Pangram classified a fifth of examined Dutch master's theses as largely written by AI

An AI detector such as Pangram produces a screening signal, not proof that a student committed fraud. Universities may only classify a thesis as fraudulent after independent human assessment: access to versions, sources, defence and the student's explanation, recording which detector, language and threshold were used.

The occasion is research in which recent master's theses from twelve Dutch universities were scanned by the detector Pangram. According to the reporting on this research by RTL Nieuws, roughly a fifth of the theses examined were classified as largely written by AI, and 41% contained no detectable AI text. Education minister Rianne Letschert called undeclared, AI-written thesis text fraud and called on universities to intervene.

Our assessment: the core of this matter lies not in the percentage, but in the question beneath it. A detector result is a statistical estimate, not a confession. Anyone who confuses these two turns a useful triage instrument into an unreliable verdict.

What exactly did the research into AI in Dutch master's theses establish?

The research scanned hundreds of recent master's theses from twelve universities with Pangram and classified a portion as almost entirely AI-written and a larger portion as more than half AI text. Importantly: this is a detector-based screening, not a count of fraud cases confirmed by universities. The figures come from the reporting on the research; the detector estimates probabilities, it does not prove authorship.

Our interpretation of the institutional context is as follows: this makes clear why this is sensitive. Universiteit Utrecht states in its official thesis guideline that a thesis must be entirely one's own work and that having part written by a chatbot counts as fraud, with referral to the examination board. At Universiteit Leiden, the rulings on thesis fraud involving generative AI show that such disputes are already reaching the formal appeal procedures.

Why is a detector score not proof of fraud for an individual thesis?

Because detectors do not establish reliable individual authorship attribution, and because their results vary strongly by tool, language and text length. A score is a detector’s probabilistic or classificatory estimate, not reliable evidence by itself of what this student actually did.

A peer-reviewed study by the Vrije Universiteit Brussel tested four detection tools on academic texts. In it, Pangram performed best in that test, but the tools differed considerably in their estimated AI percentages. When Pangram was applied to over a thousand master's theses with no known ground truth, almost half received a flag, mostly pointing to low to moderate levels. A recent peer-reviewed analysis of the pitfalls of AI detection in academic writing points to classification errors, differences between tools and possible skewed false positives, and concludes that theoretical detectability does not yield reliable attribution to a specific student.

Here the sources visibly clash: the guidelines treat AI text in a thesis as fraud, while the methodological research emphasises that a detector score on its own does not establish authorship. In our assessment, that is not resolved by trusting or rejecting the detector, but by using it for what it is suited to: as a trigger for investigation, not as a final judgement.

What does this mean for administrators, examination boards and lawyers at universities?

Our analysis: because a detector result has no ground truth, an examination board that bases a fraud judgement directly on it may run into evidential problems on appeal; therefore every university should now record that a score only opens an investigation and is never itself the outcome. Because the available institutional guidance and ministerial response point to a need for clear and consistent expectations, in our analysis, it is prudent for administrations to publish an institution-wide regulation for the coming thesis season with a proportionate AI declaration and clear rules about what is permitted. Because detectors may show skewed false positives for certain language groups or text lengths, in our analysis, lawyers and examination boards should guard against discriminatory effects by recording that no decision rests on a single automated score and by giving the student an opportunity to explain. Because Leiden’s official records include AI-related thesis-fraud cases in formal appeal proceedings, the presentation of evidence is, in our analysis, a central point of attention. We therefore advise universities to document, where a detector is used, its version, the language, the text fragment and the threshold, so that a later decision is more traceable.

In summary, this yields a short checklist:

• Treat every detector result solely as grounds for human investigation.
• Publish one institution-wide regulation with a proportionate AI-use declaration.
• Train supervisors and examination boards in reading and weighing detector output.
• Record per case which detector, version, language, text fragment and threshold were used.
• Offer a fair appeal route with access to the underlying material.

How does a university set up a fair, human assessment of AI suspicions?

By combining the detector flag with human investigation into how the text came about, and recording that process reproducibly. The aim is not to confirm a figure, but to understand how the thesis came to be.

Our recommended assessment process combines the detector flag with a documented human investigation into how the text came about. It can consider, in a proportionate order: the student’s draft versions, research notes and use of sources; the oral defence; the disclosure; the student’s explanation; and a documented human judgement with reasons.

1. The student's draft versions, research notes and use of sources.
2. The oral defence, in which substantive command is demonstrated.
3. The disclosure: what did the student themselves declare about AI use?
4. The student's explanation of the specific flag.
5. A documented, human final judgement with reasoning.

Our analysis places this approach within broader principles of responsible AI governance: human oversight of automated decisions must be demonstrable. Anyone wishing to test whether an automated result is reproducible at all finds a foothold in the question of how many repeats a reliability test truly requires. And just as with an AI summary of a legal file, the same applies: the machine delivers a signal, the responsible professional makes the decision.

Our conclusion: the newsworthiness is that a detector flagged AI-associated text on a large scale in a sample of Dutch theses. The task that follows from it is modest but firm: respond decisively to that signal, without replacing academic judgement with an opaque automatic attribution.

Sources: The article relies on reporting about the RTL research, official guidelines from Universiteit Utrecht and Universiteit Leiden, and peer-reviewed studies from the Vrije Universiteit Brussel and in AI and Ethics (Springer).