AI detectors made the em dash a tell. Writers are dropping it

Sometime in the last year the em dash stopped being punctuation and became evidence. Bloomberg's Morgan Meaker reports that writers are now editing around the furniture of their own prose: stripping em dashes, avoiding the "not X, but Y" construction, steering clear of a short list of words that read as machine-issued, and, in some cases, planting typos on purpose to look more human. Bloomberg calls the em dash the most infamous AI tell of the moment.

That would be a curiosity about writers' nerves if the verdict weren't so close to the page. Substack wired Pangram, the most accurate AI-text detector on the market by independent testing, directly into its reader this summer: any subscriber can scan a post, note, comment or reply over 100 characters and see what share of it the model thinks a machine wrote, TechCrunch reported at launch. Authors can turn detection off per post and can write a "How I make this" statement that appears beside the score. They cannot make the scoreboard go away.

So a tool that was sold as a reader-protection feature has become a style regulator, and the people being regulated are adjusting their sentences to satisfy it. That is the part worth thinking about, because the adjustment cannot work for long.

The detector is not looking for the em dash

Pangram does not hunt punctuation. It is a classifier trained on pairs: a human document and a machine imitation of the same document, tens of millions of them, and it learns the difference in texture, as El País describes in its explainer on the Pangram debate. Nothing in that pipeline encodes "em dash equals robot". The tells writers are purging are folklore that happens to correlate with the thing the model actually measures.

How well it measures is not in dispute. Pangram's technical report for version 4 reports an AUROC of 0.9916, a false-positive rate of 0.0041 percent — roughly one human-written text in 24,300 — across more than a million human English texts, and a false-negative rate of 0.34 percent on about 520,000 outputs from 26 models. Its pitch to publishers has been a false positive around one in 10,000, the figure Bloomberg's feature cites; the published number is better than the pitch.

That gap between "measures resemblance" and "identifies authorship" is where every argument about these tools goes wrong, in both directions.

Run the base rate honestly

The standard objection is that at realistic prevalence, false positives swamp true catches. For Pangram, at the document level, the arithmetic does not support that. Assume a platform with 10 million human-written posts a month and a false-positive rate of one in 24,300: about 400 wrongly flagged documents. Set a modest 1 percent AI prevalence beside it and recall near 99.7 percent, and the flagged pile is still overwhelmingly correct. Anyone arguing that the math collapses has not done the math.

The failure is elsewhere, and it is narrower and more human. First, mixed authorship: the report's own newest and least mature capability is telling you which spans came from a model, and "31 percent AI" on a human story is not a smaller version of "100 percent AI" — it is a different claim, made with less evidence. Second, unusual workflows. A novelist who dictates and edits the transcript produces the even rhythm and repeated cadence that the classifier reads as machine. A writer working in a second language, or in a formulaic register like legal or academic prose, sits closer to the model's distribution than a stylist does.

Third, and worst, interpretation. A score that answers "how much does this resemble model output" gets displayed as "who wrote this," and the person who has to disprove it is a person. Steven Mollmann's compilation of 17 studies found 49 false positives across 36,282 human-written texts, a rate of 0.14 percent, El País reports — a rounding error in aggregate, forty-nine reputations in practice.

The loop this creates

Here is the problem with editing for a classifier. Its training data is human text, and the humans writing it have now spent a year being scored. The corpus it learned "human" from is being edited by people trying to look human to it. Retrain on the corrected corpus and the em dash returns as a sign of authenticity; models pick it up again from the same public web; the tell flips. Writers are chasing a target that updates faster than they can revise.

The visible cost is a flattening. If the price of not being accused is the removal of dashes, the "not X, but Y" rhythm, and any word that got popular in 2024, what remains is prose with its fingerprints filed off — and typos inserted as proof of life, which is a strange thing to optimize for. We already have a case study in what happens when a detector's verdict outruns the institution's ability to interpret it: the Goncourt prize dropped a novel accused of being AI-written on the strength of scores nobody could adjudicate.

The detector vendors and the platforms win either way. Pangram gets distribution on the largest newsletter network in the world; Substack gets a trust feature that costs it nothing and gives readers a reason to open the post inside its app rather than in email. The writers pay, and so do readers, who get a badge that measures resemblance and prose that has been sanded toward a moving average.

What to watch

Whether the score becomes default-visible rather than something a reader has to request — that is the difference between a tool and a regime, and Substack's Android rollout will tell you. Whether Pangram's numbers survive the shift from pre-2022 text, where the false-positive rate was measured, to text written by people who know they are being scored. And whether anyone builds the disclosure that actually scales: Substack's "How I make this" note is a better primitive than the score, because it asks authors what they did instead of guessing from what they wrote.

For a reader, the practical rule has not changed: a detector's output is a probability wearing the costume of a verdict, and the way to judge a piece is to check its claims, not its punctuation. For a writer, the honest advice is the uncomfortable one: write like yourself, keep the receipts, and expect to be scored anyway.

If a detector has ever flagged your writing, did anyone let you appeal — or did the number stand? Write to us via our contact page.