False Positive

A false positive in AI detection occurs when a piece of writing produced entirely by a human is labeled as likely AI-generated by a detector. This is not a rare edge case. Turnitin itself reports roughly one false positive per hundred documents at the document level, and says the rate is worse on short passages. That figure is the vendor's own, measured on the vendor's corpus — no independent replication of it exists, because the terms of service of every major detector prohibit the automated querying that producing one would require.

Certain writer profiles are disproportionately affected. Non-native English speakers, technical and academic writers, and students who write in formal register all produce prose with lower perplexity and more uniform burstiness — the same statistical patterns that flag AI content. This means populations already at higher risk of being unfairly judged are precisely the ones detectors struggle with most. Liang et al., writing in Patterns in 2023, ran seven GPT detectors over TOEFL essays by non-native English speakers and found the majority misclassified as AI-generated, while essays by US school students were classified almost perfectly as human.

For individuals, understanding false positives changes how detection scores should be interpreted. A high score is a probability estimate, not a verdict. Best practice is to treat any single detector's output as one data point rather than a conclusion, and to check suspected human-written content against multiple detectors to see whether they agree. Institutions that act on detector output without a human review step are building policy on a statistically unreliable foundation. Coda One publishes its own detector's measured false positive rate, the benchmark behind it, and the kinds of text it is known to get wrong, at /ai-detector/accuracy.