AI detectors are everywhere, and they are far weaker than their confident scores suggest. Here is an honest look at why, and what to do instead of chasing a detector score.
How AI detectors work
Most detectors measure statistical properties of text, like how predictable each word is given the ones before it. AI-generated text tends to be more predictable, so the detector flags low surprise as likely machine-written. That is a probability estimate, not a measurement of origin.
Why they are unreliable
- False positives. Human writing that happens to be clear and conventional gets flagged as AI. Studies have repeatedly shown detectors mislabel text from non-native English speakers, whose phrasing is often more regular.
- Easy to fool. Light editing, paraphrasing, or just varying sentence length can drop a detector's score, which means the tool is not measuring anything robust.
- No ground truth. A detector cannot know who wrote something. It can only guess from patterns, and patterns overlap heavily between careful human writing and AI output.
- Inconsistent results. Run the same text through several detectors and you often get wildly different scores.
Because of all this, serious institutions have backed away from treating detector output as evidence.
What about removing em dashes to beat detectors?
Swapping em dashes and straightening quotes changes the surface of the text, not its statistical signature, so it does little to a real detector. Those cosmetic edits are worth doing because they make writing read as human to people, which is the audience that matters, not because they reliably move a detector score.
What actually matters
Stop optimizing for a guess. Optimize for the reader:
- Clarity. Does it say what you mean, plainly?
- Accuracy. Is every claim correct and checkable?
- Voice. Does it sound like you, with real specifics and varied rhythm?
- Clean formatting. No invisible characters, no stray em dashes, no double spaces, so the text is portable and professional.
A cleaner like textscrubr handles the formatting half, removing the hidden characters and normalizing punctuation so your text is clean and looks human-edited. The rest is good writing, which no detector can fault and no reader can mistake for filler.