Is the Em Dash an AI Tell?

Nothing we can find in the vendor's own documentation supports it, and the next section shows exactly where we looked. We searched three of Turnitin's help pages on 18 August 2026 — the AI Writing Report guide, the AI writing detection FAQs, and the file requirements page. Across those three, the string "dash" occurs exactly once, inside the word "dashboard". The word "punctuation" occurs exactly once, in an answer explaining that the detector "is not tuned to target Grammarly-generated spelling, grammar, and punctuation modifications to content". And the page that answers the question "what parameters or flags does Turnitin's model take into account when detecting AI writing?" answers it entirely in terms of sequences of words. So the most confidently repeated advice about a piece of punctuation rests on nothing we can find the company that built the detector saying about it.

That is not the same as proving a dash cannot matter. A model that learns statistical patterns is under no obligation to publish what it learned, and no public material rules out a dash contributing something. What is missing is the other half: nothing we could find supports weighting one either. Meanwhile the fix people reach for — replace every em dash, everywhere — is a change to almost every paragraph of a document, made in one keystroke, with nobody reading the result.

"Em dashes are an AI tell" gets said as one thing. It is two, and they are joined by nothing:

• Generated text contains a lot of em dashes. A statement about what comes out of language models. Anyone can go and look.
• An em dash raises an AI writing score. A statement about what a particular classifier weights. Nobody outside the vendor can look at that.

The first can be entirely true while the second is false, because a feature being common in one population says nothing about how a detector trained on both populations uses it. Human writers who use the dash heavily are the population the second claim would misfire on, and nothing about the first claim tells you what happens to them.

We are not the first to say the general version of this. Our own page on what AI detectors measure makes the point that a passage does not get a high score because of one visible habit. What follows here is narrower: we went and checked whether this specific habit appears anywhere in the vendor's own documentation, because a claim this widespread ought to have a document behind it somewhere.

Three pages, retrieved 18 August 2026, read as rendered page text and searched case-insensitively: "Using the AI Writing Report", "Turnitin's AI writing detection capabilities FAQs", and "File requirements for an AI Writing Report". Those are the pages that set out the mechanism and the file rules; release notes and known-issues pages were not part of this search. Counts are occurrences, not lines.

The last row is there because a search that finds nothing and a search that never reached the page look identical from the outside. A phrase known to be on all three pages returning one hit each is what makes the zeroes above it mean something.

That leaves one real hit. Here is the whole answer it sits in, to the question of whether Grammarly's grammar checking gets flagged:

"No. Our detector is not tuned to target Grammarly-generated spelling, grammar, and punctuation modifications to content but rather, other AI content written by LLMs such as GPT-3.5. Based on tests we conducted on human-written documents with no AI-generated content in them, in most cases, changes made by Grammarly (free & premium) and/or other grammar-checking tools were not flagged as AI-written by our detector. Please note that this excludes content generated by Grammarly's generative AI-powered features, including draft generation, paraphrasing, summarizing, and other features. Content produced using these features will likely be flagged as AI-generated by our detector."

Read it for what it is. It is about a grammar checker's edits, not about your dashes, it says "in most cases" rather than never, and the last two sentences pull hard in the other direction for anything Grammarly generates rather than corrects. The border between corrected text and generated text is its own subject, and we mapped it in does Turnitin detect Grammarly. What it is useful for here is only this: in the one place across those three pages where punctuation and detection turn up in the same sentence, the sentence points away from punctuation-level tuning rather than towards it.

Then there is the question that would have answered ours directly if the answer went that way. Under the heading "What parameters or flags does Turnitin's model take into account when detecting AI writing?", the substantive part reads:

"Our classifiers are trained to detect these differences in word probability and are adept at the particular word probability sequences of human writers."

Word probability. Sequences of words. That is the level the vendor describes its own classifiers working at, in the answer it gives when its own FAQ asks point-blank what the model takes into account.

Follow the vocabulary down through the published description of how a score is produced and it stops before it reaches a character:

"When a paper is submitted to Turnitin, sentences from the submission are extracted and segmented into overlapping sections for prediction analysis. Each segment is classified by the AI detection model and given a value between 0 and 1 … Each qualifying sentence within these segments inherits the segment's score."

Segments. Sentences. Qualifying text, which the same FAQ defines as blocks written as standard grammatical sentences. Word probability sequences. Every unit named in that chain is a word or larger, and the smallest thing that ever receives a number of its own is a span of several sentences, which a sentence then borrows from.

A dash sits inside that span. Take it out and the segment still contains the same words in the same order — that is the specific thing a punctuation swap does not change, and word order is the thing the vendor named. Whether the tokeniser cares is a separate question that nobody outside Turnitin can answer, and we are not going to pretend otherwise. But the advice is usually delivered as if the mechanism were understood, and the mechanism as published does not have a slot for it.

One measurable consequence of deleting an em dash you can check for yourself in the next thirty seconds, and it has nothing to do with detection. Microsoft Word treats an em dash or an en dash as a word boundary and a hyphen as not one, so "cost—benefit" counts as two words and "well-known" counts as one. Replace forty unspaced em dashes with hyphens and your word count falls by forty.

For most documents that is a curiosity. For a submission near a hard limit it is not, and Turnitin's own file requirements put a ceiling on what will be processed at all — which is why we spent a section on dashes and word counters in can Turnitin check a whole thesis. Notice the shape of this: the effect we can actually demonstrate runs through a word counter, not through a classifier.

An em dash is doing a job in the sentence. Whatever you put in its place does a different job, and the choice is not cosmetic:

• A comma makes the break weaker. Where the dash was carrying a genuine interruption, the comma either loses it or produces a splice.
• A semicolon raises the register and requires what follows to stand alone as a clause. Often it does not.
• Parentheses demote the content to an aside. That changes what you are claiming about its importance, which in a discussion section is an argument, not a formatting decision.
• A full stop splits one thought into two. Sometimes an improvement. Sometimes it severs the link that was the point of the sentence.
• Nothing at all works more often than people expect, and is the only option on this list that does not introduce a new decision.

Now multiply by however many the document holds. Find-and-replace makes every one of those decisions the same way, in one action, in text you had no other reason to open. That is the largest edit most people will make to a finished manuscript, and it is the one made with the least attention.

And if a style guide governs your document, the rule about dashes that binds you is the one in that guide, not the one in a thread. Sorting out which of your dashes are em and which are en is at least a question with an answer.

Strip out the theory about classifiers and something is left standing, but it points at a person rather than a score.

Plenty of readers believe the rule. If your supervisor is one of them, your dashes cost you something real — an impression, formed on the first page, before any detector is involved. That risk does not depend on any account of how the software works, which is exactly why it is the sturdier half of the claim. We think it is the half worth taking seriously, and we are offering that as our reading rather than as anything published.

But notice that a find-and-replace is a poor answer to it. Somebody forming an impression of your writing is reacting to a page, not counting characters, and the same paragraph that struck them as machine-made will strike them the same way with commas in it. What moves a reader is the writing being recognisably yours and your being able to talk about it — which is the argument for keeping drafts, and we set out how that evidence actually gets used in version history as evidence.

One more asymmetry makes writing defensively a bad trade in general. In most configurations you never see the number you are writing against: Turnitin says the AI writing indicator and report are not visible to students, and an instructor has to download and share the report for it to reach you at all. Meanwhile the marks are real and the reader is real.

Whatever you decide about dashes, the useful principle is the one a global replace violates: change text at a scale where you can still read what changed. Every unreviewed change is a place a document can go wrong quietly.

That constraint, and not any view about punctuation, is what shaped HumanPen. Hand it the file plus an AI report from Turnitin or iThenticate, and the marked passages become the scope; everything outside them keeps the wording you wrote. Our own page states the granularity rule plainly: "a paragraph is the smallest unit a rewrite can touch, so a selection is expanded to full paragraphs before it is counted." A paragraph is several levels up from the thing this article is about.

Two things follow that are worth stating as limits rather than features. Billing counts the words actually rewritten, so a scoped job costs what the scope costs rather than what the file weighs. Eligible results can continue lowering AI for free.

What we cannot do is tell you a number. Turnitin's own position on its model is that individual predictions "may not always be explainable in simple feature-by-feature terms", and it has put out nothing that maps an edit onto a movement in the score, so a tool quoting you a resulting percentage is quoting something it does not have. It gets worse below the line, because scores between 1% and 19% are shown as an asterisk with no percentage and no highlights, meaning a real improvement inside that band is invisible to everyone including you. If you want the reasons two tools report different numbers on the same paragraph, that is why detectors disagree; if you are holding a report right now, start with how to read a Turnitin AI writing report.

Will deleting every em dash lower my AI score? Nobody can tell you, and anyone who says definitely yes or definitely no is going beyond what has been published. What we can report is that the three Turnitin help pages we searched contain no mention of dashes, and that the one place punctuation turns up in them at all is an answer saying the detector is not tuned to target punctuation modifications.

Do AI models actually use more em dashes than people? Quite possibly, and that is a claim you can go and test on generated text yourself. It is a different claim from "a detector weights them", and it does not imply it.

Is there any measurable effect of removing them? One you can check yourself, and it is not about detection. Word counts an em dash as a word boundary and a hyphen as not one, so swapping unspaced em dashes for hyphens reduces your word count by one per swap.

My supervisor said my dashes look like ChatGPT. What do I do? Take that as a reader's impression rather than a detector result, because that is what it is. An impression is answered by drafts, notes and being able to walk through your own argument, not by a find-and-replace they will not notice.

Should I write differently to avoid this whole problem? Deliberately writing below your own level has a certain cost in marks and an unmeasured benefit against a score you are usually not permitted to see. The vendor's own published list of what its false positives look like — text with little structural variation, text that repeats itself literally — is not a list of punctuation, and we went through the hand edits it does support in how to humanize AI text without a tool.