Remove em dashes from ChatGPT text (and the characters you cannot see)

Language models are trained on edited prose, and edited English leans on the em dash: for an aside, for a break in thought, for the summary clause at the end of a sentence. The model reaches for the construction that fits, which is often that one, and it writes the proper typographic character rather than two hyphens, because that is what published text does.

It reads as a tell because of the RATE, not because the character is exotic. Human writers use em dashes too. A draft with one every other sentence reads as generated; three in two thousand words reads as edited. That is a measurement, not a verdict, and you can take it yourself.

U+2014 EM DASH

An em dash is text you can see, so the default pass does not touch it. That is deliberate rather than an oversight: Standard removes only characters that are invisible and doing no work, and rewriting punctuation changes something a reader can read. So it is a switch, and you throw it.

1. Open the cleaner and paste the draft.
2. Set Cleanup level to Maximum. If you want only the punctuation, open Advanced and tick “Plain-text punctuation” instead — that single option is --punct on the command line.
3. Press Clean text. Em and en dashes become hyphens, curly quotes become straight ones, the one-character ellipsis becomes three dots and the minus sign becomes a hyphen. Your words are untouched.

Maximum also switches on the other rewriting passes — look-alike letters, spacing, normalisation — so if the dashes are all you want changed, take the single option in Advanced and leave the rest alone.

Open the cleaner

The dashes are the part you can argue about. The part worth cleaning is the part with no glyph: zero-width spaces, word joiners, byte-order marks, unusual spaces and look-alike letters, which arrive with any paste that has been through a chat window, a web editor or a CMS.

Those come out at Standard, without you choosing anything, because they are invisible and doing no work. They are also the ones with real consequences: a broken exact-match search, a word your spell checker splits in two, a diff on a line nobody edited.

How to remove AI watermarks

The prose-habits report counts em and en dashes per thousand words, beside sentence-length variation, stock transitions, hedges, contractions and the other habits that are actually measurable. It emits no score and no probability — it gives you the numbers and what to change about each one, and the editing decision stays yours.

Prose habits report