AI Detection False Positives: Why Human Writing Gets Flagged as AI

A false positive is human writing that an AI detector labels as machine-generated. Turnitin says it keeps that rate under 1% for whole documents, but only under specific conditions, and independent studies have measured far higher rates for some detectors and some groups of writers. That is why a detector score is a reason to look closer, never proof of cheating on its own.

This explainer collects what the vendors say, what the research found, and cases that were reported in detail, with a source for every number. If you have already been flagged, go straight to our step-by-step guide on how to prove your writing is human.

Turnitin's own false-positive figures, and the fine print

Start with Turnitin, whose software The Markup reported in 2023 was used by over 16,000 academic institutions. Its current AI detection FAQ states the headline figure this way: it aims to keep its false positive rate "under 1% for documents with over 20% of AI writing. In other words, we might flag a human-written document as AI-written for one out of every 100 fully-human written documents."

That sentence carries conditions, and Turnitin's own documents add more:

Two more lines from Turnitin are worth knowing. Seven weeks after launch, its chief product officer wrote that "real-world use is yielding different results from our lab" (23 May 2023). And the current FAQ repeats the claim after a July 2026 change that merged several models into one, "maintaining a less than 1% false positive rate." For what the percentage bands mean in practice, see what percentage of AI is acceptable on Turnitin.

Why "under 1%" can still mean hundreds of students

A small rate applied to a large number of papers is still a lot of people. Vanderbilt University did the arithmetic when it disabled Turnitin's AI detector on 16 August 2023 "for the foreseeable future": Turnitin claimed a 1% false positive rate at launch, Vanderbilt had submitted 75,000 papers in 2022, so "around 750 student papers could have been incorrectly labeled as having some of it written by AI."

Vanderbilt gave other reasons too. It said Turnitin "gives no detailed information as to how it determines if a piece of writing is AI-generated", and it cited the finding that detectors are more likely to label writing by non-native English speakers as AI.

The same arithmetic applies to one student over a whole degree. Cesare Giulio Ardito, a lecturer at the University of Manchester, calculated in a December 2023 paper that if the false positive rate were a uniform 1%, "an entirely honest student who completes 70 essays during their degrees faces a 50.5% chance of being erroneously flagged for AI use at least once." That assumes the rate applies evenly to every paper, which Turnitin's conditions complicate, but it shows why a per-document rate understates the risk to a person.

What independent research has found

Liang et al. (2023): non-native writers flagged far more often

The most cited study is Liang and colleagues at Stanford, published in the journal Patterns in July 2023. They ran seven detectors (Turnitin was not one of them) on 91 TOEFL essays written by non-native English speakers and 88 essays by US eighth-graders.

• The detectors were near-perfect on the US essays.
• On the TOEFL essays the average false positive rate was 61.22%.
• All seven detectors flagged 18 of the 91 essays; at least one detector flagged 89 of them (97.80%).
• When the researchers used ChatGPT to enrich the essays' vocabulary, the average false positive rate fell to 11.77%.

Those figures are from the preprint; the journal version reports them as 61.3% and 11.6%. The detectors were the versions available in March 2023, and Turnitin later pointed out that the TOEFL essays were all under 150 words (Turnitin, 26 October 2023), so the exact numbers date quickly. Turnitin and others have disputed how far the result generalizes; we lay out both sides in AI detection and ESL students.

Weber-Wulff et al. (2023): 14 tools, none reliable

A peer-reviewed test in the International Journal for Educational Integrity ran 14 detection tools, including Turnitin and GPTZero, on human-written, machine-translated and AI-generated documents. The authors found the tools "neither accurate nor reliable (all scored below 80% of accuracy and only 5 over 70%)." Six of the 14 produced false positives, "with the risk increasing dramatically for machine-translated texts." Turnitin scored highest overall, and the tools' main bias was toward calling AI text human rather than the reverse. The conclusion was blunt: "the systems we tested should not be used in academic settings."

OpenAI's own classifier

OpenAI launched an AI-text classifier on 31 January 2023. On its own "challenge set" of English texts, it correctly identified 26% of AI-written text while labeling human-written text as AI 9% of the time. OpenAI withdrew it: "As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy." The same page warned it was "very unreliable on short texts (below 1,000 characters)."

Park, Jeong and Kim (August 2026): style is a confound

The newest large study, Style as a Confound (submitted 27 August 2026 and presented at EMNLP 2026), used 135,389 pairs of documents from a professional English editing service, covering 2018 to 2025: manuscripts by non-native English writers, and the same manuscripts after native-speaker editing. Because content and authorship stay fixed, changes in a detector's score can be attributed mainly to the editing, not to who wrote the text.

• On human-written documents from before ChatGPT (2018 to 2022), the 13 detectors' false positive rates ranged from 0.0% to 100.0%, depending on the detector.
• The same edits raised AI scores in some detectors and lowered them in others, and the shifts grew with the amount of editing.
• The authors' practical conclusion: "AI detector outputs should not be treated as definitive evidence of AI use, especially in academic settings where editing is common."

Read it with its limits in mind. The 13 detectors are publicly available research detectors; the authors say the analysis "does not cover future systems or proprietary commercial detectors", so Turnitin and GPTZero were not tested. One co-author works at the editing company whose records were used, and the authors note that absolute rates depend on the thresholds chosen.

Why detectors flag human writing

Detectors do not know who wrote a text. They estimate how likely the wording is. Turnitin's FAQ describes classifiers trained on differences in "word probability", on the basis that human writing "tends to be inconsistent and idiosyncratic." Several kinds of honest writing fall into the gap:

• Plain, predictable wording. In Liang's study, simplifying the vocabulary of US students' essays raised the average false positive rate from 5.19% to 56.65%.
• Formulaic passages. Turnitin lists "content without a lot of structural variation, text that literally repeats itself, or text that has been paraphrased without developing new ideas." In May 2023 it also reported more false positives in "the first few or last few sentences of a document", where introductions and conclusions are often generic (release notes).
• Short texts. Turnitin says submissions under 300 words "may result in an AI writing score that is likely less accurate" (15 December 2023 note).
• Mixed documents. Most of Turnitin's sentence-level false positives sat next to genuinely AI-written sentences.
• Editing and polish. The 2026 study above found editing pushes scores in both directions, depending on the detector. The Markup reported that Turnitin's chief product officer "admitted the tool ends up learning that more complex writing is more likely to be human."
• Machine translation. Weber-Wulff's team found false-positive risk rose sharply for machine-translated human text.
• Languages the detector was not built for. Turnitin's AI detection covers only long-form English, Spanish, Japanese and Modern Arabic, and returns no report for other languages (FAQ). Other detectors do not always refuse. SupWriter's own detector, like most, is unreliable for Malay: in a small check it scored human-written Malay text as 100% AI (how we test).

Cases reported in detail

• UC Davis, 2023. Rolling Stone reported that Turnitin flagged part of a brief by senior Louise Stivers as AI-written, and her case went straight to the student conduct office. Clearing her name took more than two weeks; she did it by giving officials "step-by-step instructions on how to open Google Docs and review history." The same article describes a classmate who got a failing grade after a professor ran his exam answers through GPTZero.
• Johns Hopkins, 2023. The Markup reported that Turnitin labeled "more than 90 percent" of a student's paper as AI-generated. The student immediately "showed me drafts, PDFs with highlighter over them," said lecturer Taylor Hahn, who concluded the tool had made a mistake and noticed it flagged international students' writing more often.
• University of Waterloo. When it stopped using Turnitin's AI detection as of September 2025, Waterloo said internal testing had found that "in more than one instance the product flagged human written text as 100% generated by AI."
• Adelphi University, 2026. A New York court annulled an academic integrity finding on 28 January 2026 that began with a professor's Turnitin result of 100% AI-generated. The court did not rule on whether the essay was AI-written. It found the finding had no valid basis: the professor's own later email undermined the allegation, the student was not allowed an advisor, his evidence (two other detectors scoring the essay as human) was not considered, and the same official decided both the case and the appeal. It ordered his record expunged.

Institutions that switched it off

Vanderbilt was not alone. The University of Pittsburgh's teaching center told faculty in June 2023 that false positives "carry the risk of loss of student trust, confidence and motivation, bad publicity, and potential legal sanctions" (The Markup). Waterloo cited unreliability and bias against students whose first language is not English. Curtin University in Australia announced on 4 September 2025 that the feature would be disabled from 1 January 2026, framing the change around trust and assessment design. We keep a sourced list of these decisions in our university AI-detection policy tracker.

For balance: most institutions kept the feature, at least early on. The Markup reported that by the end of July 2023, only two percent of Turnitin's customer institutions had used the option to switch the AI indicator off.

Outside the classroom

Freelancers and content teams face the same tools with fewer rules. If the worry is search rankings, Google's stated position is that it rewards quality "rather than how content is produced", and that "appropriate use of AI or automation is not against our guidelines" (Google Search Central, 8 February 2023). What it penalizes is using automation mainly to manipulate rankings. A client pointing at a detector score is working with the same uncertain signal as a professor, so the same defense applies: show your drafts and version history.

If a detector flags your work

1. Ask which tool, and for the report. Turnitin's AI report is instructor-only, but instructors can download it as a PDF and share it (FAQ). More in can students see their Turnitin AI score?
2. Check the conditions. Was the document over 300 words of prose? Is the language supported? A *% result means the score was below 20% and deliberately not shown (FAQ).
3. Bring process evidence. Version history, drafts, notes and sources are what cleared the students above.
4. Quote the vendor. Turnitin says the percentage should not be "the sole basis for action."

The full checklist, including how to appeal, is in how to prove your writing is human.

---

SupWriter's AI detector is a pre-check built on one scoring model. It is not the detector your school uses, and like every detector it can be wrong. Try it on a draft: 300 words free, no credit card required (sign up).