Why AI's Em Dash Habit Reveals Unedited AI Output

The em dash is a character almost nobody used to type. Named for its width—traditionally one em, the width of the letter M—it has a much shorter cousin, the en dash. Printers adopted something close to it in English drama as far back as 1588, in Maurice Kyffin's translation of Terence's Andria, using a run of hyphens to mark broken, interrupted speech. A few decades later, other translations of the same play used the same trick dozens of times. Eventually the expensive folio editions settled on a single unified dash instead of a chain of hyphens, mostly because it looked better. From there the em dash moved comfortably into professionally edited prose—novels, essays, journalism—where it appeared in the Chicago Manual of Style and the AP Stylebook. American convention kept it unspaced; British and Canadian style guides often preferred a spaced en dash instead. But in everyday, unedited human writing, the em dash stayed rare because it was a small pain to produce. A standard keyboard gives you a hyphen and nothing else. Getting a real em dash meant Option+Shift+Hyphen on a Mac, an alt code on Windows, or the old Courier workaround of stringing two hyphens together, because monospaced typewriter fonts never had a dedicated glyph for it.

Large language models learned to write from the opposite end of that spectrum. Their training diet came from books, academic papers, and edited journalism—exactly the domains where the em dash shows up constantly, because it costs nothing to produce with a keyboard shortcut and carries none of the friction that keeps casual writers away from it. That training taught the models that this is what careful, "considered" writing looks like, so they reach for the same device sentence after sentence, whether or not the content actually calls for a pause or an aside. What the models don't have is what a human editor has: a sense of their own tics. A writer eventually notices they lean on a phrase too much and cuts it. A model generates one token at a time, with no memory of how many times it already used the same trick three paragraphs up. Nobody trained it to notice its own habits, so nothing ever tells it to stop.

The result is that the em dash has stopped being neutral punctuation. A person typing quickly in a DM doesn't go out of their way to produce a character that needs a special key combination. They reach for a hyphen, a comma, or nothing at all. A message full of unspaced dashes doing the exact same rhetorical move over and over isn't how people type under normal conditions. It looks exactly like what it is: text a model produced, pasted in unchanged. This creates a problem rooted in trust. A Slack DM or Notion page written entirely in that voice reads like nobody actually stood behind it. If the sender didn't write it themselves, did they even think it through, or did they just ask a model what to say and forward the reply? The message stops being proof of someone's judgment and starts being proof that the judgment got outsourced. Writing something yourself, even a couple of sentences, is a small act of care toward whoever's reading it. Pasting model output wholesale, dashes and all, without even a pass to make it sound like you, tells the reader you didn't think the message was worth that much of your time. And underneath both of those, slower and easier to miss day to day, sits homogenization. When enough people on a team lean on the same tool for the same tone, every DM, every project update, every Notion doc starts sounding like it came from the same person, and that person is nobody in particular. A team's culture lives partly in how people actually talk to each other: the shortcuts, the register, the little quirks of individual voice. Flatten all of that into one polished, hedged, dash-heavy cadence, and you lose something that was quietly doing real work. The recommendation is to read what comes back before it goes anywhere, and rewrite the parts that don't sound like you. A tool that ghostwrites your Slack messages and your Notion pages is, quietly, a tool writing your reputation too.

The em dash occupies a peculiar place in modern writing history. For centuries it belonged to professional typesetters and editors—people whose job was to polish prose. It first appeared in English drama around 1588 in Maurice Kyffin's translation of Terence's Andria, where it marked broken, interrupted speech. From there it settled comfortably into novels, essays, journalism, and style guides like the Chicago Manual of Style and AP Stylebook. The barrier to everyday writers was simple: a standard keyboard offered no em dash. On a Mac you needed Option+Shift+Hyphen; on Windows, an alt code; or you could use the old typewriter trick of two hyphens in a row. That friction kept the mark rare in unedited, casual human writing.

Large language models encountered the em dash from the opposite end of that spectrum. Their training data—books, academic papers, edited journalism—is exactly the domain where dashes are abundant and free to produce. The models learned that this punctuation pattern is what "careful, considered" writing looks like, so they reach for it repeatedly. But they lack what a human editor has: a sense of their own tics. A writer notices they lean on a phrase too much and cuts it. A model generates one token at a time, with no memory of how many times it used the same trick three paragraphs earlier, and no training to recognize or break the pattern. The result is text that reads like nothing a human would produce under normal conditions—a visible fingerprint of AI authorship.