Does AI Use Em Dashes? Why It Does and What That Means
12 min read
Yes, AI chatbots such as ChatGPT and Claude use em dashes, and many of their answers carry several. That habit is why so many writers now worry about their own dashes.
Language models pick up punctuation from the human writing they learn from, and later training may strengthen the habit. No AI company has published a confirmed cause.
People use em dashes too, so one dash can't prove who wrote a passage. Pew Research Center found that em dashes now appear about twice as often on sampled webpages as they did in 2023 (August 2026 analysis). That rise describes web writing as a whole, and it can't test any single draft.
This guide separates evidence from theory, shows when to keep or replace a dash, and explains what to do if someone questions your work.
Check Any Draft With Phrasly's AI Detector 👇
Paste at least 50 words to get one score for the whole passage.
Why Does AI Use Em Dashes?
AI models use em dashes because they learn writing patterns from huge amounts of human text, and published books and journalism use the dash often. Later training that rewards polished, conversational answers may strengthen the habit. The companies behind these models haven't explained the overuse, so every explanation for it is still a theory.
📊 What Pew Measured
On sampled English-language webpages, Pew Research Center's study counted 5.79 em dashes per 10,000 words in January 2023 and 11.19 in January 2026.
Pew adds that AI models "tend to use these dashes a lot more than humans typically do." Even so, Pew says em dashes "on their own don't necessarily mean a particular piece of writing was produced using AI."
Research papers show the same rise. In a June 2026 preprint that hasn't been peer reviewed, Przemysław Czuma found that medRxiv Discussion sections with an em dash rose from 4.23% before ChatGPT to 11.58% after.
🧠 Models Learn Punctuation From Human Writing
A large language model learns punctuation from the patterns in its training text. OpenAI's GPT-4 Technical Report says the model was pre-trained to predict the next token, meaning the next small piece of text, using public and licensed data.
OpenAI then refined GPT-4 with reinforcement learning from human feedback (RLHF), a step where human ratings shape the answers. The model doesn't copy sentences from memory. It learns which patterns are likely, and the em dash is one of them.
Why the Overuse? What's Known vs. What's Proposed
Only the general training method is confirmed, and the other explanations range from likely reasons to one expert's analogy.
Explanation | What the evidence supports |
|---|---|
Models learn punctuation from their training text | ✅ Known. This is the general training method, described in the GPT-4 Technical Report. |
Dash-heavy published writing in the training data plays a part | ⚠️ Likely, as one reason. It doesn't explain why models use more dashes than humans do. |
Human raters preferred dashes because they sound conversational | 🔍 Hypothesis. Engineer Sean Goedecke says this "could be right," but he doesn't know how anyone would prove it. |
Newly digitized print books from the late 1800s and early 1900s caused the overuse | 🔍 Goedecke's favored theory, which he calls "still largely based on speculation." |
Engineers grade sources, such as a novelist's work at 10 out of 10 | 💬 One expert's analogy from Mustafa Ocal, a computing professor at FIU, who sums up AI as "It predicts." |
An em dash proves AI wrote the text | ❌ False. People use the same mark. |
Not Every Model or Answer Uses Them
AI tools don't always use em dashes, because output changes with the model version and with what you ask it to do. OpenAI CEO Sam Altman said in November 2025 that ChatGPT now follows a custom instruction to stop using them, as TechCrunch reported.
TechCrunch also noted that the fix won't necessarily remove dashes by default. The dash is a default style, and users can switch it off.
Are Em Dashes a Sign of AI Writing?
An em dash on its own is not reliable evidence of AI writing. Writers used the mark long before chatbots existed, and AI output can contain no dashes at all. A repeated dash pattern can be a reason to review a draft's rhythm, but it can't show who wrote it.
Pew's trend comes from almost half a million webpages, and a pattern measured at that scale can't be read from one essay. People also tend to mix up three different signals when they judge a draft:
✍️ Style tendency: what a model often does, such as adding dashes.
📈 Detector estimate: a probability-style score that a tool gives for a passage.
🗂️ Authorship evidence: drafts, notes, and saved versions that show how the work was made.
What you notice | Reasonable reading | What it can't establish |
|---|---|---|
One em dash in a paragraph | An ordinary punctuation choice | Who wrote it |
Dashes in a repetitive pattern | A reason to review rhythm and clarity | That AI was used |
AI-written text with no em dashes | Punctuation varies by prompt and model | That a person wrote it |
A detector highlights a sentence with a dash | That passage deserves a closer read in context | That the dash caused the result |
Phrasly's study of common AI phrases found the highest dash averages in LinkedIn posts (3.6) and blog posts (3.15) across 50 samples. The study still concluded that dashes don't prove AI use. Phrasly's explainer on how AI detectors work gives the same rule: no single word, phrase, or punctuation mark proves AI use.
If you want one more review signal on a full passage, Phrasly's free AI checker gives a score you can compare with your notes and revision history.
Do AI Detectors Flag Em Dashes?
There's no public evidence that mainstream AI detectors flag text because of one punctuation mark. Detectors score statistical patterns across a whole passage, and a highlighted sentence doesn't show which feature drove the score. Detectors can also misread human writing.
Different tools can also score the same text differently, and why AI detectors give different results comes down to the separate models and training data each one uses.
⚠️ Turnitin's Own Warning
Turnitin's guide to Using the AI Writing Report says its model "may misidentify human-written, AI-generated, and AI-paraphrased text." For that reason, Turnitin says the report "should not be used as the sole basis for adverse actions against a student."
🔐 Are Em Dashes an AI Watermark?
An em dash is ordinary punctuation, and AI companies don't use it as a watermark. The text watermarks announced in 2026 are invisible: they adjust word choices in a pattern that only a special detector can check.
Company | What it covers | How it works | Who can check it |
|---|---|---|---|
OpenAI (October 2026) | ChatGPT and Codex text in the EU, rolling out over the coming weeks. API customers worldwide can opt in. | An "invisible statistical signal" in the model's word choices | Approved researchers and expert organizations only, at first |
Anthropic (August 2026) | Claude text worldwide | A secret key guides word choices, based on Google DeepMind's SynthID-Text, with "no hidden characters" | Eligible organizations, through a detection API in private preview |
Watermark details checked on OpenAI's and Anthropic's own pages in October 2026. Both policies can change, so confirm the current details on each company's site.
Edits weaken these signals. OpenAI's own test found that swapping 10% of words for synonyms cut detection from about 92% to 66%. OpenAI also says a missing watermark "does not prove human authorship."
FIU's Ocal has wondered whether the dash acts as a watermark, and he offered the idea as his own speculation. A deleted dash can't touch a word-level signal, so a ChatGPT watermark has nothing to do with punctuation.
Should You Stop Using Em Dashes Because of AI?
You don't need to stop using em dashes because AI tools use them. Keep a dash when it makes a sentence clearer or adds deliberate emphasis, and follow your style guide or your client's rules. If dashes appear in every paragraph, vary your punctuation so the text reads better for your audience.
AP Stylebook editor Anna Jo Bratton wrote in an August 2025 AP Stylebook post that the em dash is not an easy tell for AI writing, because writers have used it for many years.
AP has no limit on how many dashes a piece can use, but it tells writers to "avoid overuse of dashes to set off phrases when commas would suffice." AP style also puts a space on both sides of a dash, so check spacing against your own guide.
💡 When a School or Client Bans Dashes
Follow the ban as a house style rule for that project. The rule reflects their editing preferences, and it has no link to how AI detection works.
How to Edit Em Dashes Without Weakening Your Writing
Replace each em dash based on the job it does in the sentence. Use a pair of commas for a light aside, a colon to introduce an explanation, and a period when the dash joins two separate thoughts. Keep the dash when you want a strong, deliberate break, and never swap every dash for a comma.
Em Dash vs. En Dash vs. Hyphen
An em dash () is the longest common dash, about the width of a capital M, and it marks a strong break or an aside.
A hyphen joins words, and an en dash usually shows a range, such as a span of pages.
Mark | Example | Job |
|---|---|---|
Hyphen - | a well-known author | Joins the parts of a compound word |
En dash | pages 10-15 | Shows a range (rules vary by style guide) |
Em dash | I checked the draft, twice. | Marks a strong break, an aside, or emphasis |
How Do You Find Every Em Dash in a Draft?
Search for the character itself, then decide on each dash one at a time:
Google Docs: press Ctrl + F (⌘ + F on a Mac), paste an em dash into the search box, and step through each match.
Microsoft Word: open Find and type ^+ to find every em dash. Type ^= to find en dashes.
⚠️ Avoid a Blanket Find-and-Replace
Comma swaps across a whole draft create comma splices, where only a comma joins two full sentences.
✏️ Four Quick Edits
Commas, a colon, and a period each replace a dash in a different situation, and some dashes should stay.
Situation | Before | Edit | Why it works |
|---|---|---|---|
Light aside | The revised draft, which includes the new figures, is ready. | The revised draft, which includes the new figures, is ready. | Commas make a lighter interruption. |
Explanation | The launch depends on one thing, approval. | The launch depends on one thing: approval. | A colon introduces the explanation after a complete clause. |
Separate thoughts | I checked the source, it had been updated. | I checked the source. It had been updated. | Two sentences make the order of events easy to follow. |
Deliberate emphasis | We had one hour, and no backup plan. | ✅ Keep it. | The dash carries the rhythm the writer wants. |
Editing examples, not an AI-authorship test.
How to Get ChatGPT and Other AI Tools to Use Fewer Em Dashes
Put a no-dash rule where the tool saves it, such as ChatGPT's custom instructions, which ChatGPT now follows for em dashes. In any AI tool, ask for no U+2014 characters, which is the em dash's Unicode code. Ask the tool to choose grammatical replacements, and proofread the result yourself before you use it.
📋 Copy this prompt:
Write in plain, direct sentences and don't use the em dash character (U+2014).
Where a dash would go, pick a period, a colon, parentheses, or a pair of commas, based on the grammar of the sentence.
Keep every fact, number, and citation exactly as it is. Don't replace every dash with a comma.
Before you reply, scan your text for U+2014 and fix any you find.
When we ran this prompt in ChatGPT in October 2026, the rewrite came back with no em dashes and every fact unchanged. Results can vary by model and by task, so check the output each time. Phrasly's guide to prompts to humanize AI text lists more fixes for common AI habits, including overused em dashes.
What to Do if Someone Says Your Em Dashes Prove AI Use
Start by asking which passage and which rule the reviewer means, then show the record of how you wrote it: drafts, notes, sources, and version history. Older writing that already uses dashes helps too. A detector score can add context, but no score or punctuation mark decides authorship alone.
Ask for specifics. Find out which passage concerns the reviewer and which policy applies.
Share your process. Show drafts, outlines, source notes, and version history from Google Docs or Word.
Show your baseline. Point to earlier writing that uses dashes the same way, because your usual style matters.
Add a detector result if it helps. Run the passage through a detector and explain the tool's limits along with the score.
Talk about the work itself. Walk through your argument and sources, because no mark or score replaces that conversation.
🔎 A quick Phrasly check:
Paste the full passage of at least 50 words.
Read the percentage score and the highlighted sentences.
Compare the result with your drafts and notes.
Phrasly's AI Detector is free: your first three scans need no account, and a free account unlocks unlimited scans. The full walkthrough on how to scan text for AI covers file uploads and longer drafts.
Review the Draft Before You Rewrite It
Run the full passage through Phrasly's AI writing detector before you delete any dashes, then check the score against your drafts and source notes.
Keep version history turned on in your writing app, so every future draft leaves a record.
Em dashes are a shared habit of people and AI tools, so treat each one as a style choice and edit it for meaning. The 2026 watermarks from OpenAI and Anthropic work through invisible word choices. A record of your drafts and notes answers authorship questions better than any single mark can.
FAQs
Does ChatGPT Always Use Em Dashes?
ChatGPT uses em dashes in some answers and not others, because its punctuation depends on the model version, your instructions, and the task. ChatGPT also obeys a no-dash custom instruction, so neither a dash nor its absence identifies the writer.
Does Turnitin Flag Em Dashes as AI?
Turnitin hasn't published any rule that flags punctuation alone. Its report reviews sentences in full prose, and Turnitin says the result shouldn't be the only basis for action against a student. If your work gets flagged, see what to do when an essay is detected as AI.
Are Em Dashes a ChatGPT Watermark?
Em dashes aren't a ChatGPT watermark. OpenAI's text watermark, which began rolling out in the EU in October 2026, adds an invisible signal to word choices and needs a special detector. People use the visible em dash every day as normal punctuation.
How Many Em Dashes Are Too Many?
No validated number separates human writing from AI writing. If dashes appear in most paragraphs, or several show up in one paragraph, read the passage aloud and replace the ones that interrupt more than they clarify. Your style guide or editor has the final say.
Will Removing Em Dashes Change an AI Detection Result?
Em dash removal won't reliably change a detection result. Detectors weigh many patterns across a passage, so a punctuation swap rarely changes how the whole draft reads, and it can't change how the draft was produced. Edit for clarity, and follow any rules on disclosing AI help.
What Should I Do if My Human Writing Is Flagged?
Keep your drafts, notes, and version history, and ask the reviewer to go through the flagged passage with you. Detector mistakes happen, as Phrasly's guide to AI detector false positives explains, so walk through your process instead of arguing about one score.
Written by
Alina Shah
SEO Content Specialist · Karachi, Pakistan
She writes about AI so you don't have to guess. 8+ years in content strategy and editing. Now she puts AI writing tools and detection systems through real tests and shares what actually works.
