What happens when the clearest signal for "this text was written by AI" becomes "this text was written by a careful human"? That's the practical problem AI detection tools have created around punctuation.
AI language models use em dashes heavily. Detectors flag them. The fix that's spread through student communities is to remove em dashes entirely, replace them with commas and periods, deliberately leave minor grammar errors, and write in a flatter, less distinctive voice. Detection scores drop. The text looks less "AI." It also looks, to any skilled editor, like someone who was told to write worse.
The loop students are navigating is genuinely circular: write naturally, get flagged. Write plainly with deliberate errors, pass detection. Then run the de-tuned version through AI to confirm it reads as human. The AI says yes. Submit.
AI detectors measure surface features - sentence length variation, vocabulary range, specific punctuation habits. They don't measure whether a human actually thought the sentence. When institutions train students to avoid em dashes, they haven't taught them anything about writing or about AI. They've taught them to game a flawed classifier.
The deeper problem is that detection is being used to solve something that isn't actually about detection. The real concern is whether students are learning to think and write. A detector score doesn't measure that. A professor who wants to know if a student understands their material needs to find out directly - through discussion, oral exams, or assignments that require knowledge the student demonstrably has.
Policing punctuation isn't a substitute for that, and it's producing writers who are worse at their craft because they've been trained to avoid tools skilled writers use. The em dash isn't a tell. It's just punctuation. Building academic policy around its presence or absence is the wrong frame, applied to the wrong problem.