Why AI text is full of em dashes
The em dash is the most reliable typographic tell in model output. Why it happens, and why replacing it is safe.
Stylistic markers · 2 min read
Read enough model output and the pattern becomes impossible to unsee: em dashes everywhere, used as a mid-sentence pivot, often two or three in a paragraph.
Why models do it
Three things compound.
Training data skews edited. Models learn from published prose — journalism, books, essays — where a copy editor or a house style guide has already converted typed hyphens into proper em dashes. Casual writing that people actually type is underrepresented relative to how much of it exists.
The em dash is syntactically flexible. It can stand in for a comma, a colon, a semicolon, or parentheses. When a model is assembling a sentence and needs a pause whose precise grammatical role is ambiguous, the em dash is always defensible. It is the lowest-risk choice, so it gets chosen a lot.
Nothing penalises repetition of punctuation. A model avoids repeating an unusual word within a paragraph. Punctuation is not tracked the same way, so there is no pressure pushing variety.
Why it is a strong tell
Because typing one is genuinely awkward. On a US keyboard an em dash needs Option+Shift+Hyphen on macOS, or an Alt code on Windows, or autocorrect doing it for you. Most people typing quickly produce a hyphen, a double hyphen, or a comma.
So an em dash in text that is otherwise casual is out of place — the punctuation is more polished than the prose around it. That mismatch is what reads as machine-written, more than the character itself.
The related tells cluster with it: curly quotes and apostrophes (U+2018, U+2019, U+201C, U+201D) where a keyboard makes straight ones, and the single-codepoint ellipsis U+2026 instead of three periods.
Why replacing it is safe
An em dash carries no semantic content. Replacing solutions—a problem with solutions - a problem changes nothing about what the sentence means. This is a lossless substitution, verifiable by re-scanning.
That makes punctuation normalisation the highest-value, lowest-risk cleanup available: a strong signal removed at zero cost to meaning.
What to replace it with
A spaced hyphen (-) is what a person typing produces. A comma also works where the dash is doing a comma's job, but that requires understanding the sentence, so it is a rewriting decision rather than a mechanical one.
Mechanical substitution handles the character. If you also want the sentence rhythm varied, that is the rewriting layer.
Scan your text and the report will count every em dash, en dash, curly quote and ellipsis character it finds.
Related: the words that give away AI writing