Why AI Writes in Em Dashes (and Why I Banned Them)

There is a file on my machine that says, among other things: prefer commas and periods to em dashes.

That line has been there for about a year. It went in during 2025, when readers started scanning for dashes, and a piece with six of them in the first three paragraphs got read as machine output before anyone reached the argument. The rule was a defense. I had nothing against the dash itself.

The rest of the file is more useful than that line. It took teaching other people to build their own before I understood why.

Why AI uses em dashes so much

Nobody outside the labs can give you the full answer, and the labs have mostly shrugged. The best explanations I have seen come down to three things, and they probably all contribute.

The first is the training corpus. Models learn from edited, published prose, and edited prose leans on the em dash because copy editors reach for it to fix a sentence without restructuring it. The models absorbed a professional habit and applied it at a rate no single professional would.

The second is that the dash is a safe connector. A comma commits you to a grammatical relationship. A period commits you to ending the thought. A dash lets a model bolt two clauses together without deciding what one has to do with the other. When you generate a sentence one word at a time, that flexibility is worth a lot.

The third is feedback. The people rating model output during training seem to have liked the rhythm of the dash, the slight pause and swerve, and the models learned to give them more of it.

Look at the second explanation again, because it is the one that matters. The dash is what a sentence does when no one has decided how its two halves relate. A judgment call never got made, and the mark sits where the decision should have been.

Whether the model “decided” anything is beside the point. Some of this is training dynamics with no decision behind it at all. The mechanism is about the writer, and what the mark asks the writer to do next.

Once I saw that, every other tell in the file looked the same.

The signs of AI writing, read as missing decisions

Here is what the rest of the file bans, and the decision each tell stands in for. The “before” sentences are the kind of thing a model handed me. The “after” is what I kept once I made the call.

Every paragraph ends on a polished takeaway. This is the most reliable tell I know, and detectors barely see it. The missing decision is what the paragraph was for. A model does not know whether a paragraph is giving context, making a claim, or landing a conclusion, so it closes every one with a quotable sentence to be safe.

In my drafts, once I decided what a paragraph was doing, most of them stopped needing a closing line.

Not X, but Y. “It’s not about speed, it’s about judgment.” “This isn’t a tool problem, it’s a trust problem.” The construction performs a correction, and the missing decision is what you are claiming. Nobody was arguing X. The sentence exists because Y was never stated on its own and defended.

Metaphors that do not lead anywhere. A garden, a compass, a mirror, and then back to the abstraction as if the image never happened. Nobody decided what the image was for. A metaphor is a claim about how something behaves, and if you cannot say what follows from it, you have not settled what it means.

Signs of AI writing: the full list

Those four get the most space because they do the most damage. The rest of the file, with the decision each one stands in for:

Eleven tells. One question underneath all of them, which is what was decided here, and by whom.

What the tells are telling you

I did not arrive at any of this by introspecting. Every line in the file came from arguing with a draft. Not what I meant. Close, but it sounds composed. Nobody wrote this one. Each objection was a standard I already held and had never written down, and the draft was the first thing that forced me to.

That is the part I now teach. When I work with product managers and designers on their own instruction sets, the fastest section to build is the one that lists what they always reject, and the reason it is fast is that the rejections are sitting right there in their edit history. The hard part is turning each rejection into a decision. “Fewer em dashes” is a preference. “Decide how the two clauses relate, then punctuate” is a rule the model can follow, and it is also a rule you can follow, which is the point. Read that way, the tells stop being a list of things AI does. They are a map of the decisions you have been leaving to it.

I wrote about where that map ends up in The New Résumé on Substack: how you work used to be a claim you made about yourself, and it is becoming a trace you leave. The rules file is that trace. And if you want to build one from your own drafts rather than borrow mine, that is what I teach on Maven.

One last thing, since the em dash question usually comes from someone worried about a detector. Scrubbing the tells will lower a score. It will not make a piece worth reading. I built isthisoriginal.com as the other instrument: instead of guessing who typed a piece, it maps where the argument departs from what search and models already say. The argument for why that is the better question is in The Wrong Game Comes for Writing.

Delete the dashes if you want. The decisions underneath them are the part that matters.