The em dash and ChatGPT
By Sourabh Singh18 September 20265 min read
Why the most popular AI tell is mostly wrong, and what actually works.
Some time in 2025 the internet decided that em dashes prove AI authorship. The claim spread because it is memorable and because model output genuinely does use them more than the average person. It is still a bad test.
Why it fails
- Plenty of excellent human writers lean on em dashes heavily. Punctuation preference is a style, not a fingerprint.
- Plenty of model output contains none at all, depending entirely on the prompt.
- It is a single feature. Every real detector weighs dozens, and punctuation is a minor one.
- It is trivially defeated, anyone evading detection removes them in seconds.
What is actually a signal
Not the rate itself, but a change in the rate. If your last twenty pieces used two em dashes per thousand words and today's draft uses fourteen, something about how this was written is different. That comparison requires a baseline of your own writing, which is exactly what a style fingerprint is.
The same logic applies across every punctuation habit: semicolons, question frequency, contraction rate. In isolation each is meaningless. Measured against your own history, a sharp deviation is informative.
The practical advice
Do not strip em dashes from your writing to avoid suspicion. Removing punctuation you would naturally use makes your prose less like you, which is precisely backwards. If you are worried, measure your own baseline and write to it.
Measure your punctuation habits against your own baseline.
Check your rateCommon questions
No. Model output does use em dashes more than the average person, but plenty of excellent human writers lean on them heavily and plenty of model output contains none, depending on the prompt.
Only as a minor signal. Every real detector weighs dozens of features, and punctuation is a small one.
No. Removing punctuation you would naturally use makes your prose less like you. If you are worried, measure your own baseline and write to it.
A change against your own baseline, not the rate itself. If your last twenty pieces averaged two em dashes per thousand words and today's draft has fourteen, something about how it was written is different.
