← Back to blog

AI & Content · November 18, 2026 · 8 min read

Why AI Writing Uses So Many Em Dashes (And How to Stop)

AI reaches for the em dash because it is the safest way to join two ideas without stating how they relate. Deleting the character alone leaves five other tells intact.

By FluxWriter

Why AI Writing Uses So Many Em Dashes (And How to Stop)

Why AI uses so many em dashes comes down to one thing: the dash is the safest way to join two ideas without stating how they relate. A model predicting the most probable next few words reaches for that move constantly, because the polished, edited English it learned from reaches for it too. Deleting the character afterwards does not make a draft sound human, only like machine writing with the dashes taken out. This walks through where the habit comes from, the five tells that travel with it, and the edit pass that changes how a page actually reads.


The Dash Is a Probability, Not a Preference

A language model does not choose punctuation the way you do. It predicts the next chunk of text from patterns in what it was trained on, then repeats that prediction a few thousand times. The em dash is unusually useful in that job. It can stand in for a comma, a colon, a semicolon, a bracket or a full stop, and it is almost never technically wrong. When two clauses could relate in four different ways, the dash means the model never has to pick one.

The second half of the explanation is what counts as good writing in the source material. Books, magazine features, newsletters and well-edited landing pages all use the dash for emphasis and aside. Models are also tuned to sound helpful and composed, and composure at sentence level mostly means balance and rhythm. The dash buys both in one keystroke. Treat that as a working explanation rather than a documented setting, because nobody outside the labs can point to the line that causes it.

So the dash is not really the problem. It is the visible edge of one habit: the model reaching for its highest-probability version of "well written". Every other tell below comes from the same place, which is why removing a single character changes so little.

Five Tells Travel With the Dash

Paired-clause rhythm. Two balanced halves, sentence after sentence. The first half sets up, the second completes. Four of those in a row and the prose sounds like a metronome, whatever punctuation sits in the middle of them.

The triad. Three items where two would do. Three adjectives, three examples, three benefits. Human writers use threes as well. Models use three as the default length for any list, including lists with only two real members.

The hedged opener. "It's worth noting that." "Importantly." "That said." "In many cases." Each one opens a sentence without adding anything to it, and they cluster at the front of paragraphs.

The mirrored negation. "It isn't about speed, it's about consistency." One idea is rejected so the next sounds earned. Use it twice in a post and a reader feels the shape even if they cannot name it.

Flat variance. Every paragraph three sentences, every sentence about the same length. Human drafts are lumpy. A four-word sentence lands next to a forty-word one, because the writer got impatient, or got interested.

None of these are errors. Each is defensible alone. Density is what gives a page away, and the dash is simply the easiest member of the set to find with ctrl-F.

Deleting the Character Leaves the Sentence Intact

Swap every dash for a comma and you get comma splices, which read as sloppy rather than human. Swap them for semicolons and the page turns stiff, because the semicolon carries a formality the dash does not. Swap them for brackets and every aside gets demoted to a whisper. In all three cases the clause count is unchanged, the balance is unchanged, and the relationship between the two halves is still unstated.

That is the part worth understanding before you run a find-and-replace across a year of posts. The dash was never carrying the machine-written quality on its own. It was marking the spot where the sentence declined to make a decision. Take out the marker and the decision is still missing, only now the reader cannot see where.

A page that has had its dashes stripped and nothing else changed reads slightly worse than the original. The rhythm survives, the hedging survives, and the punctuation is now doing a job it was not built for.

Naming the Relationship Is the Actual Fix

The repair is to decide what the two halves have to do with each other and say so. Usually one of four words does it: because, so, but, after. If none of them fit, the two ideas probably belong in separate sentences.

Take an illustrative example. "The post ranked for six months. Then the traffic went away." Both halves are true, the link between them is implied, and the sentence sounds finished without being informative. Rewritten with a decision in it: "The post ranked for six months. Traffic then fell by about half in three weeks, after two competitors published longer versions of the same guide." Longer, less elegant, and far more useful to a reader, because now it says why.

Do that ten times in a draft and something else happens. Sentence length starts varying on its own, because "because" clauses run long and split sentences run short. The triads thin out, since a list of three is harder to justify once each item needs a stated reason. The habit you are actually breaking is the withheld relationship, and the punctuation follows it out of the door.

A Full Pass Costs About Twelve Minutes

The times below are illustrative rather than measured, based on roughly how long each check takes on a 1,500-word draft. Your own numbers will differ with how heavily the draft was prompted.

Tell How to find it The fix Minutes
Dash-joined clauses Search the character, count per 1,000 words Split the sentence, or name the link 4
Triads Scan for any list of exactly three Cut to two, or extend to five 2
Hedged openers Search "worth noting", "Importantly", "That said" Delete the opener, keep the sentence 2
Mirrored negation Search "isn't about" and "not just" State the positive claim only 2
Flat rhythm Read the first sentence of each paragraph aloud Break one paragraph, merge another 2

Twelve minutes on a 1,500-word post is roughly one minute per 125 words. On four posts a month that is about 48 minutes, which is less than most people spend choosing a featured image. The pass is cheap because none of it requires rewriting the argument, only the joins.

Detectors Are the Wrong Thing to Optimise For

AI detectors produce false positives on human writing and false negatives on machine writing, and none of them can see how a document was made. A score is an opinion about style, generated by another model with the same blind spots. Rewriting a page until a detector is satisfied optimises for that detector, not for a reader.

Search engines are a separate question, and the published position from Google is that it aims to reward helpful, original content however it is produced, while its spam policies target content generated at scale primarily to manipulate rankings. That is a statement about intent and quality, not about punctuation. Nobody can promise you how any individual page will perform.

The thing worth optimising for is whether the page contains something only you could have written. A real price, a real process, a number out of your own account, an objection a customer actually raised last week. A draft with three of those in it survives a heavy dash habit. A draft with none of them fails whether or not it reads as machine-written.

Prompt Instructions Get You Part of the Way

"Never use em dashes" works on the character, mostly, and models drift back over long outputs. It also leaves the habit intact: cleaner punctuation, same shape. Ban lists are a floor, not a style.

Structural instructions do more. Ask for the relationship between clauses to be stated. Ask for sentence length to vary. Ask for one specific number in each section. Ask that no list contain three items unless there are exactly three. These are harder for a model to satisfy while still producing its default paragraph, which is the point.

The most reliable input is usually a sample. Paste 300 words of your own writing and ask for that register, rather than describing it with adjectives. Models match a sample far more closely than they follow an instruction like "sound conversational".

FAQ

Does an em dash mean a page will be flagged as AI-written?

Not on its own. Plenty of experienced human writers use the dash heavily, and detector scores are unreliable in both directions. What draws suspicion is several tells at once: dashes, triads, hedged openers and uniform paragraph length in the same 1,500 words.

Will Google penalise a post because it reads as AI-written?

Google's published position is that it rewards helpful, original content however it is produced, and that its spam policies target content produced at scale mainly to manipulate rankings. So the question to ask about any post is whether it is genuinely useful and original, not which tool typed it.

Is it worth editing old posts for this?

Only the ones that matter. Sort by traffic or by commercial value, take the top ten, and spend the twelve minutes each. Pages nobody reads are not worth the pass, and rewriting a hundred posts to remove punctuation is a week you could spend publishing five new ones.

The Practical Takeaway

Pick your three most recent published posts and search each one for the em dash. Count them per 1,000 words so the posts are comparable. Then take the five sentences with the highest dash density and rewrite them by naming the relationship: because, so, but, after, or a full stop. Read all three posts aloud afterwards and note where your voice flattens, since that is where the rhythm is still machine-shaped. The first action is small and takes two minutes: open the post you published most recently and count its dashes.

The sample advice carries over to FluxWriter as Brand Voice training, up to five of your own posts per site on every plan, which works on the register of a draft rather than only on its punctuation.



← All posts