Is the Em Dash an AI Tell? We Measured Dash Density Across Human vs AI Texts

1Why does ChatGPT use em dashes in the first place?

Before measuring anything, it helps to know why the pattern exists. The "why does ChatGPT use em dashes" question has a boring, well-documented answer, and it isn't "the AI loves elegant prose."

Two things stack up. First, training data. According to McGill University's Office for Science and Society, more than 60% of GPT-3's training corpus came from web crawls — essays, articles, long-form blog posts. That's exactly the register where the em dash lives. Everyday texts and chat messages barely use it; published essays lean on it constantly. The model learned the distribution of the writing it was fed, and that writing was dash-heavy.

Second, the em dash is genuinely useful to a next-token predictor. It's a flexible connector. It can open an aside, tack on an afterthought, or swap in for a comma, colon, or parenthesis without breaking a sentence. A system optimizing for fluent, hedge-friendly continuations reaches for a tool that fits almost anywhere. So it reaches for that tool a lot.

There's a third factor: reinforcement. OpenAI has publicly acknowledged tuning ChatGPT's em dash habit specifically. In November 2025, Sam Altman announced that if you tell ChatGPT not to use em dashes in your custom instructions, "it finally does what it's supposed to do" — a tidy admission that, until then, the model kept producing them even when asked to stop. When a company ships a fix so the model will quit doing something, that something was a pattern.

That's the origin of the "ai em dash pattern." Now let's see how big it actually is.

2How we measured it

We wanted a number, not a hunch. Here's exactly what we did, so you can reproduce or challenge it.

The counting rule. Typography is a trap here, so we counted two ways. The strict count is the em dash character itself (—, Unicode U+2014). But older typesetting frequently rendered the same parenthetical break as a double hyphen (--) or an en dash, and a reader perceives all of them as "a dash." So we also counted dash constructions — em dash, en-dash-as-dash, or double hyphen — normalized per 1,000 words. This is the same family of signal our detector's structure module looks at when it flags em dash density.

Why this matters: it exposes a quiet bias in the whole debate. When people say "humans don't use em dashes," they're often looking at the character. Austen's Gutenberg edition scores 0.00 on the em dash character — and 3.47 per 1,000 once you count the double hyphens that are em dashes in disguise. The mark was always there. The Unicode codepoint is the new part.

AI baseline. For the machine side we lean on the most rigorous public measurement we could verify rather than pretending we ran our own model farm. That study is described below, and it uses the same per-1,000-words unit, so the numbers are directly comparable.

3The numbers

Here is what 700,000 words of human prose actually looks like, measured.

Two findings jump out.

Human prose is not a dash desert. Twain and Melville sit above 8 per 1,000. If you drew a naive "AI line" at 5 or 6 dashes per 1,000 words, you'd convict two of the most celebrated American novels of being written by GPT.

And the character-vs-construction split is real. Dickens and Austen read as 0.00 if you only count the modern em dash glyph, then jump to 5.33 and 3.47 when you count the double hyphens their typesetters used for exactly the same purpose. The dash was never absent. The typographic em dash — clean, curly, U+2014 — is what changed.

What AI actually scores

Now the machine side. The cleanest public measurement we found is E. M. Freeburg's 2026 analysis, "The Last Fingerprint: How Markdown Training Shapes LLM Prose," which ran multiple frontier models on matched prompts and counted em dashes per 1,000 words against a human control.

So the headline "3x" figure is real: GPT-4.1's 10.62 per 1,000 is about 3.3 times the study's human baseline of 3.23. That lines up with our own human-prose reading of 4.8 to 6.5 per 1,000 — our literary sample runs a touch higher than a contemporary-functional-prose control, exactly what you'd expect from 19th-century novelists.

But look at the spread. Llama produces zero. Gemini 2.5 Pro sits at 3.53 — indistinguishable from a human. "AI" is not one dash rate; it's a range from 0 to 10-plus depending on the model. Meanwhile a real, independent corpus check found the same drift outside the lab: an analysis of 10,000 ecology paper abstracts on OpenAlex found the em dash's relative frequency more than doubled between 2021 and 2025, with no other punctuation mark coming close to that jump. The mark really is surging in machine-adjacent text. It's just not a clean on/off switch.

Twain's Huckleberry Finn scores 10.13 em dashes per 1,000 words. GPT-4.1 scores 10.62. On this one metric, one of the most distinctively human voices in American literature is statistically indistinguishable from a language model.

4So is the em dash an AI tell — yes or no?

Here's the honest answer, and it's the reason we don't lead our product with punctuation.

The em dash is a weak signal, not a fingerprint. Presence tells you almost nothing — humans have used the mark for centuries, and some of the best writers use it heavily. What carries a little information is density in the wrong register: a stream of clean U+2014 em dashes in short, modern, functional prose (a LinkedIn post, a cover letter, a product FAQ), where you'd normally see a comma or a period. That's the pattern that shifted the ecology-abstract numbers and the pattern Freeburg measured.

Even then, it's easy to fool in both directions:

That last point is the whole ballgame. A dash is a keystroke, not a thought. You can delete every em dash from a slop paragraph and it's still slop: still vague, still substance-free, still saying nothing in a confident voice. Punctuation is the easiest signal to game and the least connected to whether the writing is any good.

This is why our tool judges quality, not authorship. We're not an "is it AI" detector — those chase a moving target and punish humans for their commas. We flag whether text reads like slop: low-substance, templated, cliché-dense content, whoever produced it. Em dash density is one small input to our Structure score, weighted deliberately low. To see what heavy-handed AI output looks like end to end — the dashes plus the empty confidence plus the cliché stacking — our AI slop examples page annotates real cases so you can calibrate your own eye.

5How our detector actually treats the em dash

Since this is a data piece, here's the real logic, not a marketing summary.

Inside our engine, em dash density feeds the Structure dimension — one of five signals, alongside Vocabulary, Cliché, Diversity, and Substance. The rule is deliberately conservative. It only starts adding points when em dash density climbs above 2 per 100 words (20 per 1,000) — roughly double the rate of the most dash-happy human novel in our sample. Below that, the mark contributes nothing to the score.

And in our Unicode module, the em dash is explicitly filed as a warning-level stylistic signal — "not forensic on its own" — sitting a full tier below the actual smoking guns, which are invisible characters like zero-width spaces that a human essentially never types by hand. Our own code agrees with the finding in this article: a dash is a hint, not proof. What moves the needle is the combination of signals, and above all whether the content has anything to say.

This is the difference between a black-box "87% AI" verdict and an explainable quality score. We can show you the exact reason a text scored the way it did — this many cliché phrases, this much sentence-length uniformity, this dash density — and you can argue with any single line of it.

6What to do with this if you write

A few practical takeaways, whichever side of the keyboard you're on.

Check your own text

Curious where your writing lands? You don't have to trust a punctuation rumor. Paste any text into our free, 100% local slop detector and you'll get a five-dimension breakdown — including how your structure signals, em dash density among them, actually score. Nothing leaves your browser; the analysis runs on your device. It won't tell you "a robot wrote this." It'll tell you something more useful: whether the writing is worth reading — and exactly why.