Em dash AI: why ChatGPT uses it and the prompt to stop it

The short answer
ChatGPT uses em dashes heavily by default, and OpenAI hasn't published a reason. The 3 leading explanations are training text, punctuation flexibility and rater preference. In November 2025 OpenAI said a custom instruction against em dashes works better than before, though not reliably in every context. A rule plus a find-and-replace pass removes them.
Why does ChatGPT use so many em dashes?
ChatGPT uses so many em dashes by default, and OpenAI hasn't published a reason, so training text, punctuation flexibility and rater preference stay unproven guesses. An em dash is the long dash (Unicode U+2014) that sets off a thought in the middle of a sentence. Here's what's known and what's guessed:
| Claim | Status |
|---|---|
| ChatGPT uses em dashes as a default style | Known. OpenAI said a custom instruction can reduce them, and that the default hasn't changed. |
| Edited books and articles are full of em dashes, so training text taught the habit | Plausible, unproven |
| One mark can replace a comma, colon, bracket or full stop, so the model reaches for it | Plausible, unproven |
| Human raters liked the style, so training rewarded it | Plausible, unproven |
Read the guessed rows as hypotheses, not findings. Ask a site that states one as fact where it got the claim, since it can't have come from OpenAI.
Is the em dash proof that AI wrote a text?
No, the em dash isn't proof that AI wrote a text, and only a dash in nearly every paragraph, next to other tells, deserves a second look. Writers have used em dashes for a long time, and they're often chosen on purpose. Wikipedia's field guide for spotting AI text files em dash overuse under signs that age as models change, and it calls its signs potential indicators only.
So the proof value of a single dash is close to nil, and frequency in context is what counts. One em dash in an essay is style; a pair in every sentence is a pattern. Our guide to the signs of AI writing puts the em dash in a checklist of nine, and that's where it belongs.
Which AI models use em dashes the most?
No model maker publishes which AI models use em dashes the most, so give each model the same 500-word task and count the dashes with Ctrl+F. A rate depends on the prompt, the model version and the sample, and it doesn't stay put between releases.
One detector vendor, Pangram, publishes its own analysis on its signs of AI writing page: about 45 em dashes per 10,000 words for OpenAI models, 32 for Anthropic models and 3 for Gemini 3 Pro, against a human baseline it gives as 5 (and as 2 in its summary table). That's vendor data with no stated date, read in November 2026, so treat it as a snapshot that won't hold for the next release.
Measure your own in five minutes:
- Give each assistant the same 500-word writing task.
- Use Ctrl+F to count the em dashes in each reply.
- Divide the count by the word count and multiply by 1,000.
- Repeat with 2 more tasks, then average the rates.
Write down the model name and the date, since the numbers won't stay valid for long.
How do you stop ChatGPT from using em dashes?
To stop ChatGPT from using em dashes, add a custom instruction that bans the character and names the replacement, then run a find-and-replace pass on the output. For a first draft that's already clean, start from our write like a human prompts, which carry the same punctuation rule inside each prompt.
On 14 November 2025 Sam Altman called it a small win that ChatGPT follows a custom instruction against em dashes. OpenAI then clarified that ChatGPT is better at skipping the mark when instructed in the personalization settings, but not by default and not reliably in every context. That's the reliability limit of the rule wording, so you'll want both steps.
Custom instruction to copy
- Open ChatGPT's personalization settings, find the custom instructions box and paste: "Never use the em dash character; use commas, colons or full stops." The custom instructions generator builds a full set if you want more rules.
- Add a second line that tells ChatGPT to scan its own draft for the long dash before sending. For Application Programming Interface (API) or project use, put the same text in a system prompt from the system prompt generator.
Find and replace
Search for the character with the Unicode code U+2014, or in Microsoft Word search for the special code ^+ in Find and Replace. Fix each hit by hand, since the right mark depends on the sentence and there isn't one fix for all. The rule fails for 2 common reasons: a request made mid-chat gets buried as the conversation grows, and a long output gives the habit more chances to return. Put the rule in the settings and ask for shorter sections.
What can replace an em dash in AI text?
Four marks can replace an em dash in AI text: a comma, a colon, brackets or a full stop, picked by what the dash was doing. A comma suits a light pause, a colon introduces a list or an explanation, brackets hold an aside and a full stop splits a long sentence. In the table, the word (dash) marks each removed dash.
| Use | Before | After |
|---|---|---|
| Comma | The launch (dash) late as usual (dash) went well. | The launch, late as usual, went well. |
| Colon | One thing matters (dash) speed. | One thing matters: speed. |
| Brackets | My sister (dash) a nurse (dash) disagreed. | My sister (a nurse) disagreed. |
| Full stop | The test failed (dash) we ran it again. | The test failed. We ran it again. |
This whole article is written without a single em dash, and it doesn't read any worse for it.
Em dash AI before and after
The same request, asked two ways. Read the second one slowly. Each line does a job.
Write me a short product update (dash) keep it friendly.
Write a 120-word product update for customers. Keep it friendly and specific. Don't use the long dash character (U+2014) anywhere. Use commas, colons, brackets or full stops instead. Before you answer, scan your draft for that character and rewrite any sentence that has one.
Why it works
The rewrite names the exact character, lists the replacements and adds a self-check step, which covers the two main reasons a plain "no dashes" request gets ignored.
Em dash AI examples
Custom instruction that bans the em dash
It names the characters, gives a replacement for each use and adds a final check.
Never use the em dash character (U+2014) or the en dash character (U+2013) in any reply. For a pause, use a comma. To introduce a list or an explanation, use a colon. For an aside, use brackets. To split a long sentence, use a full stop. Before you send each reply, check it for those characters and rewrite any sentence that has one. Keep this rule for the whole conversation, including long outputs.
How to use
- Copy it with the button above.
- Fill the blanks in {{double braces}} with your own details.
- Paste it into ChatGPT, Claude or Gemini and send it.
- Then follow up: ask for one change at a time.
Before you run “Custom instruction that bans the em dash” for the first time, the lesson on how to use AI prompts step by step walks through copying, filling and sending it.
Clean up a text you already have
Limiting the edit to punctuation keeps the voice intact, and the count lets you verify the pass.
Rewrite the text below so it contains no em dashes and no en dashes. Change as few words as possible. For each removed dash, pick a comma, colon, brackets or full stop based on the sentence, and keep the meaning and the numbers the same. Then list how many dashes you removed.
Text:
{{text}}What is the difference between an em dash, an en dash and a hyphen?
The difference between an em dash, an en dash and a hyphen is length and job: the em dash breaks a sentence, the en dash marks a range and the hyphen joins words. The names come from width: the em dash is about as wide as the letter M, the en dash about as wide as the letter N and the hyphen is the short mark on your keyboard.
- Hyphen: joins parts of a word, as in well-known.
- En dash: a range, such as 2020 to 2026 or pages 10 to 15.
- Em dash: an interruption or a pause inside a sentence.
Most complaints about em dashes in AI text concern the third use; nobody's bothered by a hyphen. For a wider fix, edit the draft by hand, or browse the prompt guides. The prompt optimizer helps with the first draft from ChatGPT or Claude.
Em dash AI FAQ
Still stuck? Build it in Prompt Studio. It'll score each part as you go.
Why does ChatGPT put em dashes everywhere?
ChatGPT puts em dashes everywhere as a default style that OpenAI hasn't explained, and a custom instruction reduces it without removing it every time. The 3 common guesses, training text, punctuation flexibility and rater preference, are all unproven.
How do I turn off em dashes in ChatGPT settings?
You turn off em dashes in ChatGPT settings by writing 'Never use the em dash character; use commas, colons or full stops' in custom instructions. OpenAI said in November 2025 that this works better than before but not reliably everywhere, so scan the output too.
Does Claude use em dashes like ChatGPT?
Yes, Claude uses em dashes like ChatGPT, at a lower rate in the vendor analysis the page cites, and the same instruction reduces it. Put the rule in a project or in Claude's settings, then scan the output.
Should I stop using em dashes in my own writing?
No, you don't need to stop using em dashes in your own writing; keep the ones that do a job and cut the rest. They're correct English punctuation, and this site's house style bans them only to keep clear of the AI association.