Why does AI use em dashes so often, and is one dash proof of AI writing?
AI models use em dashes because they learned English from text that's full of them, and tuning toward a polished chat style made the habit stronger. A single dash proves nothing. Human writers use the mark at very different rates. So the em dash AI reputation depends on how densely and how predictably the dashes appear across a whole text, and one mark can't carry that verdict.
Jevgenia PogadajevaCFO, Snoika · 4 min read
Where the habit comes from
Nobody outside the AI labs has confirmed the cause. The strongest explanations point to the training data and to the way models are fine-tuned. Sean Goedecke, a software engineer who has written at length about the question, notes that GPT-3.5 didn't overuse the mark and later models did. His theory is that labs began digitizing old print books for high-quality training text. He writes that these models "rely on late-1800s and early-1900s print books," which use about 30% more em-dashes than contemporary English prose.
A 2026 arXiv paper, "The Last Fingerprint," gives a second explanation. Its authors argue that heavy training on markdown taught models to join ideas with structural marks. When formatting gets stripped out, the em dash survives because the model doesn't treat it as formatting. In that study, dash use ranges from 0.0 per 1,000 words in Meta's Llama models to 9.1 in GPT-4.1. That spread suggests each lab's fine-tuning leaves its own mark. Both theories can be true at once, which is why ChatGPT's dash habit is hard to pin on one cause.
How often models use it
Models use em dashes more than most human writers, but the gap depends on the model and the genre. ProofreaderPro, an academic editing tool, compared 200 published papers with AI drafts on similar topics. It counted 3.8 em dashes per 1,000 words in the AI drafts, against 0.3 in human STEM papers and 1.1 in the humanities.
Output length makes the numbers climb. In the "Last Fingerprint" tests, GPT-4.1 reached 14.03 dashes per 1,000 words in 5,000-word essays when nothing held it back. Newer models have also moved the other way. Graphite's research counted just 0.015 em-dashes per 1,000 words in articles by Anthropic's Claude Opus 5.5, down from 2.92 for Opus 5. Any figure you read is a snapshot of one model at one moment.
Is one dash proof
No. A single em dash tells you nothing about who wrote a text, because people used the mark long before chatbots did. What matters is density and pattern, and even those only suggest an answer.
What one dash tells you
The human baseline in the arXiv study averages 3.23 per 1,000 words, but individual writers range from 0.33 to 17.12 per 1,000 words. That's a 50-fold spread. A novelist or a columnist can easily out-dash a chatbot. If you flag every dash, you'll accuse the writers who know punctuation best.
Signals that carry more weight
The pattern around the dash tells you more than the mark itself. Watch for the same structure repeated paragraph after paragraph, like "it's not X, it's Y" built around a dash. Check whether the dashes appear with no spaces around them, which is ChatGPT's default. Then look for other tells, such as stock vocabulary and lists that always run to three items. Several of these signals together give you a reasonable suspicion, while one points to a writer who likes dashes.
Why prompting rarely fixes it
Prompting reduces dashes in some models but doesn't remove them reliably, because the habit was built in during training. The "Last Fingerprint" authors tested this directly. Telling models to avoid markdown cut dash use a little. Naming the em dash cut it further, yet GPT-4.1 still kept 3.86 per 1,000 words under a direct ban, and DeepSeek V3 kept 1.57. Claude Opus 4.6 dropped to zero.
OpenAI has worked on the problem. In November 2025, OpenAI CEO Sam Altman posted on X: "If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it's supposed to do." Without that instruction, ChatGPT keeps its old habit.
If you want dash-free text, don't stop at the prompt:
- Name the em dash in your custom instructions or system prompt. A general "avoid AI style" instruction won't catch it.
- Search the finished draft for the ", " character, and for spaced hyphens used the same way.
- Rewrite each sentence that contains one. Swapping in a comma usually leaves the same AI rhythm behind.
- Reread the text for the AI writing patterns that travel with dashes, since removing the mark doesn't remove them.
How Snoika handles dashes
Snoika treats the em dash as a style signal to check at review, because prompts alone don't settle it. The platform creates and publishes content in 135+ languages, and the same team that runs the platform adds human review on top. That review stage is where the evidence says dash habits get caught: a person reading the draft notices the repeated structures that no single instruction removes.
You can apply the same logic to your own writing without any tool. Write your chosen house rule for dashes into your instructions and check every draft against it before you publish. Treat the em dash AI debate as a question of editing, and leave AI detectors out of it. Your readers judge the sentences in front of them, and those sentences are what you can fix.
FAQ
The rate varies by model and writing genre. The 2026 paper “The Last Fingerprint” reports model rates from 0 to 9.1 em dashes per 1,000 words, against a human average of 3.23. Individual human writers vary widely, so these averages aren't authorship tests.
Explicitly ban em dashes in your writing instructions, then check the finished draft. Some models still use them despite a direct ban. Search for the marks and rewrite affected sentences rather than relying on the prompt alone.
Compare em dashes per 1,000 words rather than raw totals. Divide each text's dash count by its word count, then multiply by 1,000. Compare similar genres, since academic papers and other prose can have different punctuation habits.
Rewrite the sentence to fit its meaning rather than choosing one replacement mark for every dash. Splitting it into two sentences can help. A comma may work, but automatic substitution can leave the same repetitive sentence structure intact.
No. The “em dash ai” association isn't a reason to remove punctuation that fits your house style. Review dashes for repeated sentence patterns, and rewrite where those patterns make the prose feel formulaic.
Keep reading
All articles
