The em dash is innocent: AI-detection folklore, debunked

Sometime in the last few years, the em dash got arrested for a crime it didn't commit. Social media decided that the long dash — this one — was the signature of machine writing, and suddenly writers who had used it for decades were being accused of outsourcing their own voice. Emily Dickinson built a body of work out of dashes. Tristram Shandy is practically punctuated with them. The em dash is not evidence; it is punctuation.

The same goes for the rest of the folk taxonomy: “delve,” “tapestry,” “testament to,” starting a reply with “Great question,” tidy bulleted lists, the phrase “in today's fast-paced world.” Each of these became a meme because it was genuinely overrepresented in early chatbot output. But there are two problems with treating them as a detector, and both problems get worse every year.

Problem one: the tells are trivially removable

Every surface marker — a word, a punctuation habit, a greeting — can be edited out in seconds, by the model itself. “Rewrite that without em dashes and don't use the word delve” is not a jailbreak; it's a normal instruction that current systems follow perfectly. Any tell that can be removed by asking is not a tell at all once the writer has any motivation to hide. The only signals worth training your ear on are the ones that are expensive to remove — structural habits, which we catalogue in our guide to the seven real tells.

Problem two: the false positives are people

The folklore has a body count, and it's human. Writers with formal educations, non-native English speakers taught the classic five-paragraph structure, autistic writers with precise mechanical styles, and anyone who simply likes semicolons now get accused of being machines on the regular. Teachers run essays through detection tools that have documented false-positive rates high enough to indict several innocent students per classroom, and the accusation is nearly impossible to disprove — how do you demonstrate that you wrote something the old-fashioned way? The em-dash panic isn't just wrong; it's a small ongoing injustice.

There is a deeper irony here. The stylistic habits people flag as “AI-sounding” — balance, hedging, signposting, polite structure — are the habits of edited prose. The models write that way because they were trained on our most groomed text: published books, journalism, documentation. Accusing a careful writer of sounding like a machine is accusing them of sounding like the canon.

What actually holds up

If the surface tells are folklore, what's left? Structure and stakes. Machine text under-commits: it hedges toward the average opinion, keeps its emotional manners perfect, and builds paragraphs that resolve too cleanly. It also under-specifies: real writers anchor claims to boring particulars — a date, a price, a street name, the model number of the radio — because they are remembering, not generating. When a passage is all texture and no ballast, that's a real signal. When it merely contains a long dash, that's a Tuesday.

The honest summary: you cannot reliably detect machine text by scanning for tokens, and neither can the commercial detectors, whose independent evaluations remain grim. What you can do is develop taste for the structural patterns — and the only way to develop taste is reps. That is, frankly, why the daily puzzle exists. Guessing, being wrong, and reading the tell afterward trains a detector no tool currently matches: an attentive human who knows what the failure modes look like.