Steno / Em dash

The em dash is not the tell any more

A lot of people are still hunting for a signal that stopped working. Worse, some of them are editing perfectly good punctuation out of their own writing to avoid being accused.

Updated 6 August 2026 · About five minutes

Around 2023 the em dash became shorthand for machine writing. ChatGPT used it heavily, most people used it rarely, and the gap was wide enough that a page full of them was a reasonable thing to be suspicious about. The heuristic spread until it reached people who had never thought about punctuation before, and writers who had used the em dash their whole careers started getting accused.

The gap has closed.

What the 2026 analysis found

The Economist ran a comparison in 2026 between its own articles and versions of the same material generated by ChatGPT, Claude, Gemini and Grok, alongside newspaper writing from other outlets and excerpts from novels published between 1950 and 2022. The set came to roughly 55,940 sentences and 1.2 million words.

On the em dash specifically, the finding was awkward for the popular heuristic. Of the models tested, only Claude used em dashes more often than human writers. ChatGPT, the model that started the whole panic, now uses them less often than the humans in the sample.

The clearer separators were verbosity and light use of punctuation. Machine prose said more words to carry the same content, and used fewer marks to break it up.

That second point is the one worth sitting with. The problem was never a specific mark. It was that the sentences were long, even and thinly punctuated, and the em dash happened to be the most visible symptom.

Why every vocabulary signal decays

Each of these follows the same arc. A pattern gets identified, the identification spreads, model providers see the complaints, and behaviour shifts. Delve went the same way. So did tapestry. So will whatever replaces them.

There is a second force pushing in the same direction. Human writers now avoid the flagged patterns deliberately, so the human baseline moves too. When both distributions move, the gap you were measuring is gone.

Structural habits are stickier. Sentence rhythm, forced triplets, participial clauses bolted to the end of sentences, conclusions stated on the reader's behalf. These come from how the model composes rather than from which word it picks, and they do not shift because someone posted a thread about them.

What to look at instead

If you are reading rather than writing, four things carry more information than punctuation.

• Rhythm. Sentences landing in the same length band for a whole page. Paragraphs of identical size, one after another. Human prose is lumpy because humans get bored.
• Word count against content. Take a paragraph and write down what it actually asserts. If three sentences produce one fact, that is the verbosity signal.
• Triplets. Groups of three where the third item is visibly weaker than the other two.
• Endings. Sections that close by restating the point in warmer language rather than stopping.

None of these is proof. All of them together, in a document with no specifics and no opinions, is a strong smell.

If you write with an em dash, keep writing with one

The honest position is that a punctuation mark is not the problem, and removing it from your own prose to avoid suspicion is a bad trade. You lose a useful mark and you fix nothing structural.

Steno tells the model to avoid em dashes, and the reason is narrower than the popular one. In machine output the em dash tends to arrive as a stacking device: a clause bolted on, then another, until a sentence is carrying three ideas that should have been two sentences. Banning the mark forces the split. That is a rule for a model. It is not a rule for you.

For the habits that have not decayed, the full list is here. If you want them applied automatically, the instructions go in custom instructions or a Claude style.