IAPQB

AI writing detectors on campus

How universities, an ombudsman and a court have treated AI detector scores in misconduct cases

IAPQB Report No. 10 · Practice report · Published 17 November 2026
International AI Practice Qualification Board (IAPQB)

In one minute

  • Turnitin switched on AI writing detection on 4 April 2023 for more than 10,700 institutions. It says its score should not be the only basis for action [1, 2].
  • In 2023, Stanford researchers reported that seven detectors labelled on average 61.3% of 91 TOEFL essays by non-native English writers as AI-generated [3].
  • Vanderbilt University, the University of Cape Town and Curtin University decided to turn the detector off [4, 5, 6]. The student ombudsman for England and Wales and a New York court found for students, fully or in part, where evidence was not shared or not weighed [7, 8, 9, 10].
  • Turnitin and the ombudsman treat the score as one piece of evidence. The ombudsman asks for fair process and bias checks, and Australia's regulator for assessment redesign [2, 11, 12].

1. A familiar situation

A lecturer marking essays sees a high AI score on one of them. The student writes in English as a second language, in short, plain sentences. The number looks exact. Nobody has yet asked the student how the essay was written.

2. Why detectors get it wrong

A detector sees only the finished text and estimates whether a language model wrote it. Most detectors use perplexity, a measure of how easily a language model can guess the next word. Second-language writing, with its narrower range of words and grammar, can be easy to guess too [3].

Every detector balances two errors. To avoid marking human writing as AI, Turnitin accepts that it may miss some AI text: a document it scores at 50% AI could contain as much as 65% [2]. At the 1% false positive rate Turnitin claimed at launch, Vanderbilt calculated that around 750 of the 75,000 papers it submitted in 2022 could have been wrongly labelled [4]. OpenAI withdrew its own classifier in July 2023 for its "low rate of accuracy"; in its tests on English texts it wrongly flagged 9% of human-written text [13].

IAPQB view. In the ombudsman and court cases below, students won, fully or in part, where the process around the score failed. Evidence was not shared, drafts and explanations were not weighed, or nobody asked whether the tool could misread a non-native speaker [7, 8, 9, 10].

3. Four cases

3.1 Stanford, 2023: second-language essays flagged as AI

In an opinion article in the journal Patterns, Stanford researchers reported a test of seven widely used detectors on 91 TOEFL essays and 88 essays by US students, all written by people [3]. On average the detectors labelled 61.3% of the TOEFL essays as AI-generated; at least one flagged 97.8%. The US essays were classified almost perfectly. The authors "strongly caution against the use of GPT detectors in evaluative or educational settings" [3]. Turnitin was not tested. Its own later study found no statistically significant bias against English language learners in documents of 300 words or more, and a larger gap in shorter ones [14].

3.2 United States, South Africa, Australia, 2023 to 2025: three universities drop the detector

In August 2023 Vanderbilt University said it would disable Turnitin's AI detector for the foreseeable future, citing less than 24 hours' notice before launch, no detail on how the tool decides, and research on bias against non-native English speakers. It told instructors with concerns to talk to their students [4].

The University of Cape Town (UCT) said it would stop using tools such as Turnitin's AI score from 1 October 2025. The director of its teaching innovation centre said "AI detectors are simply not reliable" [5]. Curtin University in Australia said Turnitin's AI detection would be disabled from 1 January 2026, while text-matching checks would remain [6].

3.3 England and Wales, 2025: the student ombudsman's cases

The Office of the Independent Adjudicator for Higher Education (OIA) reviews student complaints about higher education providers in England and Wales [15]. Its July 2025 casework note and case summaries cover AI and academic misconduct [11].

In one case, Turnitin flagged an international student's coursework, and after a viva the student got a zero. The OIA found the complaint justified: none of the evidence relied on had been shared before the viva, which gave no chance to explain how the student had worked [7]. In another, the provider had not considered whether Turnitin "might be less reliable for non-native English speakers", or the notes and drafts its procedures said it would ask to see. The OIA found that complaint partly justified, and a separate collusion finding reasonable [8]. In a third, a student with autism said the provider's detection software had earlier flagged another of their essays, which was then cleared. The panel had not properly considered the student's evidence, including planning and preparation, and the OIA found the complaint justified. The student later told the OIA that the provider, on reconsideration, found no misconduct [9].

3.4 New York, 2026: a court annuls a finding after a 100% score

In November 2024 an Adelphi University professor gave a first-year student's essay a zero and reported him, relying on a Turnitin report that gave it an AI-generated score of 100% [10]. The student denied using AI and offered results from two other detectors. The university's academic integrity officer issued a finding of violation and later denied the appeal. In between, the professor wrote to the student that the office's system seemed to treat him as responsible "unless you contest the claim again".

On 28 January 2026 the New York State Supreme Court in Nassau County held that this email undermined the claims. It annulled the violation report and the appeal denial as "without valid basis and devoid of reason". The officer had "failed to even consider" the student's evidence, the student could not confer with an advisor of his choice, and the same officer heard the appeal. It ordered the record expunged, and did not rule on whether any detector is reliable [10].

4. How practitioners handle it now

The score is one piece of evidence. Turnitin says its percentage "should not be used as the sole basis for action" [2]. Australia's higher education regulator, the Tertiary Education Quality and Standards Agency (TEQSA), says detection tools by themselves "do not provide sufficient assessment security" [16]. The OIA puts the burden of proof on the provider, which should consider a range of evidence [11].

The student knows the rules, sees the evidence and explains the work. The OIA says providers should make clear what AI use is acceptable, share all relevant evidence, including detector reports, and let students explain how they prepared the work, for example with notes or drafts [11].

Second-language and other distinctive writing gets a bias check. The OIA asks providers to consider whether assumptions about AI use "could be biased against a student's writing style", for example when English is not the student's first language [11]. TEQSA warns of "unintended bias" in detection tools [16].

Assessment is redesigned. A September 2025 TEQSA resource calls detecting AI use with certainty "all but impossible". One of its three pathways puts at least one secure task, such as an oral presentation, in every unit. It warns that falling back on invigilated exams "may reduce assessment variety and authenticity" [12]. UCT is promoting oral exams, observed group work and assignments in which students reflect on their AI use [5].

5. What to take from this

For teachers and markers. A high score is a reason to look again, ask for notes or drafts, and talk with the student, as Turnitin, Vanderbilt and the OIA advise [4, 11, 17].

For people who run misconduct cases. The OIA's findings and the New York decision point to a minimum: all relevant evidence is shared before any meeting, the student's evidence is weighed, reasons are given, and a different person hears the appeal [7, 10, 11].

For universities and training providers. Where the detector stays, the sources point to a written rule that the score never decides a case alone, a bias check, and assessment of how work was made [2, 5, 11, 16]. Redesign has its own risks: TEQSA warns that falling back on exams may make existing inequities worse [12].

6. Limits of this report

The Stanford paper is an opinion article reporting a small 2023 test that did not include Turnitin, whose own bias test was run by the company [3, 14]. This report uses the journal version; the arXiv preprint gives the average false positive rate as 61.22% [3]. The OIA summaries are short and anonymous. The New York ruling applies one university's code. Decisions to turn detectors off come from the universities' own announcements. Rules elsewhere differ, and no detector was tested for this report.

References

[1] Turnitin (4 April 2023). Turnitin turns on AI writing detection capabilities for educators and institutions (NA and EMEA). Turnitin press release. https://turnitin.ca/press/turnitin-turns-on-ai-writing-detection-capabilities-for-educators-and-institutions

[2] Turnitin. Turnitin's AI writing detection capabilities FAQs. Turnitin Guides. https://guides.turnitin.com/hc/en-us/articles/28477544839821-Turnitin-s-AI-writing-detection-capabilities-FAQs (accessed 16 November 2026)

[3] Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779 (Opinion; online 10 July 2023). https://doi.org/10.1016/j.patter.2023.100779 (journal text read at https://pmc.ncbi.nlm.nih.gov/articles/PMC10382961/; preprint version, with the names of the seven detectors: arXiv:2304.02819v3, https://arxiv.org/html/2304.02819v3)

[4] Vanderbilt University (16 August 2023). Guidance on AI Detection and Why We’re Disabling Turnitin’s AI Detector. Brightspace, Vanderbilt University. https://www.vanderbilt.edu/brightspace/2023/08/16/guidance-on-ai-detection-and-why-were-disabling-turnitins-ai-detector/

[5] Mashinini, G. (24 July 2025). UCT scraps flawed AI detectors. UCT News, University of Cape Town. https://news.uct.ac.za/article/-2025-07-24-uct-scraps-flawed-ai-detectors

[6] Curtin University (4 September 2025). Update on Turnitin AI-Detection Tool. News at Curtin. https://www.curtin.edu.au/news/?p=29208

[7] Office of the Independent Adjudicator for Higher Education (July 2025). AI and academic misconduct - CS072502 [case summary]. https://www.oiahe.org.uk/resources-and-publications/case-summaries/ai-and-academic-misconduct-cs072502/

[8] Office of the Independent Adjudicator for Higher Education (July 2025). AI and academic misconduct - CS072504 [case summary]. https://www.oiahe.org.uk/resources-and-publications/case-summaries/ai-and-academic-misconduct-cs072504/

[9] Office of the Independent Adjudicator for Higher Education (July 2025). AI and academic misconduct - CS072501 [case summary]. https://www.oiahe.org.uk/resources-and-publications/case-summaries/ai-and-academic-misconduct-cs072501/

[10] Matter of Newby v Adelphi Univ., 2026 NY Slip Op 26021, Supreme Court, Nassau County, Index No. 615397/25, decided 28 January 2026. https://www.nycourts.gov/reporter/3dseries/2026/2026_26021.htm

[11] Office of the Independent Adjudicator for Higher Education (July 2025). Casework note: Complaints relating to AI and academic misconduct. https://www.oiahe.org.uk/resources-and-publications/learning-from-our-casework/ai-and-academic-misconduct/casework-note-complaints-relating-to-ai-and-academic-misconduct/ (announced on 15 July 2025 in "OIA publishes a casework note and case summaries relating to AI and academic misconduct": https://www.oiahe.org.uk/resources-and-publications/latest-news-and-updates/oia-publishes-a-casework-note-and-case-summaries-about-artificial-intelligence-ai-and-academic-misconduct/)

[12] Lodge, J. M., Bearman, M., Dawson, P., Gniel, H., Harper, R., Liu, D., McLean, J., & Ucnik, L., with associates (September 2025). Enacting assessment reform in a time of artificial intelligence. Tertiary Education Quality and Standards Agency (TEQSA), Australian Government. https://www.teqsa.gov.au/sites/default/files/2025-09/enacting-assessment-reform-in-a-time-of-artificial-intelligence.pdf

[13] OpenAI (31 January 2023; update of 20 July 2023). New AI classifier for indicating AI-written text. https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/

[14] Turnitin Staff (26 October 2023). New research: Turnitin's AI detector shows no statistically significant bias against English Language Learners. Turnitin blog. https://www.turnitin.ca/blog/new-research-turnitin-s-ai-detector-shows-no-statistically-significant-bias-against-english-language-learners

[15] Office of the Independent Adjudicator for Higher Education. Office of the Independent Adjudicator for Higher Education - OIAHE [home page]. https://www.oiahe.org.uk/ (accessed 16 November 2026)

[16] Tertiary Education Quality and Standards Agency (November 2024). Gen AI strategies for Australian higher education: Emerging practice. TEQSA. https://teqsa.gov.au/sites/default/files/2024-11/Gen-AI-strategies-emerging-practice-toolkit.pdf

[17] Chechitelli, A. (23 May 2023). AI writing detection update from Turnitin's Chief Product Officer. Turnitin blog. https://www.turnitin.ca/blog/ai-writing-detection-update-from-turnitins-chief-product-officer

How to cite

IAPQB (2026). AI writing detectors on campus: how universities, an ombudsman and a court have treated AI detector scores in misconduct cases. IAPQB Report No. 10. International AI Practice Qualification Board. https://iapqb.org/research/010-ai-writing-detectors.html

About this report

Prepared by IAPQB. Each fact comes from the sources listed, and each source was opened and checked on 16 November 2026; anything that could not be checked was left out. AI tools were used to search and to draft, and IAPQB checked every claim against its source and is responsible for the text. IAPQB runs an AI work qualification; this report does not promote it. Corrections: contact@iapqb.org or an issue at github.com/iapqb/iapqb.github.io. Licence: CC BY 4.0.

Archived with the series on Zenodo: doi.org/10.5281/zenodo.24248199 · Markdown sources: github.com/iapqb/research