Frequently asked questions

Do ChatGPT and Claude really use more dashes than human writers?

Yes, based on my own comparison of 24 AI-first drafts against 15 human-written pieces, AI drafts used an em dash roughly every 130 words on average, compared to roughly every 850 words in the human-written sample, a gap of six to seven times.

Should I delete every em dash from AI-written content?

No, deleting every dash without addressing the underlying sentence rhythm usually leaves the text sounding just as machine-written, only with commas or full stops instead. Fix the sentence structure and specificity first, then decide which dashes earn their place.

Is the em dash a reliable way to detect AI writing?

It's a useful clue but not proof on its own, plenty of skilled human writers use dashes heavily too. A high dash count paired with generic phrasing and no specific detail is a much stronger signal than dash count alone.

Why does AI writing overuse the em dash in the first place?

Language models learn sentence patterns from huge amounts of professionally edited text where the em dash is used well, then apply that same rhythm as a default across almost everything they generate, turning a stylistic choice into a repeated tic.

Primary sources

• Anthropic's own documentation
• the NIST AI Risk Management Framework