Pangram classified a fifth of Dutch theses

You must now treat any AI detection score as a screening signal that triggers investigation, never as proof of authorship or fraud. Before classifying a thesis as fraudulent, you must conduct independent human assessment: examine draft versions, sources, the student's defence and their explanation, and record which detector, language and threshold were used.

The prompt is an analysis of 14 October 2026 of AI detection scores in academic writing, which argues that a detector result is a statistical estimate, not reliable evidence of individual authorship. The occasion is research in which a detector scanned master's theses from twelve Dutch universities and classified roughly a fifth as largely AI-written. In our assessment, the core issue is not the percentage flagged, but the methodological gap between what a detector can signal and what it can prove. Confusing these two turns a useful triage tool into an unreliable verdict.

Why detector scores alone cannot establish fraud

A detection tool produces a probabilistic or classificatory estimate, not reliable attribution of authorship to a specific student. Peer-reviewed testing shows that tools differ considerably in their estimated AI percentages when applied to the same texts. When one detector 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 of detected AI text. Research into the pitfalls of AI detection in academic writing points to classification errors, differences between tools and possible skewed false positives, concluding that theoretical detectability does not yield reliable attribution.

Because detector results have no ground truth, an examination board that bases a fraud judgement directly on a score may run into evidential problems on appeal. The institutional context sharpens this risk: official thesis guidelines at multiple Dutch universities state that undeclared AI-written text counts as fraud, whilst those same disputes are already reaching formal appeal procedures. That collision between institutional policy and methodological reality 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 risks does automated classification introduce?

• Evidential weakness on appeal — a fraud finding resting solely on a detector score lacks the ground truth needed to withstand institutional challenge.
• False positives skewed by language or text length — detection tools may show discriminatory effects on certain populations or document types.
• Tool and threshold variation — the same text receives different scores from different detectors and at different sensitivity settings.
• Confusion of signal with proof — treating a probabilistic estimate as reliable authorship attribution.
• Absence of student explanation — automated classification bypasses the student's opportunity to account for how the text came about.

Which concrete controls must you now demonstrate?

1. Record the detector used, its version, the language setting and the threshold — document these parameters so that any later decision is traceable and reproducible.
2. Examine draft versions, research notes and source use — establish the student's working process before and independently of the detector flag.
3. Conduct the oral defence and record its content — the student's ability to defend the thesis and explain its development is material evidence.
4. Document the student's explanation — give the student a formal opportunity to account for how the text came to be and record their response.
5. Record a human judgement with reasons — ensure that no decision rests on a single automated score and that the examination board's reasoning is documented.

How should you structure the investigation?

Combine the detector flag with a documented human investigation into how the thesis came about. Consider, in 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. The aim is not to confirm a detector figure, but to understand how the thesis came to be. This approach embeds the detector within broader principles of responsible AI governance: human oversight of automated decisions must be demonstrable, and the machine delivers a signal whilst the responsible professional makes the decision.

What should your institution publish now?

Because the available institutional guidance and ministerial response point to a need for clear and consistent expectations, 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. This regulation should specify that a detector score opens an investigation and is never itself the outcome. It should also guard against discriminatory effects by making clear that no decision rests on a single automated score and by ensuring the student has an opportunity to explain.

Tooling can carry the screening, the documentation and the reproducibility of the process. What remains your own professional judgement is whether the investigation you have conducted answers the question of how the thesis came to be, and whether the student's explanation, combined with the evidence you have gathered, supports a finding of fraud or permits the thesis to proceed.