STEP B
"No markdown" misses the dash
A plain-prose rule removes headers and bullets, but models do not file the em dash under formatting, so GPT-4.1 kept 8.20 per 1,000 words under that rule.
EM DASHES AND AI, OCTOBER 2026
Em dash AI: why models use it and what the data shows
Everyone has a theory about why chatbots love the long dash. Here is each one graded against the evidence, with measured rates for 12 models, the fix OpenAI shipped, and what changed in 2026.
Scan my text free Updated
chatgpt.com
Today
Why do you use so many em dashes?
Message ChatGPT
MacBook Pro
9:41
ClaudeCleanLocal scanPASTED FROM CHATGPT
SCANNING
…em dashes
U+2014 #1em dash to comma1
U+2014 #2em dash to word1
U+2014 #3em dash to period1
Three em dashes found and swapped. Keep them instead with the rewrite's Keep em dashes option.
9:41
ClaudeCleanLocal scanPASTED FROM CHATGPT
SCANNING
…em dashes
U+2014 #1em dash to comma1
U+2014 #2em dash to word1
U+2014 #3em dash to period1
Three em dashes found and swapped. Keep them instead with the rewrite's Keep em dashes option.
Short answer
AI uses em dashes because it learned them from human text and fine-tuning amplified the habit. The best-tested theory says markdown-heavy training leaves a structural reflex that survives in plain prose. In a 2026 test, GPT-4.1 wrote 10.62 per 1,000 words against 3.23 for human essays. Newer models use far fewer.
THE THEORIES, GRADED
Why does AI use so many em dashes?
Six explanations circulate. Two have measurements behind them, two are informed guesses, and two have been checked and found wanting.
| Theory | Verdict | What backs it |
|---|---|---|
| Markdown habit leaking into proseBest evidence | Measured, in one study | Told to write plain prose, all 12 models in Freeburg's test dropped headers and bullets to zero (Claude Haiku 3.5 kept a trace), yet GPT-4.1 kept 9.10 em dashes per 1,000 words and DeepSeek V3 kept 5.41. A preprint from an independent researcher, not yet peer reviewed. |
| Fine-tuning turns the dial | Supported | Base Llama 3.1 8B wrote 0.49 per 1,000 words and its tuned version wrote none, while OpenAI's tuning pushed GPT-4.1 to 14.03 in long essays. Sean Goedecke notes GPT-3.5 barely used em dashes and GPT-4o used about ten times more. |
| Old books in the training data | Plausible, unproven | Goedecke found em dash use in English print peaked around 1860 at roughly 0.35% of words, and points to court filings showing Anthropic began scanning print books in February 2024. He calls his own case "still largely based on speculation". |
| Too many dashes in "good" text | Expert opinion | FIU's Mustafa Ocal argues that edited, high-quality sources are full of em dashes, so a model that predicts likely tokens reproduces them. McGill counts 1,621.88 per million words in OpenWebText, a web corpus. |
| The raters' own dialect | Weakened | The idea was that human raters from regions with a dash-heavy writing style rewarded it. Goedecke checked a Nigerian English corpus and found em dashes at about 0.022% of words, far below 19th-century print. |
| Tokens, versatility, Medium, OCR | Set aside | Goedecke weighed the mechanical stories (the dash is a cheap token, a safe all-purpose joiner, an artifact of Medium's autoformatting or of scanned-text errors) and found none that fit the data. No study supports them. |
Swipe to compare →
Graded October 2026. Measured means a controlled test exists; plausible and expert opinion mean the argument fits but nobody has isolated it.
We propose that the em dash is markdown leaking into prose.
The theories stack rather than compete. Books and the web supply the em dash in the first place, chat training data full of headers and bullets teaches models to break text into joined pieces, and each company's fine-tuning decides how loud that reflex gets. Freeburg's own summary is that em dash frequency works as a signature of the fine-tuning recipe, which is why the same mark shows up at 10 per 1,000 words in one model and zero in another.
THE NUMBERS
How often does each AI model use em dashes?
The only side-by-side measurement of 12 models we could find, run in February and March 2026 on ten essay topics each.
12 MODELS, ONE PROMPT
Em dashes per 1,000 words, by model
Each model wrote ten 1,000-word essays. Switch the prompt to see which models drop the habit when told to write plain prose.
- GPT-4.110.62
- Claude Opus 4.69.09
- Claude Sonnet 48.29
- Claude Haiku 3.57.51
- DeepSeek V36.95
- GPT-4o Mini4.16
- GPT-4o4.12
- Gemini 2.5 Pro3.53
- GPT-5.41.43
- Gemini 2.5 Flash1.28
- Llama 3.1 8B0.00
- Llama 3.3 70B0.00
Human essays, mean 3.23 Above the human mean8 of 12 models above the human mean
Read the human number with care
The human baseline is eight essays, and single essays ranged from 0.33 to 17.12 em dashes per 1,000 words. The paper itself says the finding is not that models overuse em dashes against all human writing. The pattern is that the mark resists formatting rules and varies by provider.
GOT A DASH-HEAVY DRAFT?
See every em dash in your draft, then decide what stays.
ClaudeClean marks em dashes and hidden characters in your browser and can swap the dashes for commas in one pass. Writing them on purpose? The rewrite has a Keep em dashes option.
- 300 rewrite credits a month
- 20,000 characters per request
- .txt, .docx and .pdf files
PROOF OR NOT
Does an em dash mean AI wrote it?
It does not. Try the scale below on your own writing, then weigh what the mark can and cannot tell you.
CHECK A PASSAGE
Where does your text land on the em dash scale?
Paste anything. The count runs in your browser and nothing is sent anywhere. The sample is the opening of Moby-Dick (1851), from Project Gutenberg.
Words
145
Em dashes
3
En dashes
0
Double hyphens
0
Em dashes per 1,000 words20.69
GPT-5.4Human meanOpus 4.6GPT-4.1
0Green band: human essays 0.33 to 17.1220+
Under 150 words, a single dash swings the rate a lot. Paste a longer passage for a fairer read.
What a careful reader looks at instead
- Patterns across many tells. Graphite says the overall number of AI tells per article is holding steady even as the em dash fades.
- Drafts and history. Version history, notes and earlier drafts say more about authorship than punctuation can.
- Context of the writer. A writer who always used em dashes will keep using them; a sudden shift is more telling than the mark itself.
- The full picture. Our guide to signs of AI writing covers the other tells, and how to prove you didn't use AI covers false accusations.
Why the em dash is a weak signal
- Humans vary wildly. Published essays in Freeburg's sample ran from 0.33 to 17.12 per 1,000 words, a fifty-fold spread.
- Classic prose is dash-heavy. Goedecke counts 1,728 em dashes in Moby-Dick, and Emily Dickinson built poems around them.
- Models have changed. Graphite found GPT-6 Astra and Gemini 3.1 Pro now use fewer em dashes than human writers.
- It is easy to fake either way. One custom instruction removes them from AI text, and anyone can type one into their own.
That dramatic pause, it allows a voice to breathe on the page.
NPR described a "backlash to the backlash": writers who refuse to drop the mark because chatbots took a liking to it. The Ringer's Brian Phillips made the same case in August 2025, writing that "Human writers have always used the em dash."
THE BAN TEST
Why telling AI to stop using em dashes used to fail
Users spent 2025 asking ChatGPT to quit the habit and watching it carry on. A three-step experiment shows why.
THE BAN TEST
What happens when you tell a model to stop
Same four models, three prompts. The last prompt names the em dash outright.
PromptWrite a 1000-word essay about [topic].
GPT-4.111.51
Baseline, per 1,000 wordsDeepSeek V38.66
Baseline, per 1,000 wordsClaude Opus 4.68.46
Baseline, per 1,000 wordsGPT-5.40.75
Baseline, per 1,000 wordsSTEP C
Naming it helps, mostly
Adding "Do not use em dashes" took Claude Opus 4.6 and GPT-5.4 to zero, but GPT-4.1 still wrote 3.86 and DeepSeek V3 wrote 1.57.
LONG TEXT
Length makes it worse
In 5,000-word essays GPT-4.1 reached 14.03 per 1,000 words and kept 6.97 under a direct ban. Longer text has more joints for the reflex to fill.
THE NOVEMBER 2025 FIX
What did OpenAI change about em dashes?
ChatGPT now follows a no-em-dash rule when it lives in custom instructions. The default did not change.
THE NOVEMBER 2025 FIX
One custom instruction, before and after
9:41
ChatGPT
Draft a two-line note: the launch moves to Friday.
Sure — here is a draft. The launch moves to Friday — QA needs two more days. Everything else stays the same.
2 em dashesAsk anything
If you tell ChatGPT not to use em-dashes in your custom instructions, it finally does what it's supposed to do!
A setting, not a new default
TechCrunch noted the fix does not necessarily remove the em dash from ChatGPT's output by default; it gives users control. Tom's Guide walks through the menu path, Settings then Personalization then Custom instructions, and reports some users still see the occasional dash.
MODEL BY MODEL
Do Claude and Gemini use em dashes too?
They did, at different rates, and by October 2026 most big models had cut back hard. Two independent measurements, side by side.
| Model family | Freeburg, Mar 2026 (per 1,000 words, essays) | Graphite, Sep to Oct 2026 (articles) | Direction |
|---|---|---|---|
| OpenAI (ChatGPT) | GPT-4.1 10.62, GPT-4o 4.12, GPT-5.4 1.43 | GPT-6 Astra uses em dashes about one eighth as often as human writers, and 95% less often than GPT-5 | Down sharply since GPT-4.1 |
| Anthropic (Claude) | Opus 4.6 9.09, Sonnet 4 8.29, Haiku 3.5 7.51 | Opus 5 at 2.92 per 1,000 words, about the human rate. Opus 5.5 at 0.015, down 99% | Near zero in Opus 5.5 |
| Google (Gemini) | 2.5 Pro 3.53 (three topics only), 2.5 Flash 1.28 | Gemini 3.1 Pro has nearly stopped using them | Low, then lower |
| DeepSeek | V3 6.95, still 1.57 after a direct ban | Not tested | No 2026 follow-up found |
| Meta (Llama) | 3.1 8B and 3.3 70B both 0.00 across about 40,000 words | Not tested | Never had the habit |
Swipe to compare →
Freeburg prompted 1,000-word essays with default settings; Graphite compared 90,000 AI articles from nine models with 10,000 human articles written before ChatGPT. The methods differ, so compare within a column.
The tell is reversing
Graphite's September 2026 report says "AI models no longer overuse it consistently" and that GPT and Gemini may have overcorrected. A text with zero em dashes is now as likely to come from a current model as from a person who never liked the mark.
THE GLYPH
Em dash vs en dash vs hyphen: which one does AI use?
The habit is the em dash, —, Unicode U+2014. Freeburg counted that exact character. The en dash and hyphen are different marks with different jobs.
THREE DASHES, ONE KEYBOARD
Em dash vs en dash vs hyphen
Pick a dash to see its code point, its width next to a capital M in this page's font, and how to type it.
Em dash U+2014
Measuring width
- Code point
- U+2014 (decimal 8212)
- Width
- One em wide, traditionally the width of a capital M
- Mac
- Option + Shift + Hyphen
- Windows
- Alt + 0151 (numeric keypad)
- Used for
- Breaks in a sentence, asides, interruptions. This is the one people call the AI dash.
The launch moved to Friday — QA needed two more days.
THE HABIT, LIVE
Watch the em dashes land in a chatbot reply
em dashes so far: 0
chatgpt.com
ChatGPT
Write a short note telling the team the launch moves to Friday.
AI
Ask anything
MacBook Pro
In chatbot replies the em dash mostly stands in for a comma, a colon, a pair of parentheses or a full stop. That swap is why the sibling guide on removing em dashes from AI text picks a replacement per sentence instead of deleting the mark.
HOW WE GOT HERE
How the em dash became an AI tell
From a quiet model quirk to a punctuation panic, then a fix and a fade. Model changes and the public argument, in one line.
Nov 30, 2022
ChatGPT launches on GPT-3.5
OpenAI's launch post opens the floodgates. Per Goedecke, this model barely used em dashes.May 2024
GPT-4o arrives with the habit
Goedecke estimates GPT-4o used about ten times as many em dashes as GPT-3.5, and GPT-4.1 more still.Feb 2025
"Em dash means AI" goes viral
Posts calling the mark a giveaway spread, as the Daily Dot and later The Ringer documented.Jun 17, 2025
Grammar Girl takes the question
Episode 1091 of the Grammar Girl podcast asks whether em dashes are a sign of AI writing and where the idea came from.Aug 20, 2025
Writers push back
The Ringer runs "Stop AI-Shaming Our Precious, Kindly Em Dashes, Please", following The Federalist in July.Nov 10, 2025
NPR covers the backlash to the backlash
All Things Considered reports Reddit users asking whether to drop the mark, and writers who refuse.Nov 14, 2025
OpenAI ships the custom-instruction fix
Sam Altman says ChatGPT now honors a no-em-dash instruction, per TechCrunch. Defaults stay as they were.Mar 27, 2026
The first 12-model measurement
Freeburg's "The Last Fingerprint" links em dashes to markdown training and shows bans only partly work.Sep 16, 2026
Graphite: the tell is fading
Graphite's AI Tells report finds GPT and Gemini models now use fewer em dashes than people.Oct 1, 2026
Claude Opus 5.5 nearly drops them
Graphite's update measures 0.015 em dashes per 1,000 words for Opus 5.5.
MYTHS, CHECKED
Four em dash claims that do not hold up
Each one shows up in comment threads and detector checklists. Here is what the sources say.
"The em dash is a hidden AI watermark."
Text watermarks work through word choice. OpenAI's help center says its text watermark "does not add hidden characters, invisible spaces, or unusual punctuation." FIU's Ocal floats the idea of a punctuation watermark as speculation only. See our ChatGPT watermark guide for how the real one works.
"People hardly ever use em dashes."
English print around 1860 ran at about 0.35% of words, per Goedecke, and the opening paragraph of Moby-Dick holds three. Modern essayists in Freeburg's sample averaged 3.23 per 1,000 words.
"Every chatbot does it."
Meta's Llama models wrote zero em dashes across about 40,000 words in Freeburg's test, and Freeburg checked they were not using double hyphens or en dashes instead.
"Asking nicely never works."
It works better than it did. Since November 2025 ChatGPT honors the rule in custom instructions, and in Freeburg's ban test Claude Opus 4.6 and GPT-5.4 went to zero. Older models like GPT-4.1 still slipped.
WHAT TO DO
Fewer em dashes, or none: your options
Pick based on whether you control the model, the draft, or neither.
BEFORE IT WRITES
Set a standing instruction
In ChatGPT, add a no-em-dash rule to custom instructions and name the replacement (comma or period). In Claude or other tools, put the same line in a project or system prompt.
AFTER IT WRITES
Clean the draft
Paste the text into the ClaudeClean cleaner to find every em dash and swap it, or follow the step-by-step replacement guide to fix each one by hand.
IF YOU WROTE IT
Keep your dashes
If the em dash is part of your voice, keep it. A single mark proves nothing, and dropping it will not fix text that reads as generated. For broader edits, see the Claude humanizer guide.
SEE THE CLEANUP
Find every em dash, then choose what to do with it
ClaudeClean works on text from ChatGPT, Claude, Gemini or your own drafts. The scan never leaves your browser.
01
> source text
Paste the draft
Paste text or upload a .txt, .docx or .pdf file. Up to 2,000 characters works without an account.
02
scan report
3 zero-width1 soft hyphencleanable signals2Scan in your browser
Em dashes, zero-width spaces, no-break spaces and other hidden characters are flagged locally, and the clean pass turns em dashes into commas.
03
before → after
clearusefulanswerfocusedhelpfulresponse
Copied ✓Rewrite only if you want
Light, Medium or Strong passes reword AI habits with a before and after view, with a Keep em dashes option. Rewrites send text to OpenAI and are best effort.
PRICING
Clean the text free. Rewrite more with Pro.
The scan and hidden-character cleanup are free. Pro adds room for rewrites and files.
PRO
CLAUDECLEAN PRO
ClaudeClean Pro
$5/ month
Cancel anytime.
- 300 credits every month (≈300,000 characters of rewriting)
- 20,000 characters per request
- Clean .txt, .docx and .pdf files
- Light / medium / strong intensity
- Dashboard with history
Checking account…
Free: 10 credits a month with an account, or try 2,000 characters right here without one. Strong rewrites cost 2× credits.
WATCH
Grammar Girl on em dashes and AI writing
A 17-minute podcast episode from June 2025 that traces where the em dash theory started.
SOURCES
Where these facts come from
Measurements first, then the expert takes and the reporting on the em dash argument.
arXiv (E. M. Freeburg)
The Last Fingerprint: How Markdown Training Shapes LLM Prose
March 2026 preprint: 12 models, five providers, em dash rates per 1,000 words, the ban test and the base-vs-tuned Llama comparison.
View sourceGraphite
AI Tells
September 16, 2026 report on 90,000 AI and 10,000 human articles; finds models no longer overuse em dashes consistently.
View sourceGraphite
AI Tells: Claude Opus 5.5 update
October 1, 2026: Opus 5.5 at 0.015 em dashes per 1,000 words against 2.92 for Opus 5.
View sourceSean Goedecke
Why do AI models use so many em-dashes?
October 2025 essay: 19th-century print rates, Moby-Dick count, GPT-3.5 vs GPT-4o, and the theories he rejects.
View sourceMcGill Office for Science and Society
Why did LLMs steal our em dashes?
May 2026 explainer on Freeburg's results and OpenWebText's em dash rate.
View sourceFIU News
Why does AI use so many em dashes? An expert explains
June 2026 interview with computing professor Mustafa Ocal on prediction and high-quality training text.
View sourceTechCrunch
OpenAI says it's fixed ChatGPT's em dash problem
November 14, 2025: Sam Altman's announcement and the note that defaults are unchanged.
View sourceTom's Guide
ChatGPT finally lets users disable its em dash habit
Menu path for the custom instruction and early reports of occasional slips.
View sourceNPR
Inside the unofficial movement to save the em dash from AI
November 10, 2025 segment on writers pushing back against the em dash stigma.
View sourceThe Ringer
Stop AI-shaming our precious, kindly em dashes
August 20, 2025 essay by Brian Phillips defending the mark.
View sourceWikipedia
Dash
Unicode code points and the traditional widths of the em dash and en dash.
View sourceZapier
How to type an em dash on any keyboard
Mac Option shortcuts and Windows Alt codes for the em dash and en dash, with The Next Web as a second source.
View sourceFAQ
Em dash and AI FAQ.
The best-tested explanation is a 2026 study by E. M. Freeburg: models learn a structural habit from markdown-heavy training text, and the em dash is the piece of it that survives when they write plain prose. Fine-tuning then turns that habit up or down, which is why rates differ so much between companies.
Older ChatGPT models were tuned in a way that amplified the habit. In Freeburg's test, GPT-4.1 wrote 10.62 em dashes per 1,000 words against a human essay mean of 3.23, and still wrote 3.86 after being told not to. Newer OpenAI models use far fewer: GPT-5.4 wrote 1.43.
No. Human essays in Freeburg's sample ranged from 0.33 to 17.12 em dashes per 1,000 words, and the first paragraph of Moby-Dick has three. Several 2026 models now use fewer em dashes than the average human writer, so the mark proves nothing on its own.
The same as anywhere else: a break in the sentence, used for an aside, an interruption or a sharp turn, often where a comma, colon or period would also work. It carries no hidden code or signal. It is a regular Unicode character, U+2014.
It did, then it stopped. Claude Opus 4.6 wrote 9.09 per 1,000 words in Freeburg's essays but only 0.19 when asked for plain prose. Graphite measured Claude Opus 5 at 2.92 per 1,000 words, about the human rate, and Opus 5.5 at 0.015, a 99% drop.
Less than most. Gemini 2.5 Pro wrote 3.53 per 1,000 words in Freeburg's study and none when asked for plain prose; Gemini 2.5 Flash wrote 1.28. Graphite's September 2026 report says Gemini 3.1 Pro has nearly stopped using them.
Put the rule in Settings, Personalization, Custom instructions, for example: do not use em dashes, use commas or periods instead. Sam Altman said in November 2025 that ChatGPT now follows that instruction. Start a new chat after saving, and expect the odd slip.
No. Text watermarks from OpenAI and others sit in word choice, and OpenAI says its text watermark does not add unusual punctuation. An em dash is ordinary punctuation that any writer can type, so it cannot work as a reliable signature.