AI tell · grade B
Should you stop using em dashes because of AI?
No. A single em dash never trips underslop's scorer, and our social-sentiment review found the single-dash version of this warning does not hold up: editors have used the mark for a century, and it shows up across genres and dialects with no tie to AI. What the scorer actually counts is density: two or more dashes landing inside the same 150-word span. That pair is what our checklist calls the chain, the pattern a reader actually notices. Keep one dash where it earns its place, let more than 150 words pass before the next one, and the question stops mattering.
What it is
The em-dash chain is a rate, not a single mark. underslop's linter tracks every em dash in a passage, then measures the words between each pair. Two dashes landing 150 words or fewer apart both count toward a chain, and the tell scores. A dash sitting more than 150 words from its nearest neighbor gets logged as a single dash and stays out of the score entirely. That split, chain counted, single dash logged only, is how the scorer under this page actually runs.
How often it shows up
Raw model output runs the mark hard, but not evenly: in our own 12-model logs (ai-tells-empirical.json, entry emp-008), the DeepSeek variants alone show em dashes roughly once every 80 to 90 words, well inside the 150-word chain window. The same entry logs a clean exception at the other end: anthropic/claude-3-haiku carries zero em dashes anywhere in its corpus, so that 80-to-90-word rate is a DeepSeek figure, not a house average. Nine of twelve models reach for the mark at all, to append a clarifying, causal, or pivoting clause mid-sentence, which includes but is not limited to pivoting straight out of an opening validation line. That is raw, unedited output. Our own eval set of 41 unedited AI drafts (research/eval/results/baseline/cells.jsonl) shows the pattern is not uniform: some drafts carry no em dash at all, others carry a full chain. The chain, not the raw count, is what separates the two.
Why readers flag it
The single-dash panic is loud, but it traces mostly to social posts warning that any dash means a machine wrote the line. Our social-sentiment review checked that claim against writers, editors, and dialect studies of dash frequency, and it did not survive: word processors have inserted the character by default for decades, professional editors leaned on it long before generative models existed, and measured dash rates in African-English corpora actually run lower than average, undercutting even the fallback theory. What a reader notices is the drumbeat, dash after dash, hinging every clause the same way and never landing on a plain full stop, not any single mark on the page.
The fix
Count before you cut. Scan any paragraph for em dashes and check the gap between each pair. Two dashes inside 150 words: convert one to a period and let the sentence end there, or turn a hinged aside into its own sentence. A single dash sitting alone: leave it. It is doing real work, and the mark itself was never the problem.
Examples
Our new onboarding flow is simple—almost deceptively so—because every step you see today was rebuilt from scratch—based on six months of user interviews—to remove the friction nobody wanted to admit was there.
Our new onboarding flow looks simple, almost deceptively so. Every step you see today was rebuilt from scratch, based on six months of user interviews, to remove friction nobody wanted to admit was there.
I've always been drawn to problems that don't have easy answers—the messy ones—the ones most people avoid—and I think that's exactly why I ended up in this field—not by accident, but by design.
I have always been drawn to problems without easy answers, the messy ones that most people avoid. I think that is why I ended up in this field, not by accident but by design.
Questions
Should I avoid em dashes because of AI?
No. underslop's own scorer never flags a single em dash, and the claim that any one dash proves AI authorship did not hold up in our social-sentiment review. What actually scores is density: two or more dashes inside the same 150-word stretch.
How do I use an em dash without looking like AI?
Use it once and let it go. If the next dash you reach for is under 150 words from the last one, that is the moment to reach for a period or a comma instead. The rhythm is what a reader clocks, not the mark.
What's a good alternative to the em dash?
Depends on the job the dash was doing. A brief aside becomes a comma or a set of parentheses. A hinge between two related claims becomes a period, two sentences instead of one stitched together. A sudden interruption is the one case an em dash still does better than anything else.
Do people actually use em dashes?
Constantly, and long before language models existed. Professional editors have relied on the mark for a century, word processors insert it automatically, and it appears across genres from journalism to fiction at rates that have nothing to do with who or what wrote the sentence.
underslop marks this tell and the rest of the consensus set in any draft you paste, using the same open lint that scores it here. The free tier caps at 500 words a run, and your text is held for 0 seconds.