AI detector false positives are not a theoretical edge case.
A false positive occurs when human-written text is classified as AI-generated. The risk varies with the detector, threshold, language, genre and test data.
What published evaluations can—and cannot—show
Liang et al. (2023) tested seven detectors on a specific English-language dataset that included TOEFL essays. The study found substantial misclassification of non-native English writing in that setup. This does not provide a universal rate for every detector, language or current model.
Tufts et al. (2025) evaluated detectors across unfamiliar models, data sources and practical evasion prompts. Performance varied sharply across conditions, supporting the need to report error rates at a defined operating threshold rather than one headline accuracy.
Institutional guidance from Freie Universität Berlin and the University of Potsdam likewise cautions against treating AI detectors as hard proof. Apply your own institution’s current policy.
Why human prose can look “AI-like”
- formal academic genres constrain vocabulary and structure;
- short texts provide little evidence for classification;
- language learners may use more regular, predictable constructions;
- translation and intensive proofreading change surface patterns;
- technical sections repeat terminology and conventional phrases.
Why the displayed score can mislead
A user interface may look like a probability, but the number may represent an internal classification score or aggregated passage result. Without calibration data, a threshold and the relevant base rate, it should not be read as “the probability that this student used AI”.
If your work is flagged
- Ask which detector, version, threshold and passages were used.
- Preserve drafts, notes, sources and version history.
- Explain the argument and how the cited evidence supports it.
- Document permitted language, translation or editing tools.
- Request a human review under the institution’s published process.
If you are reviewing someone else’s work
Look for corroborating process evidence and content problems, not stylistic stereotypes. Give the writer a meaningful opportunity to respond. If policy is unclear, pause the allegation and resolve the procedure first.
Sources and scope
- Liang et al., Patterns (2023) — English datasets and detectors available at the time.
- Tufts et al., Findings of NAACL (2025) — cross-model and cross-domain evaluation.
- Turnitin AI Writing Report guide — vendor documentation, not an independent evaluation.
Use the detector as one input to a human review.
Read the highlighted passages and retain evidence of the writing process.
Run an AI-writing check