False positives
When honest writing gets read as machine writing
Honest writing gets misread for reasons that are understandable once you see the mechanism.
Classifiers judge how predictable text is, not who wrote it. Careful, formulaic, or second-language writing can be highly predictable and still be entirely the work of the person who submitted it, which is where false positives come from. These pages explain the mechanism for each situation and what is genuinely worth doing. None of them will change a result you have already been given.
By situation
- student, any assignment, first time it happens
- A flagged essay is a probability score, not a finding of fact. What it means, what it doesn't yet, and the order to act in before you say anything.
- a person has accused you; the conversation itself
- What the misconduct process actually looks like, why asking for the specific flagged passages helps more than denying, and which phrases sound evasive.
- the evidence file, for authorship you genuinely hold
- Which records hold up in a real hearing: version history, timestamped drafts, and a search trail. Neither side can prove authorship from the text alone.
- the mechanism, vendor-neutral
- Detectors score statistical predictability, not authorship. The mechanism, and which human writing styles score badly by construction, not by accident.
- a college application essay
- The small-moment opening and 'that is when I realized' turn is the arc essay coaching teaches, and the arc a model defaults to. Which move to drop first.
- a graduate school personal statement
- A personal statement is judged on specificity, not polish. Detectors flag sentences with nothing concrete underneath. Read flags as a map of thin spots.
- non-native English writing flagged as AI
- Detectors score predictability and sentence variation. Liang et al. (2023) found detectors flagged 61.3% of TOEFL essays by non-native writers for exactly that reason.
- a postgraduate thesis or dissertation
- Methods and literature review chapters are formulaic by design, which is what a classifier reads as machine-shaped. Which chapters score this way, and why.