One in five master’s theses largely written by AI, time to assess differently
RTL Nieuws investigated hundreds of recent master's theses from twelve Dutch universities using the detection tool Pangram. In about one in five theses, the contribution of ChatGPT, Claude or Gemini is greater than that of the student. Six percent are almost entirely written by AI.
Minister Letschert calls this fraud and wants universities to intervene. Students themselves told RTL that "everyone" does it.
I am not surprised by this. I previously wrote that we should have students explain their work orally. Not to ban AI, but because AI reveals how vulnerable our current method of assessment is. In this piece, I explain why, and what the science says about it.
How reliable are these figures
That question applies to any research using an AI detector. The first generation of detectors turned out to be unreliable. Debora Weber-Wulff and colleagues tested fourteen tools in 2023 and concluded that they were not accurate and not reliable enough.
Researchers at Stanford showed that same year that detectors wrongly labelled texts by non-native speakers as AI, on average in 61 percent of essays (Liang et al., 2023). A painful outcome for an international university.
Pangram, the tool used by RTL, scores much better in more recent research. Economists at the University of Chicago found almost no false positives in 2025, even with texts edited using so-called humanizers. Sociologists at the Vrije Universiteit Brussel reached a similar conclusion in 2026.
But those same Brussels researchers explicitly warn that a detector should never be the sole evidence in major decisions. So the RTL figures are a serious signal about the sector. They prove little about an individual student. And that is precisely the problem.
The real problem lies in the assessment
A thesis was meant to demonstrate that a student can conduct research and substantiate it independently. We assess a document, while what we actually want to assess is a skill. As long as that document is the main evidence, it says less and less about what the student can do themselves.
This is not just a theoretical risk, as the University of Reading showed in 2024. Researchers submitted exam work on behalf of 33 fictitious students, all written entirely by GPT-4. No less than 94 percent went unnoticed. The AI answers even received higher marks on average than the work of real students (Scarfe et al., 2024).
Australian assessment researchers Phillip Dawson, Margaret Bearman, Mollie Dollinger and David Boud therefore focus on a different question. Not how to prevent fraud, but how to ensure an assessment measures what it is supposed to measure. Validity, they say, is more important than fighting fraud (Dawson et al., 2024).
Corbin, Dawson and Liu go a step further in 2025. They analysed AI guidelines from educational institutions and saw that they mainly consist of rules and agreements with students. They call this an enforcement illusion. Rules you cannot check do not protect the value of a diploma. What does help are structural changes in the assessment itself.
Three ways to assess differently
The solution is therefore not more checking for AI use, but different assessment. Three steps that any programme can take tomorrow.
Have students justify their choices during the research. Discuss methodological choices at fixed points in the process. This makes the thinking visible before there is a final text.
Assess not only the final product, but also the process and substantiation. Programmatic assessment shows how this works. Van der Vleuten and colleagues described in 2012 how to base major decisions on many measurement moments instead of one product.
Have students defend their research orally. Ask critical follow-up questions and probe choices, contradictions and conclusions. Research at Griffith University shows that interactive oral assessments protect academic integrity and at the same time develop skills demanded by the professional field (Sotiriadou et al., 2020).
This argument does not stand alone. The Australian report Assessment reform for the age of artificial intelligence, written by Lodge, Howard, Bearman and Dawson on behalf of regulator TEQSA, comes to the same conclusion. Reliable judgements about what students can do require multiple and varied moments of evidence.
And what about AI
AI does not belong outside, but in the right place. Use it especially in preparation. Have students critically examine their own arguments and practise their defence.
But with an important condition. In a large field experiment with nearly a thousand pupils, those who practised with a standard chatbot performed 17 percent worse without AI afterwards than pupils who had never used AI. A version that gave hints instead of answers largely eliminated that negative effect (Bastani et al., 2025).
The difference is therefore not in the technology, but in how it is designed. That is why we built EduGPT with its own critical thinking layer. It does not provide ready-made work, but asks questions in return. Why did you choose this method? What is the strongest counterargument? How do you explain this outcome?
Exactly the kind of questions a student should be able to answer during an oral defence. Guidance instead of delivery.
Ask a different question
Stop asking whether this student has used AI. Ask whether this student can demonstrate mastery of research skills.
We do not need to wait for better AI detectors. Especially in an age of AI, we need to develop better ways to assess what students can really do.
I am curious how your programme deals with this. Will the thesis remain as it is, or is this the start of a real conversation about assessment? If you would like to explore how AI can support preparation for an oral defence, please feel free to contact us. We are happy to think along with you.
Pascal Mariany is senior lecturer at Hogeschool Utrecht, founder of CivAI and founder of Onderwijs Orakel.
Sources
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. and Mariman, R. (2025). Generative AI without guardrails can harm learning. Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26). doi.org/10.1073/pnas.2422633122
Corbin, T., Dawson, P. and Liu, D. (2025). Talk is cheap. Why structural assessment changes are needed for a time of GenAI. Assessment & Evaluation in Higher Education, 50(7), 1087-1097. doi.org/10.1080/02602938.2025.2503964
Dawson, P., Bearman, M., Dollinger, M. and Boud, D. (2024). Validity matters more than cheating. Assessment & Evaluation in Higher Education, 49(7), 1005-1016. doi.org/10.1080/02602938.2024.2386662
Jabarian, B. and Imas, A. (2025). Artificial writing and automated detection. Working paper 2025-116, Becker Friedman Institute, University of Chicago. bfi.uchicago.edu
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E. and Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. doi.org/10.1016/j.patter.2023.100779
Lodge, J. M., Howard, S., Bearman, M. and Dawson, P. (2023). Assessment reform for the age of artificial intelligence. Tertiary Education Quality and Standards Agency (TEQSA). teqsa.gov.au
RTL Nieuws (14 October 2026). Research into AI use in master's theses at twelve Dutch universities, as reported by the Reformatorisch Dagblad. rd.nl
Scarfe, P., Watcham, K., Clarke, A. and Roesch, E. (2024). A real-world test of artificial intelligence infiltration of a university examinations system. A "Turing Test" case study. PLOS ONE, 19(6), e0305354. doi.org/10.1371/journal.pone.0305354
Sotiriadou, P., Logan, D., Daly, A. and Guest, R. (2020). The role of authentic assessment to preserve academic integrity and promote skill development and employability. Studies in Higher Education, 45(11), 2132-2148. doi.org/10.1080/03075079.2019.1582015
Van der Vleuten, C. P. M., Schuwirth, L. W. T., Driessen, E. W., Dijkstra, J., Tigelaar, D., Baartman, L. K. J. and Van Tartwijk, J. (2012). A model for programmatic assessment fit for purpose. Medical Teacher, 34(3), 205-214. doi.org/10.3109/0142159X.2012.652239
Van Vlasselaer, M., Van Droogenbroeck, F. and Spruyt, B. (2026). Who wrote this? Evaluating the reliability of AI detection tools in higher education. International Journal for Educational Integrity, 22(1). doi.org/10.1007/s40979-026-00226-w
Weber-Wulff, D., Anohina-Naumeca, A., Bjelobaba, S., Foltýnek, T., Guerrero-Dib, J., Popoola, O., Šigut, P. and Waddington, L. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19(1). doi.org/10.1007/s40979-023-00146-z
Image for illustration, created with AI