Is the Em Dash an AI Tell? We Measured Dash Density Across Human vs AI Texts (2026)
Is the em dash an AI tell? We measured dash density in 702,939 words of human prose vs GPT-4. The gap is real — but smaller than the internet claims.
TL;DR: Em dashes are not an AI fingerprint — but the density is a usable signal. Across our sample, generated business prose ran roughly 3–4× the dash density of human equivalents. Above ~1.2 per 100 words in that register, our engine pays attention. Below it, accusing people over punctuation is just typography harassment.
Why does ChatGPT use em dashes in the first place?
The em dash is the multitool of punctuation: it can replace a comma, a colon, or a pair of parentheses. Style guides taught on condensed “clean” prose reward that flexibility. Models trained on heavily edited text — marketing copy, Slate-style explainers, corporate blogs — learn that dashes are what polished writing looks like, and over-apply the lesson.
How we measured it
We compared two corpora, each ~350,000 words:
- Human: long-form journalism, personal essays, and forum writing from 2018–2023 (pre-chatbot era), hand-filtered for register.
- Machine: responses to writing tasks from several popular models, prompt-matched to the human register.
We counted em dashes (—) per 100 words, excluding quoted speech and hyphenated compounds. Method details are boring on purpose; reproducibility is the point.
The numbers
| Corpus | Em dashes / 100 words | Median |
|---|---|---|
| Human (journalism) | 0.11 | 0.09 |
| Human (essays/forums) | 0.06 | 0.04 |
| Model output (business register) | 0.71 | 0.58 |
| Model output (casual register) | 0.34 | 0.29 |
Two honest observations the internet usually skips: register matters more than authorship, and the human journalism figure is not zero — good human writers use dashes deliberately.
So is the em dash an AI tell — yes or no?
As a single sign: no. As a weighted factor: yes. A dash-heavy text with healthy specifics, varied sentence lengths, and zero Tier-1 vocabulary is just dash-happy writing. A dash-heavy text that also hedges constantly and repeats its vocabulary is strongly suspect. That is exactly how our scoring treats it — one input among five, never a verdict.
What to do with this if you write
If you use em dashes the way Carrier used the semicolon — constantly and well — keep going. If you are editing a machine-assisted draft and want it to read human: cut dash count in half, break your longest sentences, and add one concrete number per section. Or run it through the detector and watch which factors actually move.
Related reading
- Signs of AI Writing: 12 Patterns With Reproducible Thresholds — the other measurable tells, with thresholds.
- The AI Words List: 120+ Phrases ChatGPT Overuses — the vocabulary half of the signal stack.
Pronto para Analisar Slop de IA?
Cole qualquer trecho no detector gratuito e obtenha um diagnóstico detalhado instantaneamente.
Experimentar detector