The em dash AI debate follows a pattern: someone posts a screenshot, a thread fills with verdicts, and a writer deletes a mark they have used since school.
Punctuation is a weak signal that costs you ten seconds to check and never decides whether a draft is worth publishing.
I run this blog with an agent and a set of deterministic rules, so I have measurements to work from. When we rigged a draft to sound exactly like us while carrying nothing of our own, the rigged draft scored 1.89 out of 2 on brand voice and 0.04 on first-hand facts, against 1.91 and 0.63 for the real one.
Below I also explain why models reach for the mark so often, what came out of a pass over our own 37 tool descriptions, and the two checks that stop my drafts.
- An em dash is a weak signal worth ten seconds
- Models picked up the mark from edited human prose
- Taking the dashes out of 37 tool descriptions was a voice decision
- A draft can sound on voice and still carry nothing
- The checks that decide a draft read links and evidence
- What to check before you blame the punctuation
An em dash is a weak signal worth ten seconds
An em dash tells you almost nothing about who wrote a draft, so the only sane use for it is a ten-second prompt to read the text more closely. DataForSEO reported 3,600 US searches a month for "em dash ai" when seodraft measured it on September 27, 2026, and on October 3, 2026 Google answered it with an AI Overview of its own.
In that October 3 read, the first results were Reddit threads and opinion essays arguing over whether the mark is proof of manipulation. They are worth reading as culture, and they are weak ground for a publishing decision.
The people who write the cleanest long dashes include copy editors and anyone who learned typography before 2022. Accuse on punctuation and you will be wrong about them, with no way to measure how often.
If you want the broader list of surface tells, we keep a free one at the seodraft AI slop checker.
Models picked up the mark from edited human prose
ChatGPT does use em dashes heavily, and the most defensible explanation is that the edited English it learned from is full of them. Nobody outside the labs can prove the cause, and I am not going to pretend otherwise.
Scale is what turned a house habit into a public tic. On August 31, 2026, the European Commission designated ChatGPT a Very Large Online Search Engine, and OpenAI reported about 159.1 million average monthly active recipients of ChatGPT search in the EU for the six months ending March 31, 2026. One model's habit reaches that many readers.
The mark itself is old. What changed is the volume of text arriving in front of readers every day, which is why a punctuation habit became a thing people argue about.
Taking the dashes out of 37 tool descriptions was a voice decision
The dashes came out of our tool descriptions because our voice is short sentences and plain words, and the work note for that pass reads: Chorch rewrote the text of all 37 tools so they sound like the product, with no internal jargon, no dashes and no shouting caps.
A long dash invites a subordinate clause, and a tool description that an agent has to parse is a bad place for one. Those descriptions carry more weight since seodraft killed most of its web UI: the agent reads them to decide what to call.
Nobody reading those descriptions can tell whether a model drafted the first version. The edit was about the reader, and the reader does not care which keyboard produced the draft.
Removing every dash to dodge an accusation is a different job from removing one because the sentence reads better. I did the second, and the first never came up.
A draft can sound on voice and still carry nothing
Voice and substance move independently, and we measured how far apart they can sit. We rigged a draft to imitate our style while stripping out everything only we know, and the rigged draft scored 1.89 out of 2 on brand voice and 0.04 on first-hand facts, against 1.91 and 0.63 for the real one.
A style score of 1.89 against 1.91 is a rounding error, while 0.04 against 0.63 is the entire article.
That gap is what punctuation hunting misses. A draft with every dash removed can still score 0.04 on anything first-hand.
The habits that put substance back in are the ones we wrote up in our post on getting an agent to write without sounding generic.
The checks that decide a draft read links and evidence
The rules that stop our drafts never look at punctuation, and the failures they do catch are all checkable without judgment. The current workflow saves the brief before the body and requires every evidence item in the brief to reach the article, so a fact I planned and then dropped stops the draft before a reader ever sees it.
The second check counts links and length. My seodraft profile requires at least two internal links and targets 900 to 1,500 words when the topic justifies the length, and a link to a product page never counts toward that minimum.
Each of those has a yes or no answer. A link resolves to a published post or it does not. An evidence line appears in the finished body or it stayed a promise the brief made.
Punctuation fails that test. There is no count of dashes at which a draft becomes defective, so there is no threshold to write into a rule, which is why it stays a curiosity.
The failure that does deserve a hard check is invented fact, and how to detect AI hallucinations shows the check we use.
What to check before you blame the punctuation
Decide a draft on what only you could have written, and let the punctuation be whatever the sentence needs. The fast version of that check takes one pass: open the draft, mark every number, name and first-hand detail, then count the marks. Three of them in 1,200 words is your real problem, whatever the dashes are doing.
Then fix the input. A draft that comes back empty needs facts from you, and editing cannot put them there. The rest of the pre-publish pass is in our AI content quality checklist.
If you want that check to run on every draft, seodraft does it with your own agent (Claude, Cursor, or ChatGPT on Business, Enterprise or Edu): your agent writes, the rules decide, you approve. You can start at seodraft.app/sign-up.
