The em dash

—. The punctuation mark that became an accusation. This is what it is, why generated prose is so full of them, what an em dash does and does not prove about who wrote a passage, and how to swap them out when you want to.

Three dashes, three jobs

All three are ordinary, correct punctuation. None of them is a mistake. The en dash is the one most often used where an em dash was meant, and the hyphen is what most people actually type because it is the only one on the keyboard.

Why generated text is full of them

Nothing mysterious, and no watermark. Three ordinary reasons:

• The training text is edited prose. Books, magazines and professionally sub-edited journalism use em dashes freely, because a typesetter put them there. Casual writing — email, forum posts, notes — mostly does not, because nobody wants to hunt for the key.
• It is a safe connector. An em dash joins two clauses without committing to how they relate. A semicolon, a colon and a full stop each make a claim about the logic; a dash makes none. A model optimising for fluent, uncontroversial prose reaches for it constantly.
• Typing one is free. A person weighs the dash against the cost of inserting it. A model does not have a keyboard.

The result is a density difference rather than a difference in kind. Human prose uses em dashes. Generated prose uses them several times more often per thousand words, and — this is the part people notice — uses them the same way each time.

Does an em dash mean AI wrote it?

No. It is not evidence, and treating it as evidence gets ordinary writers accused. People who write a lot of edited prose use em dashes heavily. So does anyone typing on a Mac, where ⌥⇧- produces one without thinking, or anyone whose word processor turns -- into — automatically — which Word has done by default for over twenty years.

What a high em dash density genuinely tells you is that the text has a signature common in generated prose. That is a reason to read the passage more carefully, and it is not a finding about who wrote it. The same goes for curly quotes and the single-character ellipsis, which arrive for exactly the same reasons.

A more useful signal, if you are trying to read a paragraph honestly: how much the sentence lengths vary. Generated paragraphs hold a steady rhythm; human ones lurch. That is harder to fake and harder to trigger by accident. Humanizing AI text goes through the rest of the patterns, and the page on detectors covers why any single signal is a bad basis for an accusation.

Replacing them

Sometimes you want them gone for reasons that have nothing to do with any of the above: the destination is a plain-text field, a CSV, a code comment, a system that mangles anything outside ASCII, or a house style that says hyphens only.

Find-and-replace works for this one, because you can paste an em dash into the search box. What it will not catch is the en dash sitting among them, the curly quotes that came along with them, or the … character that looks like three full stops and is not.

In humanaze this is the Normalize punctuation switch in the settings. It is off by default and deliberately so — folding — to - changes how the text reads, and that is your decision rather than a cleanup. Turn it on and press Clean, and every em dash, en dash, curly quote and ellipsis character becomes its plain ASCII equivalent in one pass. The report counts them first, so you can see how many there are before deciding.

How to type one

• Mac: ⌥⇧- for an em dash, ⌥- for an en dash.
• Windows: Alt+0151 on the numeric keypad, or Ctrl+Alt+- in Word.
• Linux: Ctrl+⇧+U then 2014.
• HTML: — or the character itself in a UTF-8 file.

Spacing is a house-style question with no correct answer. American style closes the dash up—like this. British style usually spaces it — like this. Pick one and be consistent; the inconsistency is far more noticeable than either choice.

Common questions