ChatGPT Em Dash Tell: Surface Habits That Give a Draft Away

The ChatGPT em dash tell is real but not decisive. The dash itself is neutral; what reads as machine-made is frequency plus uniform placement. Three other surface habits do more damage: triads used as rhythm, stacked hedges, and bolded list intros that announce a point before making it. Fix them by replacement, not deletion.

Why does ChatGPT overuse em dashes?

Because the model was trained on prose where em dashes do a lot of work — magazine features, essays, business writing — and it has no ear for when the dash earns its keep. It reaches for the dash the way a nervous speaker reaches for “you know”: as a pause that avoids committing to a comma, a colon, or a full stop. The result is a draft where every third sentence has an interruption in the same syntactic position.

That uniformity is the tell. One em dash in a paragraph is invisible. A dash in the same slot — clause, dash, aside, clause — across four consecutive paragraphs is a fingerprint. Readers notice rhythm before they notice vocabulary.

The same is true of the other surface habits in this article: the problem is not the device, it is the regularity. A human writer uses a dash when the aside doesn’t warrant a parenthesis and a comma would be misread. A model uses a dash because the token was likely.

If you want the mechanical version of this argument, the pillar page on humanizing AI text walks through how detection signals accumulate across a document rather than sitting in any single sentence.

What are the four surface habits that give a draft away?

Most flagged drafts share four patterns. They show up together often enough that finding one should send you looking for the others.

Habit	What it looks like	Why it reads machine-made
Em dash overuse	Clause—aside—clause, repeated in the same slot	Uniform placement; the writer never chooses a different pause
Triads	“clear, concise, and compelling”; “faster, cheaper, and more reliable”	Rhythm substitutes for argument; the third item is often filler
Stacked hedges	“may potentially suggest”; “could arguably be seen as”	Two hedges where one would do; the writer is not committing
Bolded list intros	Speed. Then a sentence explaining speed.	The bold announces the point before making it; the sentence repeats the label

Each of these is a surface habit — it lives in punctuation, rhythm, or formatting rather than in the ideas. That matters, because surface habits are the easiest thing to fix without touching your argument, your data, or your citations.

If you want to see which sentences in your own draft are carrying the signal, run it through the DeAIze AI detector first. It scores the whole document and highlights each sentence by AI probability, so the dash-heavy paragraphs sort themselves out before you start editing.

Why do triads and stacked hedges read as machine-made?

Triads feel complete. Three items satisfy a pattern the ear recognises from speeches, slogans, and headlines, so a triad can pass for a finished thought even when the third item adds nothing. Models produce them constantly because three is a common shape in the training data. The problem for a reader is that the triad closes a sentence that should still be open — it signals “this list is exhaustive” when the writer only had two real items and a synonym.

Stacked hedges are the opposite failure. Where a triad over-commits, hedging under-commits. “This may potentially indicate that the results could suggest” has four hedges in one clause. A human writing carefully hedges once, in the place where the uncertainty actually lives: “This suggests the effect is small.” Two hedges in the same clause usually means the writer did not decide what they believed.

Both habits read as machine-made for the same underlying reason: they are decisions made at the level of form rather than content. A triad is chosen because three sounds good. A hedge stack is produced because the model is trained to avoid overclaiming. Neither is chosen because the sentence needed it.

Why do bolded list intros feel like a machine wrote them?

The pattern is familiar from model output and from a certain kind of SEO blog: a bolded phrase, a full stop, then a sentence that restates the phrase in longer form.

Latency. Latency is the delay between a request and a response.

The bold label and the sentence say the same thing. The reader learns nothing from the second half. This is not a formatting crime — plenty of human writers use bolded run-in heads — but it becomes a tell when every item in a list follows the identical shape, and when the bolded phrase is a category label rather than a claim.

A human version makes the bold do work the sentence cannot:

Latency is the cost you pay per request. Add a hundred calls and it compounds.

The bold carries the argument. The sentence extends it. If your bolded intro could be deleted without losing meaning, it is a label, not a heading.

What does our own measurement say about surface tells?

We ran our detector over two corpora, using the local ensemble at the lite tier with a 30% flag threshold. The first corpus was 46 hand-written DeAIze guides — human prose, written by people, on deadline. Mean AI score: 27%. Median: 25%. Range: 0–55%. Seventeen of the 46 were flagged at the threshold.

That is the number worth sitting with. Nearly two in five human-written guides crossed a 30% line. If a detector can flag human writing that often on prose that was never near a model, then a surface tell is not evidence of authorship. It is a signal — one of several — and a weak one on its own.

The second corpus was 12 passages generated from fixed prompts and left unedited. Mean AI score: 35%. Median: 34%. Range: 25–49%. Unedited model output scored higher than human writing, but not by a wide margin, and the ranges overlap heavily.

We also counted AI-tell frequency: 3 per 100 words in the human sample versus 1 per 100 words in the unedited model output. That result surprised us. The human writers used more of the so-called tells — dashes, triads, hedges — than the model did, because the human writers were writing in a house style that leans on those devices.

Which is the point. A tell is a tell only when it is out of character for the writer. If your drafts have always run on em dashes, the dash is not evidence; it is your voice. If they have not, and the model introduced them, that is a different story.

How do I tell a tell from a legitimate style choice?

Ask three questions about any habit you find in your draft.

1. Is it mine? Look at something you wrote before you used a model. If the habit is there, keep it. If it is not, it arrived with the draft and probably should leave with it.
2. Is it doing work? An em dash that replaces a comma you would have misread is doing work. A dash thrown in because the sentence felt long is not.
3. Is it uniform? One triad in a piece is a flourish. Four triads in four paragraphs is a pattern. Variety is the cheapest signal of a human hand, and it costs nothing to add.

If the answer to all three is “it’s mine, it’s working, and I vary it,” you have a style, not a tell. Do not strip it out to appease a detector — you will flatten your writing and probably not move the score much anyway, given the overlap in the numbers above.

How do I replace each habit with my own construction?

Replacement is easier than deletion. For each habit, here is a swap you can make without rewriting the sentence around it.

Em dash → pick the right pause. If the aside is a genuine interruption, keep the dash. If it is an explanation, use a colon. If it is a mild qualifier, use a comma. If it is a full second thought, use a full stop. The goal is not zero dashes; it is dashes that were chosen.

Triad → two or four. Cut the weakest item. “Clear, concise, and compelling” becomes “clear and compelling” — and now the sentence claims something instead of listing. If you genuinely have four items, use four; the ear will not expect the fourth, which is exactly what you want.

Hedge stack → one hedge, placed where the uncertainty is. “May potentially suggest” becomes “suggests.” If you need to soften it, do it once: “The data suggests.” One hedge reads as care. Two read as evasion.

Bolded label → bolded claim. Rewrite the bold so it says something the following sentence cannot. If the bold is a noun, it is a label. If it is a sentence, it is an argument.

Worked example. Here is a passage in the unedited style, then the same passage after a surface pass.

Before (illustration):

The results—which may potentially indicate a broader trend—suggest that latency, throughput, and reliability are all important. Scalability. Scalability matters because systems need to grow.

After:

The results point to one thing: latency is the constraint. Throughput and reliability follow from it. Scalability is what you get when the first problem is solved.

The second version is shorter, makes a claim, and uses one colon and one full stop instead of two dashes. The ideas are unchanged. Only the surface moved.

What pass should I run over my own draft?

Run this before anyone else reads the draft. It takes about ten minutes on a normal-length article.

1. Read it aloud. Every place your voice trips — a clause that is longer than your breath, a hedge you would not say — is a candidate. Mark it, do not fix it yet.
2. Count the em dashes. If you have more than a handful in a piece of a few hundred words, and they cluster in one syntactic position, you have the tell. Replace the ones that are not doing work.
3. Hunt the triads. Search for the word “and” in list contexts. Any three-item list where the third item is a synonym for one of the first two gets cut.
4. Scan for hedge pairs. “May potentially,” “could arguably,” “might possibly” — collapse each to one word.
5. Check the bolded intros. If the bold is a noun, rewrite it as a claim or delete it.
6. Re-score. Run the draft through the detector again and compare sentence-level scores before and after. You are looking for the paragraphs that moved, not the overall number.

A sentence-level view matters here because surface habits cluster. A paragraph with three dashes and a triad is worth rewriting; a paragraph with one of each is probably fine. The draft editor is built for exactly this — you see which sentences carry the signal, edit the ones that do, and leave the rest alone.

What if I write this way on purpose?

Then keep it. This is the honest part of the article, and it is the part most editing guides skip.

If your house style runs on em dashes, your triads are a rhetorical habit you have had for a decade, and your bolded intros are a formatting convention your publication requires, then stripping them out to chase a detector score is the wrong trade. You will lose the thing that makes the writing yours, and — based on our own numbers, where human writing and unedited model output overlap heavily — you probably will not move the score much.

What you should do instead is document it. If a client, an editor, or a committee flags your draft, the strongest defence is not a lower score. It is a version history that shows the draft evolving, plus a style guide that shows the habits were always yours. Detector output is a probability estimate, not proof of authorship, and a writer who can show their working is in a much better position than a writer who cannot.

For the broader question of how these signals are produced and what they mean, how AI content detectors work is the reference to read next.

Ready to see which sentences read machine-made?

The surface habits in this article are the visible layer. Underneath them, the signal that actually moves a detector score is sentence-level: rhythm, predictability, and how uniform the construction is across a paragraph. You cannot eyeball that reliably, which is why the scoring step matters.

Start with the AI detector and read the sentence highlights before you change a word. If the flagged sentences cluster around the habits you found in your own pass, you have confirmed the diagnosis. If they do not, you have learned something more useful than any list of tells.

When you are ready to fix them, the draft editor runs the rewrite loop: score the text, rewrite only the flagged sentences, keep a rewrite only when semantic similarity says the meaning survived, then re-score and show you before and after. On the example paragraph on our site, a 92% AI score becomes 4% after that pass. Across our own tests, the site reports an average AI-score reduction of 60%+ after humanizing, with a best case of 71% — our own predicted scores, not a verdict from any third-party detector.

The free tier gives you 200 words on signup with no credit card, which is enough to test the loop on one flagged paragraph before you commit. After that it is pay-as-you-go: Starter $20.99 for 1,000 words, Standard $44.99 for 3,000, Pro $99.99 for 10,000, Max $199.99 for 30,000. Credits never expire and there is no subscription. You can see the full breakdown on pricing, or create an account and paste in the paragraph you are least sure about.

One last thing, and it is the part that matters. No surface pass, no detector, and no humanizer will make a draft yours if the ideas, data, and conclusions are not. Fix the dashes, cut the triads, collapse the hedges — then make sure the argument underneath them is something you would defend in a room. That is the only version of this that holds up.