There's a paper in front of you, and something about it just doesn't add up. You're almost certain AI wrote it. So what do you do next? It's a question more and more lecturers are having to face: how do you prove a student used AI? The answers people tend to suggest are either shaky on legal grounds or simply impractical. This article accompanies episode 1 of our new video series on academic integrity in the age of AI. We'll walk through five steps that stand up to scrutiny, and look at what court rulings tell us about what actually counts.
Table of Contents
ToggleThree approaches that get you nowhere
- The detector percentage. It looks convincing. It feels like proof. But it isn't. We'll get to the numbers below.
- The confrontation. Raising suspicions without first securing the records or preparing your questions doesn't strengthen your case. It just makes an already awkward situation worse.
- A new assessment format. For academic writing, there still isn't a pedagogically equivalent alternative. And holding an oral exam for every submitted paper simply isn't realistic: imagine a hundred students and one examiner.
And here's the problem: none of these approaches helps with the paper sitting on your desk right now.
Why this is happening more often than you might think
In Germany and Austria, 46.5% of students say they use AI in their studies every day (CHE survey, 28,584 respondents). That, in itself, isn't a problem. Using AI to help with your writing is not the same as having AI do the writing for you. But how often does that distinction actually get noticed?
Researchers at the University of Reading decided to find out. They registered fictitious students and had a language model complete their exams from start to finish: 63 submissions to unsupervised take-home exams across five modules. The result? 94% went undetected. On average, the AI-written submissions even received higher grades than those of real students.
Five steps to take when you suspect a student used AI
So, if you need to establish whether academic misconduct has taken place, what should you actually do? These five steps are a good place to start, and the order matters. Step four is where much of the evidence comes together. Step five is what can make or break the decision if it is challenged.
- Secure the evidence before you speak. Start with the basics: save the submitted version, the submission date, any available version history and every timestamp recorded by your assessment system. Don't overwrite anything, delete anything or add comments to the original file. It may feel like routine admin, but later on it could determine whether you have any usable evidence at all.
- Don't confront the student with a number. A score from an AI detection tool might give you a reason to open a procedure. It should never be what closes the case. Whether your decision stands up to an appeal may depend on that distinction.
- Have a conversation, not a hearing. Instead of asking whether someone used AI, ask them how the work came together. Which source changed their argument? Why does chapter three now come before chapter two? How did they find the study cited on page eleven? Someone who wrote the paper should be able to walk you through it. Someone who didn't may fall back on a general summary. Make a note of the date, everyone present, the questions asked and the answers given. That record can become evidence in its own right.
- Look for independent evidence. There are four checks that are almost always worth making. First, do the cited sources actually exist? Look up at least two. Second, does the timeline make sense? Did a nearly finished paper appear just days after the student was clearly struggling with the material? Third, are there noticeable shifts in style within the text? And fourth, does the writing match what the student can explain about their own topic? One clue on its own proves very little. Several pointing in the same direction are much more meaningful.
- Keep the decision-making process independent. Check your regulations to see who is responsible for deciding an appeal. If the same person who made the original finding also decides the appeal, the decision may be open to challenge.
There's one principle running through all five steps: the burden of proof remains with the institution, not the student. That doesn't change, no matter how much evidence you gather. Which brings us to the question at the heart of the whole issue: what would you actually accept as proof?
What have the courts said how to prove a student used AI?
In January 2026, a court in New York overturned a university decision that had relied mainly on a detector score of 100%. Two other tools had rated the very same text at 0%. But the decision was overturned for a different reason, one that brings us back to step five: the same person had made the original finding and decided the appeal. The student's counter-evidence hadn't even been examined. The court therefore never had to rule on the reliability of the tools themselves.
And this isn't just a theoretical debate in Germany either. On 25 February 2026, the Administrative Court of Kassel dismissed two claims brought by students challenging findings of serious academic deception. In the first case, the student had admitted using AI. In the second, the university had drawn its conclusion from several things taken together: a gap between the written work and what the student could explain orally, linguistic irregularities, and a timeline that simply didn't add up. The student had struggled considerably with the material, then produced an almost complete draft in a remarkably short time. That's steps three and four playing out in an actual judgment.
The court describes this as prima facie evidence (Anscheinsbeweis): a combination of indications that, taken together, points towards a typical sequence of events. The burden of proof still rests with the institution, as the court explicitly states. And that conclusion only stands if there is no serious evidence pointing to a different explanation. The threshold may also be lower than many people expect: according to these judgments, even a single undeclared use of AI can constitute academic deception.
Two further developments show just how much is still unsettled. An appeal has been permitted in the Kassel cases. And in May 2026, the Administrative Court of Berlin also found deception in a bachelor's thesis, but expressly left open a broader question: can work containing AI-generated passages still be considered the student's own work?
What AI detector scores can and cannot tell you
A European research group led by Debora Weber-Wulff of HTW Berlin tested 14 AI detection tools. Depending on how accuracy was calculated, the best-performing tools reached between 76% and 81%. A Stanford study looked at seven detectors and 91 English essays written by people for whom English is a second language. On average, more than 61% of those essays were wrongly classified as machine-generated. The researchers issued a clear warning: they "strongly caution against the use of GPT detectors in evaluative or educational settings".
Both studies date back to 2023. Since then, the best tools have become noticeably better at avoiding false positives in independent tests. We look at that in more detail in our article What If AI Detectors Actually Work?. But there is another side to the story. In 2025, the US Federal Trade Commission took action against a vendor that had advertised 98% accuracy. In testing, its tool achieved around 53% on general-purpose content. We develop an AI detector ourselves, so we're hardly dismissing the technology. Quite the opposite. We simply know why a percentage should never be the only thing standing between a student and an accusation.
The strongest evidence is often in the process itself
So where do you find the evidence for step four when all you have in front of you is the finished paper? Our answer is simple: look at how the work came together. A source is discovered in week one, set aside, found again in week three and finally understood. The argument changes direction two thirds of the way through. A paragraph gets rewritten three times because something about it still doesn't feel right. Think of it like maths: the final answer doesn't tell you who did the calculation. The working does. The Orélien case is a good example of this in practice. The detector score raised the accusation; the drafts helped settle the question.
Now, there's something vendors don't always mention: process documentation isn't a magic fix either. People whose writing is documented may feel as though they're being watched, and that's a perfectly legitimate concern. Students who work differently, or who write in a second language, mustn't be put at a disadvantage. Any system needs to be designed with that in mind from the outset. Could someone get around the documentation? In principle, yes, but it would take considerable effort. And comprehensive process documentation is still far from standard practice in everyday teaching in 2026. If someone tells you otherwise, ask the same question you'd ask about an AI detector score: how do you know?
How Mentafy helps
This is exactly the problem we're working on at Mentafy. Mentafy Trail documents how a piece of writing develops by comparing snapshots of the document roughly once a minute. Our analyses show that this level of detail is sufficient, so we don't collect more data than necessary. The content of deleted text is never stored; we only record that work was done on the text. It's the combination of time resolution and data filters that makes the difference between documenting a process and monitoring a person. We explain this in more detail in Academic Integrity and Student Privacy. And, like every other indicator in step four, process documentation is evidence, not a verdict. But as things stand today, we believe it's the strongest evidence available.
The video series: what's coming next?
We'll publish a new episode on 1-2 weekly basis, first in German and then in English two days later. Each episode will also have a companion article here on the blog. Episode 2 follows by 20 November: "Are AI Detectors Accurate Enough to Accuse a Student?" After that, we'll - among other things - look at how to talk to a student you suspect, whether the essay is still worth using as an assessment, and what your examination regulations should say about AI.
- Download the five steps as a one-page handout to print
- Watch all episodes on our video page
- Subscribe to our YouTube channel
And we'd love to hear from you: tell us under the video which of the five steps you find most difficult.
Selected sources
- CHE Centre for Higher Education (2026): DatenCHECK 9/2026; press release of 22 July 2026 (in German).
- Scarfe, P., Watcham, K., Clarke, A., & Roesch, E. (2024): A real-world test of artificial intelligence infiltration of a university examinations system, PLOS ONE 19(6), e0305354.
- Weber-Wulff, D., et al. (2023): Testing of detection tools for AI-generated text, International Journal for Educational Integrity 19, 26.
- Liang, W., et al. (2023): GPT detectors are biased against non-native English writers, Patterns 4(7), 100779.
- Federal Trade Commission (2025): In the Matter of Workado, LLC, File No. 232-3092.
- Administrative Court of Kassel, judgments of 25 February 2026, 7 K 2134/24.KS and 7 K 2515/25.KS; court press release (in German).
- Administrative Court of Berlin, judgment of 20 May 2026, VG 12 K 305/24; report in beck-aktuell (in German).
- Matter of Newby v Adelphi University, 2026 NY Slip Op 26021, Supreme Court of the State of New York, Nassau County, decision of 28 January 2026.





No comment yet, add your voice below!