CopyClean
AI text

The em dash panic: how one punctuation mark became an AI tell

The em dash got branded the ChatGPT hyphen. How a legitimate punctuation mark became an AI tell, what OpenAI changed, and what to do about your own writing.

April 14, 2026 · 3 min read · CopyClean Blog

No punctuation mark has ever had a stranger couple of years. The em dash - a perfectly legitimate piece of typography used by centuries of writers - got itself branded as the signature of a chatbot, nicknamed "the ChatGPT hyphen", and turned into a public accusation: post one on the internet and someone will reply "AI wrote this".

Here is how that happened, what is actually true, and what you can do about your own text.

the tells

The results were clearalmost too clear Ship it, they said.

normalized

The results were clear - almost too clear... "Ship it," they said.

How a dash became a tell

Large language models trained on edited, published prose - books, journalism, essays - and edited prose loves the em dash. So models reproduce it, generously. By early 2025 readers had noticed the pattern; by April it was mainstream news coverage; through the year the discourse escalated from observation to allergy. The complaint was never just that models used em dashes, but that they would not stop: users reported that even explicit "do not use em dashes" instructions were ignored, which is what made the mark feel like a fingerprint rather than a style choice.

OpenAI eventually treated it as a real product problem. In November 2025 the company announced that custom instructions telling ChatGPT to avoid em dashes finally work. The default behavior, though, still reaches for the dash - and by then the cultural damage was done.

Is the em dash actually proof of AI?

No. Researchers who study machine-generated text keep saying the quiet part: the em dash shows up in model output because humans wrote that way in the training data. Plenty of humans still write that way, unassisted, and always have. Treating one character as a verdict misfires constantly - and the people it misfires on are, ironically, the most well-read writers.

But "not proof" does not mean "not a signal people react to". Perception runs ahead of statistics. In 2026, an em dash in a cold email, a review, or a social post gets read through a filter it did not face three years ago. Fair or not, typography now carries a vibe.

The em dash rarely travels alone

The dash gets the headlines, but pasted AI text usually carries a whole typographic accent: curly quotes rather than straight ones, a true ellipsis character rather than three periods, occasionally styled Unicode letters, plus an invisible layer of non-breaking and narrow no-break spaces you cannot see at all. Any one of them is innocent; together they read as a paste from somewhere.

What to actually do

If you love the em dash, keep it. Write like yourself; the mark is innocent, and surrendering good punctuation to a moral panic is a sad trade.

If you use AI tools in your drafting and want the output to read as yours, normalize the typography: em dashes to plain dashes, curly quotes to straight, the real ellipsis to three periods, and strip the invisible passengers while you are at it. Doing that by hand for every paste is a chore; doing it with a web tool means mailing your drafts to a stranger.

CopyClean does it at the clipboard, on your Mac, automatically: copy anything, and the AI-tell typography is normalized to plain keyboard characters before you paste, with every category individually toggleable (keep the dash conversion, skip the quotes - your call). The invisible characters go too. What lands in your document is text with your name on it, in characters a keyboard would have produced.

Clean every copy, automatically CopyClean removes hidden characters, AI typography and link trackers the instant you copy. Free 7-day trial, then 12.99 USD once. macOS 14+.
Download on theMac App Store