Why Does ChatGPT Use Em Dashes? And How to Stop It

Ask ChatGPT for a cover letter, a product description or a quick email, and there's a good chance you'll get at least one long dash back. Sometimes three in a paragraph. It's become such a familiar habit that plenty of readers now treat the mark as a sign a machine wrote the text.

So why does ChatGPT use em dashes so much, is the suspicion fair, and how do you stop it? Here's what's known, what's still a guess, and the fixes that work.

Em dash, en dash, hyphen: a 30-second refresher

Three marks look alike on screen but do different jobs:

• The hyphen is the short one on your keyboard. It joins words ("well-known", "part-time") and often stands in for ranges ("pages 10-12").
• The en dash is a little longer, roughly the width of a letter n. Formal style guides use it for ranges and links between things, like a London to Paris route or the 2020 to 2024 period.
• The em dash is the long one, about the width of a letter m. It marks a sharp break in a sentence or sets off an aside, doing work a comma, colon or pair of parentheses could also do. In HTML it's written as —.

There's no em dash key on a standard keyboard, so in quick, casual typing it's easier to reach for a comma or a full stop. That's part of why a chatbot that produces them effortlessly, over and over, stands out.

Why does ChatGPT use em dashes so much?

Nobody outside the AI labs knows for certain. What we have is good evidence that the habit grew, plus a few credible theories about why.

The habit got stronger over time

The Washington Post analyzed 328,744 messages from OpenAI's GPT-4o model in publicly shared conversations from May 2024 to July 2025. It found that fewer than 1 in 10 ChatGPT responses used an em dash a year earlier, but by summer 2025 more than half did. (The Post discloses a content partnership with OpenAI.)

Sean Goedecke made a similar observation in an October 2025 blog post: GPT-3.5 didn't lean on em dashes, GPT-4o used roughly ten times more, and GPT-4.1 used even more. He also noted that models from Anthropic and Google use them too, so it isn't just an OpenAI quirk.

Theory one: older printed books in the training data

Goedecke's best guess is training data. As labs ran short of high-quality text, they started digitizing print books, and older books use dashes more than modern writing does. He points to a punctuation study that found dash use peaked around 1860 and then declined. Train heavily on 19th century prose and you inherit its punctuation.

He's careful to call this speculation, and he tested and rejected another popular theory: that the habit came from the dialect of the human raters who score model answers during training. A corpus of Nigerian English he checked actually used fewer em dashes than standard English, not more.

Theory two: Markdown leaking into prose

A March 2026 preprint, "The Last Fingerprint" by E. M. Freeburg, offers a different angle. It tested twelve models from Anthropic, OpenAI, Meta, Google and DeepSeek and asked them to write natural prose without Markdown. Headers, bullet points and bold text mostly disappeared. Em dashes didn't. Rates ranged from zero per 1,000 words for Meta's Llama models up to 9.1 per 1,000 words for GPT-4.1 under the suppression condition.

The paper argues that the em dash is the last surviving piece of the structured, list-heavy formatting models absorb in training, and it found the tendency already present before the human feedback stage of training. It's a single-author preprint that hasn't been peer reviewed, so treat it as an interesting lead rather than a settled answer.

Our plain-English take

Whatever the origin, the em dash can replace a comma, a colon, a semicolon or a pair of parentheses, so it's almost never grammatically wrong. For a system picking the next likely piece of text, a mark that fits nearly anywhere is an easy choice. That's our interpretation, not a proven mechanism, but it matches what you see: dashes joining clauses a person would have split into two sentences.

Why readers now see the em dash as an AI tell

By spring 2025 the mark had picked up a nickname. Rolling Stone documented people calling it the "ChatGPT hyphen", including a LinkedIn post warning about "weird long dashes" and a Reddit user who grew suspicious of a colleague's emails because they used so many.

That suspicion landed hard on people who had used em dashes for years. The same Rolling Stone piece pointed out that writers like Emily Dickinson made the dash part of their voice long before any chatbot existed. In December 2025, Air Mail ran "In Defense of the Em Dash", gathering writers and public figures including Michael Lewis and Sofia Coppola to stand up for it.

Both sides have a point. A page dense with long dashes, especially paired with other habits like "it's not just X, it's Y" phrasing, does read like default chatbot output. But one punctuation mark proves nothing about who wrote something. Plenty of editors and novelists use em dashes on purpose, and judging someone's honesty by their punctuation is unfair. Sentence rhythm and word choice matter far more, as we explain in why AI writing sounds robotic.

How to stop ChatGPT using em dashes

1. Put it in your custom instructions

For a long time this didn't work reliably. Users asked ChatGPT to stop, and it kept going. That changed in November 2025, when Sam Altman posted on X that ChatGPT would finally follow a custom instruction not to use em dashes. Engadget covered the change, and OpenAI confirmed on Threads that ChatGPT is now better at avoiding them when you explicitly ask in custom instructions. As Heise noted, that doesn't remove them by default. You have to ask.

Custom instructions live in ChatGPT's personalization settings (OpenAI's own guide to customizing ChatGPT walks through them, along with project instructions for a specific body of work). Try something like this:

Never use em dashes or en dashes. Use commas, periods, colons or parentheses instead. A plain hyphen is fine inside compound words like "well-known" and in number ranges like 5-10.

Naming the alternatives gives the model somewhere to go. You can also paste the line into a single chat instead of applying it everywhere.

Don't assume it's perfect, though. The Freeburg preprint found that even an explicit ban didn't fully eliminate em dashes in some models. Skim the output before you send it.

2. Replace them by hand

If you've already got a draft, swapping dashes yourself takes a minute and usually improves the sentence. The trick is to ask what job the dash is doing. In the examples below, [dash] marks where the AI draft used a long dash.

A quick rule: read the sentence out loud. If you'd pause briefly, use a comma. If you'd stop, use a period. If the second half explains the first, a colon usually fits. And if the dash is propping up a slogan, rewrite the sentence so it says something concrete.

3. Let our free cleaner do the first pass

For longer pieces, paste the text into the free AI text cleaner. It runs entirely in your browser, so nothing you paste is uploaded or stored. Long dashes between words become a comma, or a period when the next word starts with a capital letter, and number ranges keep a plain hyphen. It also straightens curly quotes, strips invisible characters and leftover Markdown, and tidies spacing, without changing a single word.

An automatic swap can't always tell when a colon or parentheses would read better, so give the result a quick read. And if the problem is the wording rather than the punctuation, our humanizer rewrites AI-assisted drafts so they read naturally and in your own voice while keeping every fact intact. If you're a student or writing for work, check your school's or employer's AI policy first and use these tools within it.

Should you stop using em dashes yourself?

Not if they're genuinely yours. An em dash used on purpose, once or twice in a piece, is a perfectly good tool. What gives AI drafts away is volume and sameness: a dash in every paragraph, always doing the same job. When you edit an AI-assisted draft, cut most of them, keep the one that earns its place, and spend the saved effort on rhythm, specifics and your own opinions.

Frequently asked questions

Why does ChatGPT use em dashes?

Nobody outside OpenAI knows for sure. Leading theories point to training data, especially older digitized books that use more dashes, and to formatting habits models pick up from Markdown-heavy text. The mark is also flexible enough to fit almost anywhere, which makes it an easy default.

Does an em dash mean something was written by AI?

No. Writers used em dashes for centuries before chatbots, and many still do. Heavy, repetitive use can be one hint among several, but on its own it proves nothing.

How do I stop ChatGPT using em dashes?

Add an instruction to your custom instructions telling it not to use em or en dashes and what to use instead. Since November 2025, ChatGPT follows that more reliably. Still check the output, and use the AI text cleaner to catch any that slip through.

Do Claude and Gemini use em dashes too?

Yes. Analyses of models from Anthropic and Google found they use them as well. In the Freeburg preprint, Meta's Llama models were the exception, producing none.

Will removing em dashes stop AI detectors from flagging my text?

Don't count on it. Detectors look at many statistical patterns, not one punctuation mark, and no edit guarantees a particular result. Removing dashes makes text read more naturally, which is a better reason to do it.