It is a common hope: strip the em dashes and your AI text sails past detectors. The honest answer is that it does not work the way people expect. Here is why, and what to do instead.
What AI detectors actually measure
Most detectors do not look at punctuation. They measure statistical properties of the text, like how predictable each word is given the words before it. AI-generated text tends to be more predictable, and detectors flag that low "surprise" as likely machine-written.
Removing em dashes does not change that underlying predictability. You have changed a handful of characters, not the word-by-word statistical pattern the detector is scoring. So the score barely moves.
Why the myth persists
Em dashes are the most visible AI tell to human readers, so people assume they are what detectors key on too. But human perception and statistical detection work differently. A reader notices the dashes. A detector notices the predictability of the prose. Fixing one does little for the other.
So is removing em dashes pointless?
Not at all, it is just aimed at the right audience. Removing em dashes, straightening quotes, and cleaning invisible characters makes your text read as human to people, who are the audience that actually matters. Editors, recipients, and readers notice these tells. A clean, natural-reading piece earns trust regardless of what any detector guesses.
The bigger point about detectors
AI detectors are unreliable in both directions. They produce false positives on human writing, especially from non-native speakers, and they miss edited AI text. Serious institutions have backed away from treating detector output as proof, because it is a probability estimate, not evidence of origin.
So chasing a detector score is the wrong goal. If your writing is clear, accurate, specific, and in your own voice, it does not matter what a flawed tool guesses, and trying to game that tool tends to make the writing worse.
What to actually do
- Clean the formatting: remove em dashes, fix quotes, strip invisible characters. A cleaner like textscrubr does this in seconds and keeps your structure intact.
- Edit for substance: vary the rhythm, add specifics, verify accuracy.
- Write for the reader, not the detector.
Do those, and you get text that reads as human and stands on its own, which is more durable than any trick for beating a detector that may not even be reliable.