AI detector false positives: who gets flagged and why

A false positive is human-written text that a detector labels as AI. The rate is the share of human documents so labelled. Small rates become large counts at scale, which is why one university's arithmetic in 2023 still gets quoted.

Vanderbilt University made that calculation when it switched off Turnitin's AI detector in August 2023: Turnitin claimed a 1 percent false positive rate at launch, Vanderbilt had submitted 75,000 papers in 2022, so around 750 student papers a year could have been wrongly labelled as partly AI written (Vanderbilt University, guidance on AI detection, 16 August 2023, read 9 November 2026). The post also cited the lack of insight into how the tool worked and the reports of non-native speakers being flagged.

Every vendor figure below is the vendor's own, on the vendor's own test set; the independent figures come from 2023 studies.

Turnitin states it keeps the rate "under 1% for documents with over 20% of AI writing", that it tests on over 700,000 pre-ChatGPT papers before each release, and that it does not display scores below 20 percent in the report (Turnitin, AI writing detection FAQ, read 9 November 2026). GPTZero states that its de-biasing work since April 2022 has reduced the false positive rate on TOEFL texts to 1.1 percent (GPTZero, technology page, read 9 November 2026). The gap between GPTZero's 1.1 percent and Weber-Wulff and colleagues' 50 percent is three years of model updates and two different test sets, which is the point: a false positive rate is a property of a detector, a text set and a date, never of detectors in general.

The guide on whether AI detectors work covers accuracy in both directions; this page stays with the human writers who get flagged.

The pattern in the published evidence is that detectors flag writing that is plain, even and predictable, whoever wrote it.

Liang and colleagues went a step further and showed the mechanism. When they prompted a chatbot to enrich the vocabulary of the TOEFL essays, the average false positive rate fell from 61.22 percent to 11.77 percent, and when they asked it to simplify the US essays as a non-native speaker might, the flags rose (Liang et al. 2023, Patterns00130-7), read 9 November 2026). The detectors were not reading authorship; they were reading vocabulary range.

That finding sits behind the fairness objection to detectors in education. A student writing in a second language, or a first-year writing to a template, produces exactly the text the models were trained to call machine made.

The vendors' own guidance is the strongest material a wrongly flagged writer has. Turnitin states that its percentage "should not be used as the sole basis for action or a definitive grading measure by instructors" and that it "provides data for the educators to make an informed decision" (Turnitin, AI writing detection FAQ, read 9 November 2026). GPTZero's FAQ states that "these results should not be used to punish students" and recommends a holistic assessment of student work (GPTZero, FAQ, read 9 November 2026). OpenAI wrote that its classifier "should not be used as a primary decision-making tool" (OpenAI, new AI classifier, read 9 November 2026).

No detector score should be the only evidence in a misconduct case.

Weber-Wulff and colleagues put the problem from the student's side: detectors return a statement like "likely written by AI" without any possibility of verification or evidence, so "a student accused of unauthorised content generation only on this basis would have no possibility for a defence" (Weber-Wulff et al. 2023, International Journal for Educational Integrity, read 9 November 2026).

The full procedure is on the page about what to do if you are flagged for AI. The short version: ask which detector, what reading and what date; collect your drafts, version history, notes and sources; quote the vendor's own statement that a score is not a sole basis for action; and ask for the review process your institution's policy sets out. Do not run your own writing through a humanizer to "fix" a flag; that creates the problem the detector was built for.

For Turnitin specifically, the Turnitin evidence page collects what Turnitin has published, including the 20 percent display threshold and the 300-word minimum, so you can check whether a reported figure is even one the tool would show.

The published method runs six human-written control texts, one per genre, through the same five detectors as every humanizer output, on the same day, so each detector's false positive behaviour is visible beside its readings of AI text. The controls were written by the editors for the test set and never published elsewhere. Those rows are the only false positive figures this site will ever state as its own; every other number on this page belongs to the study or vendor named beside it. The guides hub holds the rest of the series.