Why does AI use so many em dashes?
The em dash became the internet's favorite AI tell for a simple reason: the gap was real and it was big. Here are the measured rates, per 1,000 words of default output, against a human baseline from essay corpora.
| Writer | Em dashes / 1,000 words |
|---|---|
| GPT-4.1 | 10.62 |
| Claude Opus 4.6 | 9.09 |
| DeepSeek V3 | 6.95 |
| Gemini 2.5 Pro | 3.53 |
| Human baseline | 3.23 |
| GPT-5.4 | 1.43 |
At its peak, ChatGPT's flagship produced em dashes at more than three times the human rate. That's not a subtle statistical signal; that's a punctuation mark waving a flag.
Why models do it
Nobody at a lab decided models should love dashes. The best evidence points at training style. A 2026 study, "The Last Fingerprint," traced the habit to the fine-tuning stage that teaches models to be helpful assistants: the same base model produced almost no em dashes before that stage and inherited the habit after it. The assistants were trained on heavily edited, markdown-flavored prose, and they learned its punctuation as a house style. The dash also genuinely suits how models write: it glues together the balanced, clause-stacking sentences they favor, one more symptom of the cadence problem rather than its cause.
The fix, and who shipped it
For years, telling ChatGPT "no em dashes" barely worked, which became its own running joke. In November 2025, Sam Altman announced the unglamorous breakthrough: custom instructions banning em dashes finally stick. The deeper fix arrived with newer models; GPT-5.4's default rate of 1.43 per 1,000 words sits comfortably below the human baseline. The same study found suppression still isn't equal everywhere: told explicitly to avoid dashes, GPT-4.1 kept producing 3.86 per 1,000 words, while Claude dropped to nearly zero on command.
The 2026 twist: one holdout
A July 2026 analysis delivered the punchline: among the major chatbots, only Claude still uses em dashes more often than professional human writers by default. The tell that once pointed at ChatGPT now points the other way. This is the general lesson of AI tells: they're generational. Each one gets measured, memed, and tuned away, and the detection value moves elsewhere, mostly into structural patterns that survive retuning.
Does an em dash mean AI wrote it?
No, and the accusation has collateral damage: plenty of humans, especially editors and fiction writers, dash liberally by choice, and the human literary corpus runs up to 10 per 1,000 words for some authors. The honest reading is about density and company. One dash in a paragraph is punctuation. Six dashes in a page that also contains "it's not just X" sentences, triplet adjectives, and a summary ending is a pattern.
If your own drafts over-dash
The working rule from our style research: at most one em dash per 300 words, and only where a comma or parenthesis can't do the job. Our free draft checker counts them for you and rewrites the excess with one click, in your browser, without your text going anywhere.
Count your dashes. Paste a draft; get every em dash, slop word, and structural tell highlighted with a fix.
Check your draft freeSources: SlopDetector em-dash measurements · "The Last Fingerprint" (2026) · TechCrunch, the em-dash fix · Wikipedia: Signs of AI writing
Part of HumanSounding, the live index of AI writing tells. Figures quoted from their sources without adjustment; the trend data behind this site updates weekly.