0.022% of the words in a dataset of nigerian english were em dashes. that is about one in every 4,500 words. english prose in general runs at somewhere between 0.25% and 0.275%, roughly one in 400, so the nigerian sample uses the mark about a tenth as often as english at large.
sean goedecke, a software engineer, counted that in october 2025 while working on a question with no official answer: why does ai use so many em dashes? he went in testing the same theory that had been used to explain the chatbot vocabulary, that the habit came from the workers rating the model. the count killed it for the dash. he then built a better one, and called it, in his own words, still largely based on speculation.
1860, the peak
a study of punctuation in english corpora found dash use climbing after 1750 and peaking at about 0.35% of words in 1860, then falling until the 1950s before settling in the 0.25% to 0.275% band. victorian prose ran on the mark. moby-dick contains 1,728 of them.
november 2022 to july 2024
gpt-3.5, the model inside the first chatgpt, did not use em dashes. by goedecke's count gpt-4o used about ten times as many as its predecessor, and gpt-4.1 used more again. models from anthropic and google use them too, as did the chinese open models of that period.
something changed in the training mix between those releases, and goedecke's candidate is books. early models trained on scraped web text and pirated book collections, which skew toward recent popular titles because that is what people download. once labs wanted more high-quality text they started scanning print books themselves, and court filings showed anthropic beginning that work in february 2024. scanned print collections reach back into the public domain, which is the 1800s and early 1900s, and by his estimate those books carry about 30% more em dashes than contemporary prose.
two popular explanations do not survive his reading. one says the dash is token-efficient, but in the common reveal pattern a comma is just as brief. the other says the dash keeps the model's options open, but other punctuation marks are about as flexible, and gpt-3.5 managed without it.
his summary of the mechanism is plain. models learned english from texts that were full of them, which is also why prompting them out of the habit is so hard.
he lists the holes himself. nobody outside openai knows when it started scanning. models now also train on text written by earlier models, which carries the dash forward on its own. and if the books did it, it is odd that the prose sounds modern while only the punctuation went victorian.
14 november 2025
two days after gpt-5.1 shipped, sam altman posted what he called a small-but-happy win. "if you tell chatgpt not to use em-dashes in your custom instructions, it finally does what it's supposed to do!"
read the sentence for what it says. the default did not change. the official chatgpt account said the same on threads, that the model was better at skipping the dash when told to through custom instructions. ars technica's reading was that openai had tuned the new model to weight those instructions more heavily, and custom instructions are text appended to the prompt before each conversation starts. what got fixed was the model's obedience to an instruction, and the dash stays its first choice whenever nobody asks.
taking them out of a draft
if a model wrote it, the dashes came from somewhere older than the model, and they come out the way they went in, one at a time. each one is doing the job of some other mark. where it does a comma's job, it becomes a comma. where it sets off an aside, it becomes a pair of brackets. where it sets up a reveal, the "it is this, then it is that" shape, the sentence wants rewriting from the start, because the punctuation was holding up a trick.
writers who used the mark long before any of this have started cutting it from their own prose for fear of being taken for a model, which is the strangest outcome of the lot. melville is safe. moby-dick still has all 1,728.