The short answer
- Yes, detectors flag honest students, and some groups far more than others. In the Stanford-led study (Liang et al., 2023), seven detectors labeled human-written TOEFL essays by non-native English writers as AI 61.22% of the time on average; 89 of 91 essays were flagged by at least one detector, while US eighth-graders’ essays were classified near-perfectly.
- OpenAI withdrew its own classifier in July 2023: it caught only 26% of AI text and mislabeled 9% of human text as AI.
- Turnitin’s own figures: under 1% document-level false positives, but only for documents with more than 20% detected AI; about 4% at sentence level.
- Independent tests were harsh in 2023 and 2024 (six detectors averaged 39.5% accuracy, falling to 17.4% on manipulated text). A 2025 working paper found one commercial detector far better on unedited text, so the picture is not static.
- Universities have switched detection off: Vanderbilt (2023), Waterloo (2025) and Curtin (from January 2026).
- Bottom line for teachers: a detector score is a reason to talk to a student, not evidence of misconduct.
Every AI detector produces a score. The question that matters to a student is what happens when that score is wrong. This page collects the six most-cited pieces of evidence on false positives (four independent studies and two vendor evaluations) and sets out what they found, in their own numbers.
What did the six studies find?
The research is consistent on one point: detectors make mistakes that land on real students. It is less consistent on how big those mistakes are, which depends on the tool, the year and the kind of writing tested.
| Study | Year | Type | Detectors | Key finding | Source |
|---|---|---|---|---|---|
| Liang, Yuksekgonul, Mao, Wu, Zou, Patterns | 2023 | Independent, peer-reviewed | 7 widely used GPT detectors | 61.22% of 91 non-native TOEFL essays misclassified as AI on average; 18/91 flagged by all seven; 89/91 by at least one; US 8th-grade essays near-perfect | arXiv, Stanford HAI |
| Weber-Wulff et al., Int. J. Educational Integrity | 2023 | Independent, peer-reviewed | 14 (12 public tools plus Turnitin and PlagiarismCheck) | Tools “neither accurate nor reliable”; biased toward calling text human; paraphrasing made results worse | arXiv |
| OpenAI AI Text Classifier | 2023 | Vendor self-report | OpenAI’s own classifier | Flagged 26% of AI text as “likely AI-written”; mislabeled 9% of human text; withdrawn July 2023 | Decrypt |
| Turnitin validation (Chechitelli blog) | 2023 | Vendor self-report | Turnitin AI detection | Under 1% document-level false positives for documents with more than 20% AI; about 4% sentence-level; tested on 800,000 pre-ChatGPT papers | Turnitin |
| Perkins et al., Int. J. Educational Technology in Higher Education | 2024 | Independent, peer-reviewed | 6 major detectors | 805 samples: 39.5% baseline accuracy, 17.4% on adversarially manipulated text | arXiv |
| Jabarian and Imas, BFI Working Paper 2025-116 | 2025 | Independent, not peer-reviewed | Pangram, OriginalityAI, GPTZero, RoBERTa | Commercial tools beat open-source; Pangram near-zero false positives and negatives on its corpus; only Pangram met a 0.5% false-positive cap | BFI (PDF) |
Read the table chronologically and a pattern appears. The 2023 and 2024 studies are pessimistic. The 2025 working paper suggests the best commercial tools have improved sharply on unedited text. None of them shows that a score can establish who wrote something.
How badly are non-native English writers affected?
More than any other group studied. The Liang study is the most-cited evidence on false positives because the gap is so large. Researchers ran seven detectors over 91 essays written for the TOEFL by non-native English speakers, with no AI involved.
| Measure | Result |
|---|---|
| Average share misclassified as AI (7 detectors) | 61.22% |
| Essays flagged by at least one detector | 89 of 91 (97.8%) |
| Essays flagged by all seven detectors | 18 of 91 (19.8%) |
| US 8th-grade essays | Classified “near-perfectly” |
Source: Liang et al. (2023), via Stanford HAI; percentages computed from 89/91 and 18/91
Put plainly: if a teacher ran one of these essays through several detectors, the chance that at least one would call it AI was close to certain. A student who writes in clear, simple English because English is their second language is exactly the student these tools punish. Vanderbilt and Waterloo both cited this bias when they switched detection off.
What do the vendors themselves say?
Their numbers are better, but they come with conditions.
OpenAI launched its AI Text Classifier in January 2023 and withdrew it in July 2023, citing “its low rate of accuracy.” On its own challenge set, it labeled only 26% of AI-written text as “likely AI-written” and wrongly labeled 9% of human-written text as AI.
Turnitin’s chief product officer, Annie Chechitelli, wrote in May 2023 that the document-level false-positive rate is under 1%, but only for documents where more than 20% of the writing is detected as AI. The sentence-level false-positive rate is about 4%. Turnitin added an asterisk to scores under 20%, raised the minimum length from 150 to 300 words, and changed how it handles first and last sentences because false positives clustered there. K-12 Dive reported that Turnitin acknowledged a “higher incidence of false positives” below 20%, and that of 38.5 million submissions processed by May 14, 2023, 9.6% showed more than 20% AI writing and 3.5% more than 80%. The company had initially claimed 98% confidence.
What does a 1% false-positive rate mean in a real school?
At scale, small percentages become real students. Vanderbilt made this point when it disabled Turnitin’s detector in August 2023: the university had submitted about 75,000 papers in 2022, so a 1% false-positive rate could mean roughly 750 papers wrongly flagged.
The same arithmetic works for one teacher. These are estimates using Turnitin’s stated best-case rate and simple multiplication, assuming every paper is human-written:
| Scale | Papers per year | Expected wrong flags at 1% |
|---|---|---|
| One class of 30, 10 assignments | 300 | 3 |
| One teacher, 150 students, 10 assignments | 1,500 | 15 |
| Vanderbilt, 2022 (its own figure) | ~75,000 | ~750 |
And 1% is the vendor’s figure for documents above the 20% threshold. At the sentence level, Turnitin’s own 4% means a fully human essay of 50 sentences would, on average, have about 2 sentences highlighted (50 × 0.04 = 2; an estimate that treats sentences independently).
Which universities have switched AI detection off?
Several, citing exactly these problems.
| Institution | Decision | Reasons given |
|---|---|---|
| Vanderbilt University (US) | Disabled Turnitin’s AI detector Aug 16, 2023, “for the foreseeable future” | ~750 possible wrong flags a year at 1%; no opt-out at launch; lack of transparency; bias against non-native speakers; privacy |
| University of Waterloo (Canada) | Discontinued Turnitin AI detection effective September 2025 | Unreliability; bias against non-native English speakers; inconclusive internal testing; costs “outweigh the benefits” |
| Curtin University (Australia) | Turnitin AI detection off from January 1, 2026 (announced Sep 4, 2025) | Text-matching for plagiarism stays on |
A third-party tracker run by GradPilot claims more than 60 institutions in five countries have disabled or declined AI detection. We have not verified each entry, so we list only the three above, which we confirmed on each university’s own page.
Some K-12 districts are moving the same way. Orange County Public Schools in Florida adopted its first AI policy in July 2026, and it bars teachers from using AI detectors, according to Spectrum News 13. See our tracker of AI policies in 32 large US districts.
What happens when a student is wrongly flagged?
Sometimes it ends up in court. In October 2025, Plagiarism Today reported that the family of Orion Newby, a sophomore with learning differences at Adelphi University in New York, sued after Turnitin flagged his history paper. His instructor reportedly called the paper “too advanced”; he received a zero and a mandatory anti-plagiarism course. Other lawsuits have been reported, but we could not confirm their details from primary or reliable sources, so we leave them out.
Students who are flagged should gather process evidence: version history in Google Docs or Word, drafts, notes and sources, and be ready to explain their argument out loud. Our evidence review, do AI detectors actually work?, has a step-by-step list.
What should teachers do instead?
Treat a flag as the start of a conversation, never the conclusion.
- Don’t act on a score alone, especially for multilingual writers, whom the evidence shows are flagged far more often.
- Ignore low scores. Turnitin itself marks scores under 20% as less reliable.
- Ask about process. “Walk me through how you wrote this” is more informative than any percentage.
- Design work that shows thinking: staged drafts, in-class writing samples, oral check-ins and prompts tied to class discussion.
- Put the rules in writing, including what counts as acceptable AI use and how students should disclose it.
Practical guidance is in detecting AI work honestly, AI-proof assignments and our classroom AI policy template. Students can read how to use AI for homework without cheating.
Limits of this review
Detection technology changes quickly, and the newest evidence (Jabarian and Imas, 2025) is more favorable to commercial tools than older studies. That paper is not peer-reviewed and tested its own corpus, so its results may not carry over to edited or mixed human-AI writing. Vendor figures come from vendors. And no study here measured outcomes for K-12 students’ everyday homework. What the evidence does support is narrow and firm: a detector score is a probability estimate that is sometimes wrong, and it is wrong more often for some students than others.
Sources
- Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns. arXiv:2304.02819; numbers via Stanford HAI
- Weber-Wulff, D. et al. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity 19, 26. arXiv:2306.15666
- OpenAI AI Text Classifier withdrawal: Decrypt (original: OpenAI)
- Perkins, M. et al. (2024). GenAI detection tools, adversarial techniques and implications for inclusivity in higher education. IJETHE 21, 53. arXiv:2403.19148
- Jabarian, B. and Imas, A. (2025). Artificial writing and automated detection. BFI Working Paper 2025-116. PDF
- Turnitin: AI writing detection update from the CPO; K-12 Dive coverage
- Universities: Vanderbilt; University of Waterloo; Curtin University; tracker: GradPilot
- Orange County Public Schools policy: Spectrum News 13
- Adelphi case: Plagiarism Today
Checked October 14, 2026. If a study is updated or a figure is wrong, tell us and we will correct it.