← All posts

AI Cheating at University: Everyone Suspects It, Nobody Can Prove It

Swiss universities can't prove AI cheating — detectors are unreliable. The verifiable alternative: check the citations, not the intent.

6 min read AI detectorscitation checkingacademic integrityhallucinated sourcesthesis

On 22 June 2026, Swiss public broadcaster SRF published a story with a headline that quietly captures the whole problem of AI in higher education: students may be cheating with AI — but almost nobody can prove it.

The numbers reporters Fiona Zellweger and Pascal Albisser gathered are strikingly small. PH Luzern has recorded 27 cases since the start of 2023. ZHAW counts 17 in the past two years. ETH reports only a low single-digit number per year. In an era where every student has a free chatbot in their browser, those figures are not a sign that cheating is rare. They are a sign that it is barely being measured.

The detection problem

At least 11 of the 31 Swiss universities surveyed use AI detectors — and they add the same caveat: the tools are not reliable. An AI detector does not produce evidence. It produces a probability, an “AI-likelihood” score that can flag a careful human writer as a machine and wave a machine-polished essay through as human. You cannot fail a student, or clear one, on a guess.

So universities are stepping back from the detector arms race. Berner Fachhochschule (BFH) deliberately does not count AI cases at all, treating AI use as a “question of competence” rather than a disciplinary offence. The University of St. Gallen (HSG) frames it as principle: “The use of AI should support learning, not replace it. The decisive element remains the student’s own academic work.” The University of Lausanne (UNIL) puts it bluntly: it is “fairly illusory to complete an entire degree programme by delegating everything to AI.”

The emerging consensus is transparency obligations, oral examinations, and trust — not technical surveillance. That is sensible. But it leaves a gap. If you can’t prove intent, what can you check?

Flip the question: don’t prove intent, verify the citations

Here is the move. Stop trying to prove how a text was written. Check whether what it says is true to its sources.

This is fundamentally different from detection, because it is objective. A claim either is or is not supported by the source it cites. A reference either exists or it does not. There is no “73% likely” — there is checkable evidence. And crucially, this question is fair to everyone in the room. It does not accuse anyone of misconduct. It simply asks the question every examiner asks anyway: does the source actually say what the citation claims?

That reframing also protects the honest researcher — the person AI detectors punish most. If you used an AI assistant to help organise your literature and it quietly handed you a citation that does not exist, or one that points to a real paper which never made the argument you attributed to it, you are now carrying a defect you never intended and may not even know about. A detector would only tell you your prose “looks AI-written”. It would never find the actual error. Verification does.

The two kinds of AI citation damage

When AI touches a bibliography, the damage comes in two distinct shapes:

  1. The fabricated source. Author, title, or DOI is simply invented. The paper does not exist. This is the famous “hallucinated reference.”
  2. The real source that doesn’t support the claim. The paper exists and is correctly cited — but it never made the point you attribute to it, or you cited the wrong page. This is subtler, more common, and far harder to spot by eye.

These need two different layers of checking — and there is a free tool for each.

Layer one: does the reference exist? — citecheck

citecheck is an open-source npm package by Tobias Lüscher. Point it at your bibliography (.bib, .ris, CSL-JSON, or extracted from .docx/.txt/.md) and it confirms each reference actually exists against Crossref and OpenAlex, flags retracted papers, and identifies open-access versions via DOAJ.

npx citecheck <file>

No API keys, no signup. It runs locally; only reference metadata is sent to public scholarly APIs. It solves the first half of the problem cleanly: fabricated, non-existent, and retracted references.

Layer two: does the source support the claim? — Acurio

The second, subtler half is what Acurio was built for. Export your thesis as a DOCX with embedded Zotero citations, upload your source PDFs (or BibTeX/RIS), and Acurio reads each cited source and judges, citation by citation, whether it actually backs your claim. You get a colour-coded report:

DOCX in, DOCX out. Multiple language models analyse each citation independently; when they disagree, Acurio runs an overnight “second opinion” with a larger model and tighter context. Swiss data handling, FADP/DSGVO-compliant.

Acurio is not a plagiarism checker and not a style editor. Plagiarism tools find text too close to a source. Acurio finds the opposite failure — claims too far from the source, where the document does not say what you wrote. (More on that distinction in AI proofreading vs. citation verification.)

Together, the two are complementary layers: citecheck guarantees the reference is real; Acurio guarantees it earns its place in your argument. citecheck also powers Acurio’s free Quick-Check.

Limits — honestly

This approach is evidence-based, not magic. Acurio checks whether a cited source supports a claim; it cannot judge an argument you made without citing anything, and a verdict is only as good as the source PDF you upload. citecheck verifies that references exist and their status — it does not read the full text or assess argument quality. And none of this measures “did a human write this.” That, deliberately, is not the question we think is worth asking. The question worth asking is whether the work stands up to its sources — and that one, unlike intent, is genuinely checkable.

Where to start

Students at ETH Zürich, LMU München, IU Internationale Hochschule, and Berner Fachhochschule already use it.

Want more background? Read how AI hallucinates sources, how to spot predatory journals, or our comparison of thesis-checking tools.

The detector debate asks an unanswerable question. The better question is older and simpler — does your source say what you claim it does? That one has an answer.

Sources

Frequently asked questions

Can AI detectors prove a student used ChatGPT?
No. As the SRF report from 22 June 2026 shows, even universities that run AI detectors stress their limited reliability — almost nobody can actually prove AI use. Detectors return a probabilistic "AI-likelihood", not evidence. That is why a verifiable, citation-based check is the more honest approach.
How is Acurio different from an AI detector?
An AI detector guesses how likely a text was machine-written. Acurio checks something objective instead — whether each citation is actually supported by its source. A claim either is or is not backed by the cited document; that is verifiable evidence, not a probability score.
What does citecheck do?
citecheck is a free open-source tool that confirms each reference in your bibliography actually exists (via Crossref and OpenAlex), flags retracted papers, and identifies open-access versions. Run it with npx citecheck <file>. It catches fabricated references; Acurio catches real sources that don't support the claim.
Does this protect honest students too?
Yes. If you accepted a citation suggested by an AI assistant that turned out to be hallucinated or wrong, a verification check finds it before your examiner does — regardless of whether you "used AI". It protects honest work, not just polices misconduct.
Is Acurio a plagiarism checker?
No. Plagiarism tools find text that is too close to a source. Acurio finds the opposite problem — claims that are too far from the source, where the cited document does not actually say what you wrote.

The 12 most common citation mistakes

Get the free PDF guide — examples from real bachelor's and master's theses.