Turnitin has been the standard plagiarism checker in UK higher education since the early 2000s, and 2013, 98% of institutions adopted it [1]. This autumn, students and universities are protesting major changes to how Turnitin works [4], and the quuestion arises whether AI detectors are a good idea in education in the first place.
The first issue is about how Turnitin handles students' submissions and that students are forced to consent. In August, York's students' union objected to warned about changes to Turnitin's licence that could let student submissions be used for AI training. Unions at Lancaster, St Andrews, Reading and East Anglia raised similar concerns [2]. While Turnitin says it will not train its AI assistant on student work, but that it "may use anonymized student submissions to improve our detection and assessment tools" [2]. York's union pointed out that students cannot freely consent, because submitting the essay is a condition of being marked [2]. No consent, no grade. Turnitin has postponed any change until at least September 2027 [2]. Still, Southampton has decided not to renew its contract, saying Turnitin "has not fully clarified what future changes may be imposed" [2, 3].
The second complaint is about the inherent issues with AI detectors. A detector reads a finished piece of text and estimates how likely it is to be AI-written. In most cases, the score is given for the overall document, in better systems, subsections can be scored individually. The evidence that this works reliably is weak. OpenAI announced its own detector in January 2023, it correctly identified 26% of AI-written text and wrongly labelled 9% of human-written text as AI-written, leading them to eventually withdraw the tool in July 2023 because of its "low rate of accuracy" [5].
Detectors also make more mistakes for some writers than others. A 2023 test across seven detectors found that writing of non-native English speakers was labelled as AI-generated 61% of the time [6, 7], although Turnitin argued that its detector shows no statistically significant bias against English language learners on documents above its minimum length [8]. For a student who is accused, the score is hard to challenge. Even if Turnitin acknowledge that detection scores should not be the sole basis for misconduct proceedings [9], the Higher Education Policy Institute reported in July about cases from 2025 where students were wrongly accused and penalised without a chance to provide evidence in their defense.
The underlying problem is that the finished text in itself is not a viable proof of human effort. Take student A, who generated an entire report using ChatGPT, then shuffled the words just enough so it passes the detector checks. Then take student B, who hand-wrote the report, but used ChatGPT to improve the language before submission, as international students often do [12]. Student B worked and learned more, yet they are more likely to be penalised. What would settle the question is a record of how the document was written: what was typed, what was pasted, and what an AI assistant contributed.
At Orient, this is the solution we are working on. A collaborative authoring workspace with built-in AI assistance, where every change to a project is saved with its source: typed by a person, pasted in, made by the AI assistant, or made by the system. Then, a provenance report summarises this as a human-versus-AI split, a timeline per author, and a count of AI edits by type. We think the writer should have the option to choose whether edits and AI chats are included and neither their documents, nor their AI conversations are used to train AI models. If the institution hosts the system, neither of that data ever even reaches a third party. We do not allow requiring full provenance recording at an organisation level and is not in support of forcing students to consent to surveillance or data processing that they do not want. Further, provenance reports are not based on a statistical model, but a provable track of system interactions, changes to documents. It allows evaluators to interpret the histories for themselves instead of casting judgement directly. More importantly, it equips students to defend their own work with the very same provenance reports. We believe that opt-in transparency is the way forward both with human-only and human-AI co-authoring. Not only in education, but in academic research, peer-reviews, engineering, medicine and other parts of life.
Of course, provenance technology also has its limits. It cannot show that no AI was involved, because text written elsewhere and pasted in appears as a paste, not as content with a known origin [11], but these limits are temporary as AI text watermarking is slowly being introduced, and as limitations about AI use are becoming more lax, there is less reason to cheat TeXposit's provenance and more reason to use the built-in AI tools more responsibly.
References
[1] Times Higher Education. 2013. Turnitin is turning up fewer cases of plagiarism. Retrieved September 30, 2026 from https://www.timeshighereducation.com/news/turnitin-is-turning-up-fewer-cases-of-plagiarism/2002939.article
[2] Resultsense. 2026. Student unions resist Turnitin's AI training licence change. Retrieved September 30, 2026 from https://www.resultsense.com/news/2026-09-24-turnitin-student-unions-ai-training/
[3] Times Higher Education. 2026. Southampton dumps Turnitin over use of students' work to train AI. Retrieved September 30, 2026 from https://www.timeshighereducation.com/node/744508
[4] CNN. 2026. Some UK students want colleges to drop Turnitin. Retrieved September 30, 2026 from https://www.cnn.com/2026/09/23/business/video/turnitin-students-college-ai-vrtc
[5] Decrypt. 2023. OpenAI quietly shutters its AI text classifier due to low accuracy. Retrieved September 30, 2026 from https://decrypt.co/149826/openai-quietly-shutters-its-ai-text-classifier-due-to-low-accuracy
[6] Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu, and James Zou. 2023. GPT detectors are biased against non-native English writers. Patterns (2023). Retrieved September 30, 2026 from https://scale.stanford.edu/publications/gpt-detectors-are-biased-against-non-native-english-writers
[7] Times Higher Education. 2023. AI text detectors "biased against non-native English speakers". Retrieved September 30, 2026 from https://www.timeshighereducation.com/node/727071
[8] Turnitin. New research: Turnitin's AI detector shows no statistically significant bias against English language learners. Retrieved September 30, 2026 from https://www.turnitin.com/blog/new-research-turnitin-s-ai-detector-shows-no-statistically-significant-bias-against-english-language-learners
[9] Higher Education Policy Institute. 2026. Catching the wrong students: AI detection, international students and the fairness crisis in UK universities. Retrieved September 30, 2026 from https://www.hepi.ac.uk/2026/07/20/catching-the-wrong-students-ai-detection-international-students-and-the-fairness-crisis-in-uk-universities/
[10] TeXposit. Provenance and authorship certification. Retrieved September 30, 2026 from https://texposit.com/guide/provenance
[11] TeXposit. 2026. AI detectors are wrong often enough to ruin careers, and nobody can appeal a probability. Retrieved September 30, 2026 from https://texposit.com/blog/ai-detector-false-positives
[12] Arslan N, Haj Youssef M, Ghandour R (2025), "AI and learning experiences of international students studying in the UK: an exploratory case study". Artificial Intelligence in Education, Vol. 1 No. 1 pp. 1–23, doi: https://doi.org/10.1108/AIIE-10-2024-0019