Why Does ChatGPT Use Em Dashes? And Other AI Writing Tells
ChatGPT's em dashes, curly quotes and 'not X but Y' lines are easy tells, but deleting them won't fool AI detectors. See why models do it and how to clean text.

Key takeaways
- OpenAI hasn't publicly explained ChatGPT's em dash habit. Likely causes include edited prose in training data and human feedback during tuning.
- In November 2025, Sam Altman said ChatGPT now follows custom instructions not to use em dashes, though results vary.
- Em dashes travel with other tells: 'not X but Y' lines, lists of three, bold lead-ins, stock vocabulary and leftover Markdown.
- Removing em dashes won't fool detectors on its own, because they read word choice, rhythm and structure.
- Clean the surface first, then fix the patterns underneath by revising rhythm, constructions and specifics.
In this guide
- Why does ChatGPT use so many em dashes?
- What other tells do AI models leave in text?
- Do Claude and Gemini have the same habits?
- Does removing em dashes fool AI detectors?
- How do you clean up ChatGPT text properly?
- What does a cleaned-up paragraph look like?
- Can you stop ChatGPT from using em dashes in the first place?
- The bottom line
ChatGPT uses so many em dashes because of how it learned to write, not because of any rule it was given. OpenAI hasn't published an explanation we could find, but likely causes include edited prose in training data and human feedback that rewarded a punchy, conversational style. Removing em dashes alone won't fool AI detectors, which read deeper patterns.
Here's what probably drives the habit, the other tells that travel with it, and how to clean AI text in a way that actually changes how it reads.
Why does ChatGPT use so many em dashes?
Nobody outside OpenAI knows for certain, and the people who study this mostly offer educated guesses. The likely reasons:
- It learned from edited prose. Books, magazines and essays use em dashes freely, and models learn what polished writing looks like from text like that.
- Human feedback may have rewarded it. Chat models are tuned with feedback from people rating responses. A style that feels lively and conversational, with dashes for asides and punchlines, may have been preferred and then reinforced.
- It's the most flexible punctuation mark. An em dash can stand in for a comma, a colon, parentheses or a semicolon. For a model choosing the next token, it's a safe way to tack on an aside or a twist without restructuring the sentence.
Whatever the cause, the habit became so recognizable that OpenAI addressed it publicly. In November 2025, CEO Sam Altman posted that if you tell ChatGPT not to use em dashes in your custom instructions, "it finally does what it's supposed to do," and called it a "small-but-happy win." The default behavior didn't change, and some users reported the dashes creeping back.
One thing to keep in mind: em dashes aren't an AI invention. Plenty of human writers love them, and using one proves nothing. The tell is frequency. When every paragraph has two, readers start to notice.
There's also a typing angle. The em dash is its own character (Unicode U+2014), separate from the hyphen on your keyboard. Most people rarely type one, because it takes a special shortcut or an autocorrect rule. A model outputs it as easily as any other character, which is one reason a page full of them stands out.
It helps to know the em dash's relatives, too. The hyphen joins words, as in "well-known." The en dash, slightly longer, joins ranges in formal typesetting, such as page numbers. The em dash, the longest of the three, breaks into a sentence. It's the one chat models lean on.
What other tells do AI models leave in text?
Em dashes are the famous one, but they travel with a whole set of habits.
| Tell | What it looks like | Why it stands out |
|---|---|---|
| Em dash overuse | Several long dashes per paragraph for asides and punchlines | Most people use them sparingly |
| "Not X, but Y" | "It's not just a tool, it's a mindset." | Sets up a fake contrast to sound insightful |
| Rule of three | "fast, simple, and reliable" | Lists of exactly three, again and again |
| Bold lead-ins | A bolded label at the start of every bullet | Reads like a slide deck |
| Markdown leftovers | Stray asterisks and pound signs after pasting | Formatting carried over from the chat window |
| Stock vocabulary | "delve," "tapestry," "testament," "crucial" | Words models favor far more than people do |
| Curly quotes in plain text | Typographic quotes mixed with straight ones | Suggests text was pasted in from somewhere else |
| Hidden characters | Non-breaking or zero-width spaces | Invisible to you, visible to software |
| Summary closers | A final paragraph that restates everything | Tidy, predictable endings |
You can scan a draft for the vocabulary tells with our AI word checker, which highlights overused AI words and suggests plainer alternatives.
Do Claude and Gemini have the same habits?
Often, yes. Other chat assistants are trained and tuned in broadly similar ways, so their drafts can show the same tells: neat lists of three, tidy summary paragraphs, bold labels and stock vocabulary. Each model has its own favorite moves, but none of them writes like you unless you rewrite the draft. The cleanup steps below work whichever model you used, and our pages on humanizing Claude text and humanizing Gemini text cover those models specifically.
Does removing em dashes fool AI detectors?
No. Not on its own.
AI detectors don't work by counting punctuation. They're classifiers trained on large sets of human and AI writing, and they react to deeper patterns: how predictable the word choices are, how even the sentence rhythm is, and how the paragraphs are built. We explain those signals in perplexity and burstiness explained and how AI detectors work.
If you swap every em dash for a comma, the sentence skeleton stays exactly the same. So do the predictable word choices, the matching sentence lengths and the three-item lists.
Detector makers know about cosmetic fixes. When Turnitin announced detection for AI humanizer tools in August 2025, Plagiarism Today noted that humanizers typically edit out common AI markers such as certain phrases and em dashes, and Turnitin said it had studied leading humanizers and trained its model to recognize their patterns. Surface cleanup is exactly what detectors now expect.
So why clean em dashes at all? Because people read your text too. Readers, editors, clients and teachers notice the tells long before any detector runs.
How do you clean up ChatGPT text properly?
Work from the surface down.
1. Strip the formatting artifacts
Paste your text into our free ChatGPT text cleaner. It removes em dashes, hidden characters and leftover Markdown such as asterisks and pound signs, so you start from clean plain text.
2. Replace dashes with the right punctuation
Don't swap every dash for a comma. Pick the mark that fits the job:
| If the dash is doing this | Use this instead |
|---|---|
| Setting off an aside | Commas or parentheses |
| Introducing an explanation or list | A colon |
| Joining two complete thoughts | A period or a semicolon |
| Delivering a punchline | A new, short sentence |
3. Rewrite the constructions
Turn "It's not X, it's Y" into a direct claim. Break up lists of three, or cut them to the one item that matters. Drop bold lead-ins unless the format needs them.
4. Fix the rhythm
Paste a few paragraphs into the sentence length checker. If every sentence lands in the same range, combine some and split others.
5. Add what only you know
Specific examples, numbers, names and opinions are the hardest things for a model to fake and the easiest way to make text sound like you.
6. Humanize and check
For longer drafts, The Humanizer rewrites sentence rhythm, structure and word choice while keeping your meaning, facts, citations and paragraph count. It works on text from ChatGPT, Claude, Gemini and other models; see our page on humanizing ChatGPT text. Then run a quick first pass with our free AI detector, which estimates AI-likeness from writing patterns. Results vary by text and by detector.
What does a cleaned-up paragraph look like?
Here's a typical AI paragraph, with its dashes already removed:
Remote work isn't just a perk, it's a fundamental shift in how we work. It offers flexibility, autonomy, and balance. The takeaway: companies that ignore it risk falling behind.
Swapping the punctuation didn't help much. It still has the fake contrast, the list of three, the bold label and a generic warning. Now the same point, rewritten with specifics:
Our team went remote in 2021, mostly so people could skip a 90-minute commute. Two of our best hires live four states away. If we'd insisted on the office, we would have lost both.
The second version is shorter on slogans and longer on evidence. It also varies its rhythm naturally, which is what readers and detectors both respond to.
Can you stop ChatGPT from using em dashes in the first place?
Partly. Going by Altman's November 2025 post and OpenAI's own note, the place to ask is ChatGPT's custom instructions, in its personalization settings, rather than a single prompt. Results still vary, so check the output anyway.
You can also prompt for a plainer style overall: ask for short paragraphs, varied sentence length and no bold text or headings. Our list of ChatGPT prompts that sound human has examples. Prompts help with tone, but they don't remove the deeper patterns, which is why cleanup and revision still matter. For more hands-on techniques, see how to humanize AI text manually.
The bottom line
ChatGPT likely overuses em dashes because of how it was trained and tuned, though OpenAI hasn't explained the habit publicly. Deleting the dashes makes text look cleaner but doesn't fool detectors, which read rhythm, structure and word choice. Clean the surface, then fix the patterns underneath.
Start with the free ChatGPT text cleaner, then try The Humanizer free with 350 words a month.
Frequently asked
Questions, answered
Written by
The Humanizer Team
Editorial


