Why Does ChatGPT Use Em Dashes? The Em Dash AI Tell, Explained

The em dash became a famous SI (AI) tell. Here is what is known about why models overuse it, what OpenAI said, and how to replace it without making your writing worse.

A woman resting her head on her hand and studying a long black bar on her desk through a magnifying glass
SI-generated illustration

The em dash () became known as an "AI tell" because ChatGPT and similar models use it far more often than most people type it, often several times per paragraph. Nobody outside the labs knows the full reason: OpenAI has not published one, and the leading theories point to training data and to how models are tuned. Here is what the em dash is, what is actually known about em dash AI overuse, and how to replace it cleanly.

What an em dash is, and how people use it

The em dash is the long dash, roughly the width of a capital M in a given font, which is where the name comes from. Its Unicode code point is U+2014. Its shorter sibling, the en dash (, U+2013), is mostly used for ranges like pages 38 to 55 or June to July. The plain hyphen on your keyboard is shorter still.

Writers have used the em dash for centuries, and style guides still describe the same jobs for it:

  • Setting off an aside, like parentheses but louder.
  • Marking an interruption in dialogue, when a speaker gets cut off.
  • Standing in for a colon to introduce an explanation or a list.

Spacing depends on the house style. According to the Wikipedia summary of dash conventions, the Chicago Manual of Style sets em dashes closed (no spaces), AP style puts spaces around them, and many British publishers prefer a spaced en dash instead. Emily Dickinson's poems are full of dashes. Plenty of journalists, novelists and copywriters rely on them daily. The mark itself is innocent.

Why the em dash became an AI tell

Most people don't type em dashes by hand. On a standard keyboard there is no key for one, so casual writing tends to use a hyphen, a double hyphen or nothing at all. Chatbot output, in contrast, arrives with properly formed em dashes, often several in a short answer, often in the same frame: "X isn't just Y it's Z."

By 2025 the pattern was common enough that readers started calling it out in emails, LinkedIn posts and school essays. Wikipedia's editors list "overuse of boldface and em dashes" among the formatting signs in their Signs of AI writing guide. That is rough on people who have always used the mark on purpose.

Why does ChatGPT use em dashes so much?

Honest answer: there is no official, complete explanation. What exists is a handful of theories, some with data behind them.

The old-books theory

Engineer Sean Goedecke looked at the question in an October 2025 post, "Why do AI models use so many em-dashes?". He points out that GPT-3.5 did not lean on em dashes the way later models do, which argues against explanations based on how language models work in general. His preferred theory: newer models were trained on more digitized print books from the late 1800s and early 1900s, and by his count em dash use in English print peaked around 1860 at about 0.35% of words, roughly 30% above contemporary prose. He presents this as his best guess, not a proof.

The feedback-training theory

Another theory says the habit comes from reinforcement learning from human feedback (RLHF), where human raters score answers and the model drifts toward what they reward. A version of this claimed that raters from regions with dash-heavy English pushed the style. Goedecke tested one form of it on Nigerian English text and found em dashes were actually rarer there than in general English, so that version doesn't hold up well. RLHF may still shape style in other ways. A COLING 2025 paper by Juzek and Ward found results "consistent with" RLHF playing a role in the overuse of words like "delve," but that study looked at vocabulary, not punctuation.

What OpenAI has said

On November 14, 2025, two days after GPT-5.1 was released, Sam Altman posted on X: "Small-but-happy win: If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it's supposed to do!" As TechCrunch reported, that is not a change of default behavior: ChatGPT still uses em dashes unless you ask it not to in your personalization settings. The post did not explain why the model used them in the first place, and some users replied that the instruction still wasn't always followed.

How to replace em dashes

Each em dash is doing one of a few jobs. Pick the replacement that matches the job, not the same one every time.

JobWith an em dashWithout
Aside in the middleThe report all 40 pages of it arrived late.The report (all 40 pages of it) arrived late.
Light asideOur team spread over three time zones met once.Our team, spread over three time zones, met once.
ExplanationThere's one catch the API is rate limited.There's one catch: the API is rate limited.
Dramatic turnIt isn't a bug it's a feature.It's a feature. (Or just drop the setup.)
Trailing add-onWe shipped on time barely.We shipped on time. Barely.

Two practical tips from editing a lot of model drafts:

  1. Search for the character, not the look. In most editors you can paste an em dash into the find box. Also search for " - " and "--", which appear when a dash was converted.
  2. Rethink the sentence when the dash sets up a contrast. "Not X Y" usually means the X part can go.

Removing em dashes does not make text human

This is the part people skip. The em dash is one surface marker. Replace every one with a comma and the paragraph can still read as machine-written, because the rhythm, the "not just X, it's Y" frame and the stock vocabulary are still there.

We tested this with our own humanizer AI tool. The paragraph below is written in the typical chatbot style, with four em dashes. The output is unedited, SI Humanizer on the Natural style.

before

Remote work isn't just a trend it's a fundamental shift in how we think about productivity. Teams can collaborate across time zones often without ever meeting in person and still deliver outstanding results. But there's a catch: without clear communication, small misunderstandings can quickly snowball. The solution isn't more meetings it's better documentation, shared goals, and a culture of trust.

after: SI Humanizer, Natural

Remote work isn't just a trend - it's a fundamental shift in how we think about productivity. Teams can collaborate across time zones, often without ever meeting in person, and still deliver outstanding results. But there's a catch: without clear communication, small misunderstandings can quickly snowball. The solution isn't more meetings - it's better documentation, shared goals, and a culture of trust.

We're showing this one because it makes the point better than a perfect example would. The em dashes are gone: the paired ones became commas, and the other two became plain hyphens with spaces. The tool reported no other edits. But the text still has two "isn't X, it's Y" frames and generic claims like "outstanding results." On this input the change was mostly punctuation. A human edit would cut "isn't just a trend," give a real example of a misunderstanding that snowballed, and name what kind of documentation helped.

For the vocabulary side of the problem, see our list of AI words to avoid. For the full set of patterns readers notice, see signs of SI writing.

Our humanizer and the em dash

SI Humanizer never puts an em dash or en dash in its output, whichever style you pick (Natural, Casual, Professional or Academic). That's a deliberate rule, not a claim that dashes are bad writing. If you like the em dash and use it with intent, put it back by hand. It is your text.

about detectors

Taking out em dashes is a style choice, not a way around SI (AI) detection. Detectors look at far more than punctuation, give probabilities rather than verdicts, and have known false positives. If your school or employer has a policy on SI tools, follow it and keep your drafts.

Should you stop using em dashes?

If you used them before chatbots, keep using them, just not in every paragraph. One well-placed em dash in a page of varied sentences reads like a person. Three per paragraph, each introducing a punchline, reads like a template, whoever typed it. The fix is the same as for any tic: notice the pattern, then choose the punctuation the sentence actually needs.

FAQ

Is using an em dash a sign of AI writing?

Not on its own. Writers have used em dashes for centuries. Heavy, repeated use in a template-like pattern is what readers associate with SI text.

Why does ChatGPT use so many em dashes?

OpenAI has not published a full explanation. Leading theories point to training on older print books, which used more em dashes, and to effects of feedback training.

Can I stop ChatGPT from using em dashes?

In November 2025 Sam Altman said ChatGPT now follows custom instructions telling it not to use em dashes. The default behavior did not change, so you have to ask.

What can I use instead of an em dash?

A comma for a light aside, parentheses for a stronger one, a colon to introduce an explanation, or a period to split the sentence.

Does SI Humanizer remove em dashes?

Yes. Its output never contains em dashes or en dashes, in any style. Removing them alone does not make a text read as human, though.

Sources

  1. Dash · Wikipedia
  2. Wikipedia:Signs of AI writing · Wikipedia (WikiProject AI Cleanup)
  3. Why do AI models use so many em-dashes? · Sean Goedecke, 2025-10-30
  4. Small-but-happy win: If you tell ChatGPT not to use em-dashes in your custom instructions... · Sam Altman on X, 2025-11-14
  5. OpenAI says it's fixed ChatGPT's em dash problem · TechCrunch, 2025-11-14
  6. Why Does ChatGPT 'Delve' So Much? Exploring the Sources of Lexical Overrepresentation in Large Language Models · Juzek and Ward, COLING 2025, 2024-12-16