Which replacement to pick
The dropdown is the only control that matters here.
Comma, the default. Most AI em dashes sit mid-sentence, marking an aside a comma handles cleanly. The rhythm survives and the tell goes.
Period. For the dash that splices two complete sentences together. Two sentences read better without it.
Hyphen. A casual swap for chat and informal writing. It reads cheap in polished prose, so use the comma there.
Nothing at all. Delete the dash when the sentence already flows without a mark. Glance at the spacing around the edit before you send it.
En dash handling is a separate toggle, on by default. Turn it off to keep en dashes in number ranges, where they are correct. Smart quotes, ellipses and invisible characters are untouched here; the AI text cleaner takes those out in the same pass as the dashes.
Three marks, three jobs
Width separates them, and the names describe the width.
Em dash ("—", U+2014): breaks and asides, as wide as a capital M. The one models overuse.
En dash ("–", U+2013): ranges and spans, 2010–2015 or pages 40–52. Keep it wherever it means "to".
Hyphen ("-", U+002D): compound words, well-known and twenty-one. Never a sentence break.
Where the habit comes from
Edited books and prestige magazines lean on the em dash, and models absorbed that style with no keyboard nudging them back toward plain punctuation. One measurement put ChatGPT near 8 to 12 em dashes per 1,000 words, against a human rate around 3. Some models still emit about 4 per 1,000 when told explicitly to avoid them.
You can cut it off at the source. Open ChatGPT's Personalization, then Custom Instructions, and add a line such as "don't use em dashes; use commas or periods". OpenAI shipped a dedicated toggle for this on November 14, 2025. Prevention only covers text you wrote yourself, though. Anything pasted from a colleague, a PDF or a different model was never under your settings.
On a Mac, Klipto's Anti-LLM transform does the replacement on ⌘V, in any app, offline.