The Em Dash Problem: Cleaning AI Text So It Reads Like You
Why the em dash became the giveaway of AI writing, what else the clipboard smuggles in, and how to strip all of it in one pass without breaking your text.
No keyboard has an em dash key. That single fact explains most of what happened.
On Windows you hold Alt and type 0151 on the numeric keypad. On a Mac it is Shift+Option+hyphen. Hardly anyone bothers. People type a hyphen, or two hyphens, and let Word’s autocorrect decide. A language model has no such friction: U+2014 is one more token, and it spends them freely. Two or three per paragraph, sometimes twice in a sentence.
That asymmetry is the whole tell. Not vocabulary, not “delve”, just punctuation nobody can type.
What the dash is standing in for
Read an AI paragraph and swap each dash for what it replaced. Nine times out of ten it is a comma. Sometimes a colon. Occasionally a full stop that the model did not want to commit to, because an em dash lets a sentence keep going without choosing a relationship between the halves. That vagueness is exactly why it shows up so often, and why removing it usually makes the writing tighter rather than flatter.
The AI Text Cleaner does the swap, and shows you how many it found.
The part you cannot see
Punctuation is the visible half. The clipboard brings the rest.
Chat interfaces render their answers as HTML, so copying text out drags along whatever the layout used:
- Non-breaking space, U+00A0, which looks identical to a space and is not one
- Narrow no-break space and thin space, used around numbers and units
- Soft hyphen, U+00AD, invisible until a line happens to break there
- Zero-width space, U+200B, which has no width at all
- The single-character ellipsis, U+2026, instead of three periods
None of these show up when you proofread. All of them break things later. Excel’s TRIM() only strips ordinary ASCII spaces, so a copied name with U+00A0 in the middle will refuse to match in VLOOKUP while both cells look identical on screen. Slugs come out with %C2%A0 in the URL. And a few AI detectors treat a dense cluster of these characters as a signal on its own, which is a silly heuristic but one you may as well not trip.
Where find-and-replace goes wrong
Replacing every — with , in Word takes four seconds and breaks two things.
Dashes that open a line are not stylistic. In Russian, French, and Spanish typography, that is how dialogue is punctuated, and a comma there is simply an error. Dashes between digits are ranges: turn 1990—2000 into 1990, 2000 and you have changed the meaning of the sentence, not its style.
The tool leaves both alone. Line-initial dashes stay; number ranges collapse to a plain hyphen. Everything else follows the mode you pick, and you can compare a comma against a spaced hyphen before committing.
It will not fool a detector
Worth saying plainly, because plenty of pages imply otherwise. Detectors score word choice, sentence rhythm, and how predictable each next token is. Punctuation is not what they measure. Cleaning the typography removes the thing a human reader notices in two seconds; it does nothing to the sentences underneath, which still need you to cut the hedging and say something specific.
Think of it as a typography pass, the way you would run a spell check. Then clean a draft and read what is left out loud.