AI Detection Is Banned at Some Dutch Universities — Why, and What It Means for You

AI Detection Is Banned at Some Dutch Universities — Why, and What It Means for You

 

AI detector banned at Dutch universities — detector tool with a not allowed sign and a student showing draft history

In October 2026, Dutch broadcaster RTL Nieuws reported that about one in five recent master's theses at Dutch universities was mostly written by chatbots. Days later, Times Higher Education reported that three major universities — Utrecht, Leiden and TU Delft — had rejected AI-detection tools. The headline writes itself: Dutch universities ban AI detectors.

But the real story is subtler, and more useful for students. None of the three universities uses the word "ban" on its own pages. What they actually say is narrower and more interesting: a detector score cannot prove fraud, so it cannot be used as evidence against a student. This article explains what the investigation found, what each university actually said, why detectors fail the test — and, most importantly, what replaces detection: your drafts, your prompts, and your ability to explain your work out loud.

The probe that started it

On 14 October 2026, RTL Nieuws published an investigation by Eric van den Berg and Maaike Kempes. The team took a 3,000-word sample from each of hundreds of recent master's theses from 12 Dutch universities and ran them through Pangram, a commercial AI detector that recognises text from ChatGPT, Claude and Gemini. The headline claim: in one in five theses, "the contribution of ChatGPT, Claude, or Gemini is greater than that of the student."

The breakdown was:

  • 6% of theses were almost entirely AI-written
  • 14% were more than half AI-written
  • 39% were less than half AI-written
  • 41% showed no AI-detected text

So about one in five (20%) was flagged as more than half AI-written, and 59% showed at least some AI-detected text. The story made national news, and education minister Rianne Letschert reacted sharply: using AI to write a thesis amounts to "fraud", she said, because chatbot-generated text is "not a student's own work".

Why the 20% figure is not a fraud rate

Before going further, the number needs its caveats — and they are large:

  1. Only a 3,000-word sample was scanned from each thesis. A master's thesis is typically 10,000–20,000 words, so the AI share in the rest of each thesis is unknown, and the sample may not be representative of the whole document.
  2. Only theses published online were included, from 12 of the 14 Dutch universities. VU Amsterdam and Maastricht University were excluded because their theses are not published online. This is not a census.
  3. The exact sample size was never published — only "hundreds" of theses. The list of the 12 universities, the Pangram version used, and the threshold defining "almost entirely" were not published either.
  4. A detector score is a statistical pattern estimate, not proof of authorship. A high score means the text resembles AI-generated writing; it does not prove a chatbot wrote it, and it cannot say how the text was produced.
  5. Neither the universities nor their sector body keeps figures on AI-related thesis fraud, so there is no official baseline to compare the 20% against.

The 20% figure is a detector-based media finding, not an official fraud statistic. Keep that distinction in mind — it is the same distinction the universities themselves draw.

What each university actually says

Utrecht University

Utrecht's Centre for Academic Teaching and Learning states the position plainly on its AI focus-points page:

"None of the current AI detectors are reliable, and it does not look like they will be anytime soon. This means they cannot be used as official evidence, because they generate too many false positives and false negatives."

Utrecht makes a second point that is easy to miss: detectors only check whether a specific chatbot has been used, not to what extent. "It is not clear whether a detector can identify whether a piece of writing has been written entirely by a chatbot or whether it has only been spell-checked." A score cannot distinguish heavy editing help from a fully generated text.

What Utrecht does instead: "If you suspect fraud, you can, of course, always talk to a student and/or the whole group." Lecturers are advised to adapt assignments — for example, asking students to keep a logbook of how the work was created, or to submit an outline first — and to state per assignment how much AI use is allowed, using the university's five-level "AI index" scale.

For students, the practical consequence is spelled out in the Faculty of Humanities guidelines: save your versions, log your GenAI use, save your prompts, and hand them in on request. Teachers may ask you to explain your work orally if they suspect GenAI use, and random spot-check discussions can happen even without suspicion. Note what is absent: no official Utrecht page uses the word "ban", and the university's central AI policy hub does not mention detectors at all.

Leiden University

Leiden's position comes from the Faculty of Humanities — an important scope note, because it is faculty wording, not a university-wide rule. Its staff page on fraud and plagiarism says:

"There's a growing number of tools aimed at detecting GenAI. Unfortunately at present, and probably also in the near future, these tools do not meet a standard that justifies confronting the student with the result."

And: "Results should therefore always be interpreted with caution and should not be used as conclusive evidence of AI use."

Leiden is precise about what Turnitin can and cannot do: "Turnitin is not suitable for detecting GenAI-produced texts, but it can be helpful in checking for sources that were fabricated by GenAI." Turnitin remains in use for plagiarism and similarity checking, not for GenAI detection.

The faculty's expertise centre, ECOLe, tested the tools itself. Its findings cut both ways: false positives were "rare in most tools", but AI passages were often missed and the wrong passages were flagged. On that basis it does not recommend Grammarly, Scribbr AI Detector, GPTZero, HumanizeAI.pro, ZeroGPT or Paraphraser.io — partly for reliability, and partly because of legal concerns about uploading student work to commercial AI tools.

There is one striking exception. In ECOLe's own tests, Pangram — the same detector RTL used — detected many AI-generated texts, and "human-written texts were never incorrectly identified as AI-generated." Even so, Leiden does not use it as proof: the faculty is only "planning to launch a pilot in 2026-2027 with a group of instructors to further investigate its use." The Times Higher Education wording that Leiden "does not allow professors to use such tools" is the magazine's paraphrase; the faculty's own words are the cautious ones quoted above.

TU Delft

TU Delft's position is published as Teaching Support guidance for educators, not as a regulation. Its page on AI detection tools states:

"Research has shown that these tools are often inaccurate and can drive students to adopt basic 'evasion' prompting techniques that further reduce their accuracy (Liang et al., 2024)."

And:

"Relying on AI-detection tools to penalise students risks false accusations, unfair treatment, and a breakdown of trust between staff and students… false positives are common and untestable. Using them as a policing mechanism is therefore neither effective nor equitable in way of upholding assessment integrity. Instead, it is recommended that educators focus on adapting assessments to be more AI-resilient."

The research cited is Liang et al. (2024), which found that GPT detectors "consistently misclassify non-native English writing samples as AI-generated" and that simple prompting can bypass them — a finding that matters enormously in Dutch universities, where many students write in English as a second language. No official TU Delft page uses the words "banned" or "opted not to use". The accurate formulation is that the university's guidance advises educators not to use detectors as a policing or penalising mechanism.

What the sector body said

Universiteiten van Nederland, which represents the 14 Dutch universities, has no published statement or position paper on AI detectors on its website. What exists is a press-reported quote, carried by Times Higher Education: "Based on a specific score from a detection tool, it cannot yet be determined whether a student is guilty of fraud." The same reporting notes that changes to adapt to AI were "well under way" and that AI rules vary by programme — "AI plays a different role in a medical programme than in computer science or a language study."

So the minister and the universities are not saying the same thing. Letschert called AI-written thesis work "fraud". The universities say a detector score alone cannot prove fraud. Both statements can be true at once — and the gap between them is exactly where students now live.

Why detectors fail the test

The three universities' objections add up to a coherent case:

They are unreliable in both directions. False positives wrongly flag human writing; false negatives miss AI writing. Leiden's own tests illustrate the two-sided failure: false positives were rare in most tools, but AI passages were often missed entirely. A tool that misses AI text is as useless as evidence as one that flags human text.

They cannot measure extent. A detector returns a probability score for a passage. It cannot tell a lecturer whether a student used AI to fix grammar, to brainstorm, or to write the whole thesis — and those are very different offences under university rules. Utrecht makes exactly this point.

They can be evaded. TU Delft cites research showing that students adopt "evasion" prompting techniques — asking the chatbot to write in a more human style, for example — which further reduces detector accuracy. A detection system that teaches students to hide better is a poor integrity tool.

They are biased against non-native writers. The Liang et al. study cited by TU Delft found detectors consistently misclassify non-native English writing as AI-generated. An earlier Stanford study found detectors flagged 61.3% of TOEFL essays by non-native English writers as AI-written. In the Netherlands, where international students write theses in English as a second language, that bias is not a footnote — it is a fairness problem. We covered the evidence in detail in our guide to whether AI detectors are accurate: https://aistudenttoolsblog.blogspot.com/2026/11/are-ai-detectors-accurate-what-students.html

Vendor claims are not independent evidence. Pangram publishes impressive self-benchmarks — a 0.009% false-positive rate on 781,981 human documents and a 0.365% false-negative rate on 520,000 AI documents. Those are the vendor's own numbers, measured on the vendor's own test sets. They do not settle whether the tool is reliable on real student theses, in real courses, with real consequences.

They break trust. TU Delft's argument is the most human one: using an unreliable tool to police students "risks false accusations, unfair treatment, and a breakdown of trust between staff and students." A university that accuses students on the basis of a score it cannot defend has a bigger problem than cheating.

Student explaining thesis draft in oral defence to two examiners

What they do instead

The alternative is not "do nothing". It is a shift from detecting a product to examining a process:

Talk to the student. All three universities start here. Utrecht's advice is direct: if you suspect fraud, talk to the student or the whole group. A conversation about how a thesis was made is something a detector score can never replace.

Keep process evidence. Utrecht asks students to save versions, log GenAI use and save prompts. Leiden's guidance is similar: "Do not only assess the final product, but also the process… Have them disclose their use of AI (AI logbook)." Students are expected to keep a record of prompts and all GenAI output, and to hand it over on request. Revision history — for example, replaying a document's edit history — can support a student, but only voluntarily: there is no data-processing agreement that would let a university demand it.

Oral defence. Leiden's Assessment Methods and AI Toolkit recommends combining written work with oral presentations and portfolios, including a capstone with a 10-minute unscripted oral defence and a mid-course draft with a reflection paragraph. TU Delft uses oral design defences — asking a student "Why did you choose this algorithm over the AI-recommended one?" — and in-class labs where students explain their work live. Theses are staged and supervised, with oral presentations at kick-off, green-light and finalisation stages, and an AI disclosure statement is required; failing to disclose AI use may itself count as fraud.

Design assessments AI cannot do alone. TU Delft's guidance is to "design AI-resilient assessments": grade how the student modified and critiqued AI output, not just the output itself. Leiden's toolkit makes the same point — oral exams are "inherently AI-resilient" because "AI cannot be accessed during live interaction".

None of this is new technology. It is old academic practice — drafts, discussion, defence — brought back because the new technology broke the shortcut.

What it means for you

If you study in the Netherlands, or anywhere detectors are used, here is what this story actually changes:

A detector flag is not proof — but it is not a free pass either. At Utrecht, Leiden and TU Delft, a score alone cannot be used as evidence or as a basis to confront you. But GenAI-written work presented as your own is still fraud under university rules, and the education minister has said so publicly. Do not read "detectors are banned" as "AI use is allowed".

Process evidence matters more than ever. The universities' alternative to detection is your paper trail. Keep every draft version. Keep a log of your prompts and the AI output you received. Save the disclosure statement you submitted. If a lecturer ever asks how a piece of work was made, that trail is your answer — and it is also the best protection against a false accusation.

A Turnitin AI score is a starting point for a conversation, not a verdict. If your work is flagged, ask what happens next, bring your drafts and prompt log, and be ready to explain your work orally. Our guide to what a Turnitin AI score means and what to do about it walks through the thresholds and the steps: https://aistudenttoolsblog.blogspot.com/2026/11/turnitin-flagged-your-essay-as-ai-what.html — and our guide to how universities detect AI writing explains the other methods — oral defences, process portfolios and disclosure — that now matter more than any score: https://aistudenttoolsblog.blogspot.com/2026/11/how-universities-detect-ai-writing-in.html

Practical checklist:

  1. Save every draft version of essays and theses — do not overwrite them.
  2. Keep a prompt log: what you asked the AI, and what it returned.
  3. Disclose AI use exactly as your course requires — and check the requirement per assignment, because it varies.
  4. Never submit AI-generated text as your own work.
  5. If you are flagged, ask for the conversation and bring your evidence. A score is not a verdict.

FAQ

Is it true Dutch universities banned AI detectors?

Not in the way the headline suggests. "Banned" is Times Higher Education's word. What Utrecht, Leiden and TU Delft actually say on their official pages is that detector scores are unreliable and cannot be used as evidence of fraud or as a basis to confront a student. Leiden's wording is faculty-level (Humanities), and TU Delft's is guidance for educators, not a regulation.

Can a detector score fail me?

Not on its own at these three universities — a score alone cannot prove fraud. But a score can start a conversation, and what matters then is whether you can show how the work was made: drafts, version history, prompts and disclosure. A missing paper trail plus a high score is a much weaker position than a high score with full process evidence.

Does this mean I can use AI freely?

No. At Utrecht, Leiden and TU Delft, using GenAI to write gradable work and presenting it as your own is fraud. The rules vary by programme and by lecturer — some assignments allow AI for brainstorming or editing, others allow none — so check the requirement for each assignment.

What should I keep as proof of my own work?

Draft versions with dates, your prompt log and the AI output you received, your AI disclosure statement, and any outline or logbook your course asked for. Keep them in one folder per assignment, so you can produce them quickly if asked.

Why don't universities just use a better detector?

Because the problem is not only accuracy. Even Pangram — the detector RTL used, and the best performer in Leiden's own tests, where human texts were "never incorrectly identified" — is only being piloted, not used as proof. A score, however accurate, is a statistical estimate about a text. It cannot say who wrote it, how it was written, or what the student understands. That is why the universities went back to drafts, discussion and defence.

Bottom line: Dutch universities did not declare AI detection useless — they declared it unfit as evidence. The investigation that scared everyone produced a number about a tool, not a verdict about students. What protects you now is not a better detector. It is the boring, old-fashioned evidence that you did the work: your drafts, your prompts, and your ability to explain them.

Related guides