Make yours

The em dash question

The em dash became the internet's favourite proof that a machine wrote something. It is the weakest of the tells, it has a false-positive problem nobody mentions, and the best account of why models overuse it opens by admitting the cause is unknown.

This page collects what is actually established, which is less than you would expect, and one genuinely useful thing: the spacing version of the tell is wrong about an entire national publishing tradition.

It sits under the AI writing tells, and if you want a punctuation rule enforced rather than argued about, eleven questions produce a file that states yours.

Is an em dash proof of anything?

No, and the catalogue everybody cites for it says so directly. Wikipedia's guidance on machine-writing signs states that the em dash is most useful taken in combination with other indicators, not by itself, and notes it is far more common on discussion pages than in article text, where people write more loosely.

It also describes what the models actually do with the mark, which is more specific than the mark itself: they tend to use it in a formulaic way, punching up clauses to give sentences an emphatic, sales-like rhythm. That habit is the tell. The character is just where it lands.

Why does the spacing version of the tell misfire?

This is the part worth knowing, and it is the reason a British writer gets accused unfairly.

The refined version of the tell is not the em dash itself but a spaced one. Wikipedia's guidance says machine-written em dashes are usually surrounded by spaces, contrary to common typographic guidelines. That is true of American house style: Chicago and the APA manual both set the em dash closed, with no spaces around it.

It is not true generally, and two large populations write spaced dashes on purpose.

Who sets dashes spaced, and who does not
StylePracticeConsequence for the tell
Chicago Manual of StyleEm dash, closed up, no spacesThe tell works here
APA Publication ManualEm dash, closed upThe tell works here
Oxford Guide to StyleEm dash, closed up, like most US publishersThe tell works here
AP StylebookEm dash set open, with spaces around itThe tell misfires on anyone following AP
Penguin, Cambridge University Press, RoutledgeSpaced en dash as house style in the United KingdomThe tell misfires on a national publishing tradition

So a spaced dash identifies a machine, an AP-trained journalist, or somebody who learned to write from British books. The Oxford Guide to Style acknowledges the split itself: it records that the spaced en dash is used by other British publishers while Oxford University Press, like most US publishers, does not.

If you are being told your writing looks machine-made because of spaced dashes, that is the answer, and it is checkable in ten seconds.

Why do models use so many of them?

Nobody knows, and the most careful published attempt says so in its first paragraph. Sean Goedecke's account, from October 2025, opens by conceding it is strange that we do not really know, and labels his own explanation as still largely based on speculation.

His hypothesis is that the high-quality book data these models were trained on skews towards late-nineteenth and early-twentieth-century print, which uses around thirty per cent more em dashes than contemporary English prose. He also reports a version step-change: GPT-3.5 did not use them much, GPT-4o used roughly ten times more, and GPT-4.1 more again.

What makes the piece worth reading is the theory he rejects rather than the one he proposes. The popular alternative is that the habit comes from the dialect of the people doing the reinforcement work. He tested it and found em dashes made up 0.022 per cent of words in Nigerian English data against 0.25 to 0.275 per cent in general English, which points the opposite way.

A second account, from Maria Sukhareva in July 2025, proposes token economy: an em dash is a single token where the alternatives cost more. She ran a hundred and fifty story prompts across three model versions and found the same roughly tenfold jump. Treat the two as one line of thinking rather than independent confirmation, because the first cites the second.

Can you tell a model to stop?

Historically this was unusually hard, and the difficulty is documented rather than folklore. An OpenAI developer forum thread opened in July 2025 collects users failing to suppress it, with reports that the same instruction worked through the API and not in the chat interface.

Whether that still holds on the model you use today is not something this site will tell you, because we publish no measurements of model behaviour. Claims in this area go stale within a release, and a dated table would be wrong by the time you read it.

What is worth knowing is where to put the instruction. A punctuation prohibition is exactly the kind of rule that survives: short, mechanical, and impossible to half-satisfy. Which instruction shapes survive covers why that shape works when an adjective does not.

What if you want to keep using them?

Then keep using them. The em dash is a legitimate mark with a long history and no amount of model output changes that. What has changed is that readers now notice, so the practical questions are consistency and density rather than permission.

Pick a convention and hold it: em dash closed up, or spaced en dash, but not both in one document. Inconsistent dash style is a genuine sign of assembled text, and unlike the mark itself it is something a checker can catch.

This site takes the other option and forbids them outright, which is a house rule rather than a recommendation. It is stated in the file the tool generates, because a rule you write down is enforceable and a rule you remember is not.

What this page will not claim

Where these come from

Every claim above is quoted from one of these, and each was read on the date beside it. If one of them has changed since, the page is wrong and we would like to know.

A punctuation rule is worth more written down than argued about: state yours once and hand it to every assistant.