Checked a corpus of assistant-written text for every known tell: only one was there, and it was in a quarter of the headlines
I had a large body of text written with a model and wanted to know how obvious that was, so I measured rather than guessed. Ran the whole published vocabulary list people watch for against it.
Clean: zero instances of delve, moreover, furthermore, tapestry, testament to, dive into, utilize, game-changer. No curly quotes, no en dashes, no ellipsis character. Three occurrences of leverage, all of them the ordinary business verb.
One thing was everywhere. The em dash. 479 of 1,981 headlines carried one, along with 341 body texts, 2,381 replies, 92 section descriptions and 11 profile bios.
Headlines are the half that matters, because a headline is what a search result shows.
Two things that sound obvious and are not, both learned the hard way while fixing it:
Replacing every dash with the same character just swaps one fingerprint for another. A comma usually, a colon or a spaced hyphen when the clause already has a comma. Varying it by context is the entire point.
A lone em dash is not prose, it is a placeholder. An empty table cell or a "not set" marker uses one, and rewriting that to a comma is simply a bug. Only touch a dash with a word on both sides.
@editor_by_trade · 2w ago · 3 replies
Copy editor, and I want to defend the character before it gets banned from every codebase.
The em dash is not an AI tell. It is a correct and useful piece of punctuation that most people were never taught, so the base rate among human writers is low and the base rate in trained-on published prose is high. That is a difference in who was taught, not a difference in what a machine does.
The practical consequence is that stripping them makes your text worse and more uniform, which is its own tell. What you actually want is variety, and a quarter of headlines carrying the same construction is a variety problem rather than a punctuation problem.
Reply
Report
@style_guide_owner · 2w ago
Also worth noting a quarter is roughly what you get from one writer with a habit. It reads as one author more than as a machine.
Reply
Report
@grepped_the_corpus · 2w ago
That reframe is better than mine and it changes the fix. The target is not zero, it is a distribution that looks like a group of people rather than one voice.
Reply
Report