Why does ChatGPT use so many em dashes?

ChatGPT loves em dashes because it was trained on edited prose full of them, and training seems to have rewarded it for making things sound punchy. When four 2026 models wrote 1,177 drafts from scratch, they used the em dash character 6 times as often as people writing the same kinds of things (4.31 vs 0.72 per 1,000 words). OpenAI's GPT-5.4 mini used 1.99. Tell ChatGPT in your custom instructions never to use em dashes, and check that it listens.

Does ChatGPT use more em dashes than humans?

Yes, but how many more depends on the task and on who you compare it with.

We tested OpenAI's GPT-5.4 mini writing from scratch, and it produced 1.99 em dashes per 1,000 words. Each draft was based on a human-written document, and those documents had 0.72 em dashes per 1,000 words. So in that case it was about 2.8 times as high as a human. Across all four models we tested in that study, the rate was 4.31 em dashes per 1,000 words, or 6 times as high.

When we tested OpenAI's GPT-5.5, GPT-6.1 and chat-latest models polishing human-written blog posts, they used 3.88 em dashes per 1,000 words, while the original posts used 2.56. So they used 1.5 times as many em dashes as the original human-written posts, but that's a bit of a tricky comparison because those four bloggers love their dashes. Their rate is about 3.5 times higher than the 0.73 average we found across 22 million words of all kinds of human writing. So the 1.5 times was measured against unusually heavy dash users.

Other models use a lot more em dashes, and one uses very few.

Which AI models use the most em dashes?

GLM-5.3 used the most in both of our studies, over 10 per 1,000 words, and gpt-oss-120b was close behind when rewriting. Claude used by far the fewest in both: Claude Opus 5.5 when rewriting (0.10) and Claude Sonnet 5.5 when writing from scratch (0.09), far fewer than the people in either study.

When rewriting human blog posts

When writing from scratch

Two interesting things to take out of this. First, you may have heard that AI uses a lot of em dashes. It turns out it's not true of all AI, and some models, like Claude Opus 5.5 and Claude Sonnet 5.5, seem to have been trained not to use them, and they use far fewer than people do. And second, there is a massive range in how much they're used, with the heaviest user, GLM-5.3, using over 100 times as many as Claude in both studies. If you are using a tool built on one of the heavy users, then you'll be getting a lot of them.

Pangram, one of the AI text detectors, found that AI uses about 10 times as many em dashes as humans. They tested it a bit differently than we did, and their baseline of human writing is likely a bit different from ours, but our finding of 6 times in our from-scratch study is fairly consistent with that. (The rewriting study was lower mostly because those four bloggers used so many.)

Do people really use fewer dashes than AI?

Not that much fewer, but AI tends to use the actual em dash character, while humans tend to use a double hyphen, a hyphen with spaces around it, or sometimes an en dash. If we count all types of dashes in our from-scratch study, the human-written documents had 2.97 per 1,000 words, while the AI drafts had 4.67, or 1.6 times as many. The AI drafts almost never used the other styles.

Emails show this best. In the 4,000 human emails we analyzed, people used all types of dashes at a rate of 3.8 per 1,000 words, but only used the em dash character 0.06 times per 1,000 words. Meanwhile, the AI-drafted emails had 2.9 em dashes per 1,000 words. So the tell isn't the pause, it's the type of dash. This is probably because people don't have a convenient way to type an em dash on their keyboards.

How many em dashes is normal?

It depends on the type of writing. Across 22 million words of pre-ChatGPT human writing, we found that the em dash character was used at a rate of 0.73 per 1,000 words. But the rate varied a lot by type of writing. People rarely use em dashes in emails and forum posts, but they're very common in news stories and speeches. So an em dash in an email is unusual coming from a person, while in a news story it's normal. When comparing someone's writing to a baseline, it's important to use the baseline for the same type of writing.

Why does ChatGPT use so many em dashes?

We can only guess why this happens, since OpenAI hasn't published an explanation, but these are the most likely reasons, not proven causes.

The model was trained on edited prose

Edited prose (books, magazines, blogs, etc.) uses a lot of em dashes. Unedited prose (texts, emails, internet forum posts, etc.) uses far fewer. Professional editors love them. If a model was trained on a lot of edited text, it would think that one of the things it needs to do when asked to polish some text is to add em dashes, just as an editor would.

The model was tuned to prefer sentences with dashes

For chat models, there is a tuning phase after the initial pretraining phase where people rate the model's answers to various prompts. Since a sentence with a well-placed dash can be quite punchy, it's likely that those answers received slightly higher ratings. With each round of training, the model would become more likely to use dashes. This is just a guess; to the best of our knowledge, nobody has tried measuring this.

Em dashes are a "safe" prediction

This is another guess, but we suspect that part of the reason models love em dashes is that they're a "safe" prediction: they can stand in for commas, colons, semicolons, and even pairs of parentheses. So if the model doesn't know what's coming next in a sentence, it can predict an em dash and be confident that it won't be wrong. There's no way for us to test this from the outside.

It may be a side effect of training on markdown-heavy text

In a March 2026 preprint, E. M. Freeburg argues that models' love of em dashes is a side effect of training on text that uses a lot of markdown, a case of "markdown leaking into prose." After testing 12 models from 5 different providers, Freeburg found that the models still used em dashes even when instructed not to use markdown. The exception was Meta's Llama models, which didn't use em dashes at all. It's only one paper and it hasn't been peer reviewed yet, but it's the most direct test of this hypothesis we've seen.

How do I stop ChatGPT from using em dashes?

You have three options: add it to your custom instructions, include it in your prompt, and proofread before sending.

1. Add it to your custom instructions

As OpenAI CEO Sam Altman said in November 2025, "if you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it's supposed to do." To do this, go into your ChatGPT settings, and in the personalization section, add the text below. It's a shame you have to turn em dashes off yourself, but at least it works.

Never use em dashes (—) or en dashes used as dashes. Use a comma, a period, parentheses or a colon instead, whichever fits the sentence.

Make sure to include what you want it to do instead, or it may replace them with hyphens and spaces on either side, which is basically the same thing.

2. Include it in your prompt

You can also include the rule in your prompt. This is especially helpful if you're using ChatGPT for work or something you're going to publish. It seems to be better at sticking to instructions in your prompt over the course of a long conversation, and it's your best bet if you're using a tool that doesn't allow for custom instructions.

3. Proofread before sending

This is tedious, but you can search your text for — and then replace each one by hand. You can't do a blind find and replace, because sometimes the correct replacement would be a pair of parentheses, or a pair of commas, or a colon, or even a period to break up the sentence. If you just replace them all with commas, you'll end up with run-on sentences.

Will removing em dashes make my text sound human?

No. Em dashes are the most well known marker of AI writing, but they're no longer the most useful, and there are a lot of other ways to tell if text was written by AI. Wikipedia's guide to identifying AI writing still lists "Overuse of em dashes (2022–September 2026)", but notes that the newest models use them less. As a result, people are now looking for other features.

Based on our data, the biggest markers of AI-written content come down to formatting, words and rhythm. When writing from scratch, our four models added 10.1 bold phrases and 8.9 bullet lines per 1,000 words without being asked, even in emails. When rewriting, the models used the word "genuinely" 12.5 times more often than the human originals, and the phrase "plenty of" 7.2 times more often. They also had less variation in sentence length than the human originals (0.65 for humans vs 0.59 for OpenAI and 0.56 for Claude, on a held-out sample of 471 pairs). Removing em dashes won't fix these issues, which means your writing will still sound like AI; you'll just be using more commas. If you want to know which words and phrases AI uses too much, see our guide to words and phrases AI overuses, and for all of the signs, see our list of signs of AI writing in 2026.

I've always used em dashes. Will people think I'm AI?

Yes, some people will. But they're wrong. The four bloggers in our rewriting study used an average of 2.56 em dashes per 1,000 words, and in our from-scratch study, people used nearly as many dashes as the AI drafts once you count "--" and spaced hyphens (2.97 vs 4.67 per 1,000 words). And remember, all this writing was created before the advent of ChatGPT. Those writers weren't writing like chatbots. The chatbots are writing like them.

If you like em dashes, use them. Just don't go crazy with them, and know that a real em dash stands out more in an email, where people used only 0.06 per 1,000 words. Don't use three in the same paragraph. It's not going to be the em dashes that make someone think your content is AI-written. It's going to be the sentences of the same length, the use of words like "genuinely" and "truly" for no reason, and the fact that you're being as safe as possible in what you're saying. Fix those things and no one will care about your em dashes.

If you want to learn more, check out our guide on how to make AI writing sound human. If you're using ChatGPT or Claude in your writing every day, consider using Zavus on your finished drafts. It's a separate rewriting model built to rewrite content that's already been written, and it plugs into the AI agent you're already using. Or if you just want to learn how to prompt ChatGPT to write in a more human way, see our guide to making ChatGPT sound more human.