If it feels like every ChatGPT paragraph has an em dash, you are not imagining it. Here is why, and how to fix it.
The short answer
Large language models learn patterns from huge amounts of published text, books, articles, and web writing, where the em dash is a normal, even elegant, piece of punctuation. Professional writers use it to set off asides and add rhythm. The model picked up that pattern and reproduces it at a higher rate than an average person typing casually would.
So the em dash is not a bug or a hidden signal. It is a statistical echo of the polished prose the model was trained on.
Why it reads as an AI tell
The issue is density, not the dash itself. One em dash in a paragraph is normal writing. Three or four across a short message is unusual for everyday text, and readers have started to associate that density with AI. So even though each dash is grammatically fine, the frequency gives the draft away.
How to remove them
You do not have to hunt each one down manually. Paste the text into a cleaner and it swaps every em dash at once. With textscrubr you choose the replacement:
- a comma, which reads most naturally for the parenthetical asides ChatGPT writes
- a hyphen, a simple one-to-one swap
- nothing, closing the gap
It also catches the en dash, which manual find-and-replace usually misses, and keeps number ranges like 2010-2020 readable.
Should you remove every em dash?
No, the goal is to remove the unnatural density, not to ban the punctuation. In your own edited writing an occasional em dash is perfectly fine. What flags a draft as AI is a flurry of them, so converting most to commas or periods is enough to make the text read as human.
Bottom line
ChatGPT uses em dashes because its training data did. They are the most recognizable AI writing tell, and a one-click cleaner removes them faster than any manual edit.