Why Does AI Use So Many Em Dashes?

If you have read much AI-generated text, you have noticed the dashes. Not one or two — a steady drumbeat of them, several per paragraph, until every sentence has an aside bolted onto it. It is now the single most recognisable signature of machine writing, and there is a concrete reason for it.

The training data explanation

Language models learn punctuation frequency from what they are trained on, and that corpus is weighted heavily toward edited, published prose: journalism, magazine features, essays, nonfiction books. Those genres love the em dash. It is the house punctuation of the modern long-form article, where a writer needs to fold context into a sentence without losing momentum.

Ordinary writing — email, documentation, Slack, notes — barely uses it. But that register is underrepresented in training data relative to how much of it humans actually produce. The model absorbs the em dash rate of The Atlantic and then applies it to your product changelog.

The structural explanation

There is a second, subtler reason. A model generates one token at a time without a committed plan for the rest of the sentence. The em dash is uniquely forgiving punctuation: it can attach almost any continuation to almost any partial sentence and stay grammatical. A colon demands a specific relationship. A semicolon demands two independent clauses. Parentheses demand you close them. The em dash demands nothing.

So when a model has more to say and no committed syntactic structure to say it in, the em dash is the lowest-risk next token. Repeat that a few thousand times and you get the rhythm everyone recognises.

How much is too much?

Rough calibration from published prose: a heavy human user of the em dash runs about 2–5 per thousand words. Most writers use fewer. Unprompted model output frequently lands at 15–30 per thousand words — five times the rate of a dash-loving human columnist.

That gap is why editors spot it instantly. It is not that a dash looks robotic; it is that the density is outside the human range.

Does it prove text is AI-generated?

No — and be careful with anyone who claims otherwise. Em dashes are ordinary English punctuation, and a real writer with a taste for them will trip the same intuition. Plenty of people have been wrongly accused on exactly this evidence. Density is a signal, not proof, and no responsible detector treats a punctuation mark as conclusive.

Fixing it

In the prompt. Ask directly: "do not use em dashes; use commas, full stops or parentheses instead." Models comply reasonably well, though compliance decays over a long conversation, so repeat the instruction if the output drifts. Better than repeating yourself: install it once as a Claude Skill or custom instruction so it applies to every conversation automatically.

After the fact. Run the text through an em dash remover. Replacing each dash with a plain hyphen keeps the sentence rhythm intact; removing it entirely reads more naturally but is worth a reread, because the dash occasionally carries grammatical weight that a bare space does not.

By rewriting. The best fix, when you have time: most model em dashes are joining a main clause to an afterthought. Split it into two sentences. The prose gets shorter and better, which is generally what the dash was hiding.

The other tells

The em dash rarely travels alone. Curly quotes, the single-character ellipsis (…), non-breaking spaces, and occasionally genuine zero-width characters all come through the same copy-paste. The AI text cleaner handles the whole set in one pass, and the invisible character detector shows you what was actually in there.