Rules for university exams: Artificial intelligence takes notes anyway

A bachelor's thesis in computer science, a term paper in administrative law - both were written with the help of artificial intelligence. In both cases it was exposed. It was over for the students responsible: the Kassel Administrative Court had not only rated the two papers as “failed”. It also excluded the students from taking any exams again. The harsh verdict was intended to send a message: Anyone who has their exams done by AI must expect drastic consequences.

The administrative court heard the case in February. Language models have long since found their place on smartphones as conversation partners, guides and encyclopedias - especially in the lives of school children and students. ChatGPT, Gemini, Claude and Llama answer questions, help with research, help solve tasks - or do the work completely on their own. Depending on the question and perspective, teachers and lecturers are divided on this development: Should the use of AI be banned as fraud or should it be promoted as a skill?

Even in everyday school life, it is difficult to see whether a student has completed a homework assignment using their own thoughts or generated blocks of text. At universities this is even more difficult given the much larger number and length of scientific papers.

A large-scale study by the University of California shows that AI-supported language models have now found their way into all levels of science: When analyzing 2.5 million publications across multiple platforms, researchers found around 150,000 hallucinated quotes for the year 2025 alone. A systematic distortion is also noteworthy. The AI ​​disproportionately attributed the non-existent sources to male scientists, inadvertently reinforcing existing inequalities in citation practices.

The problem is getting worse and worse. The better the language models work, the harder they are to debunk. For a short time, universities hoped to detect AI-generated content using AI detectors. But they are now considered unreliable. Most recently, Claude's developers Anthropic promised to watermark their AI output and make it recognizable. It only took a few days for the first technical solutions to remove the invisible signs to appear. So when technical control is no longer effective, the task shifts to the design of the exam itself.

However, there are still few concrete rules for dealing with generative AI in exams. Only gradually are AI guidelines at universities developing from catalogs of prohibitions into enabling documents that critically reflect on the use of AI from an academic perspective and thus aim to integrate it into everyday student life.

“The dynamics are significant,” says Anja Westermann from the University Forum for Digitalization in the Center for University Development, who has examined the guidelines of universities. “At the beginning of the hype about language models, there was hardly any questioning of how they were handled, both by students and lecturers,” she says. “There is now a lot of talk about what exams should look like in the future and which skills are relevant.”

In order to adapt examination designs, it is being discussed to supplement written work with an oral defense. Additional reflections or work process reports also play a larger role. Another possibility would be to tailor the exam topics so specifically that they cannot be processed generically.

There is no uniform regulation for future examinations - nor can there be. Each federal state can make recommendations, each university decides for itself. And no university wants to restrict its faculties and institutes with requirements that may make sense for one subject but not for another. In one seminar you may be able to develop projects and invent models where a language model is of no help at all. Another course could specifically train the use of AI tools or explicitly aim to promote skills without AI help.

In Bavaria, people adapted to the changed framework conditions relatively early. The Technical University of Munich in particular developed its own AI strategy. The University of Kaiserslautern-Landau has also introduced a binding “AI compass”. “It is the responsibility of teachers to make transparent how and to what extent the use of AI systems is permitted in their courses,” it says. This information should be made known to the students at the latest when the exercises are given out, along with clear information about the permitted aids. In other federal states, universities are more hesitant about the regulations.

Overall, every guideline from German universities has its own strengths and weaknesses, says Anja Westermann. What is still missing in most of them are the ethical dimensions. For example, AI guidelines still do not adequately draw attention to the potential of AI to systematically reduce barriers in studying and teaching. “The bias of AI is sometimes cited as a risk, but the social dimension, such as the reproduction of power and inequality, remains largely underexplored.” In addition, the question of ecological resource consumption is rarely addressed. Where it occurs, it is primarily left to the users.

In general, the analysis by scientists at the University of California also shows that almost all guidelines are aimed primarily at teachers. The systematic involvement of students is often missing. “These are adults who have voluntarily decided to study,” explains Anja Westermann. “You should talk to this important group of actors instead of just formulating rules for them.” After all, students need to know to what extent they can use the language models sensibly if their work is still to be considered independently produced. However, in most guidelines the term autonomy remains vague. After all, some departments are trying to involve students, but with mixed results: The current impression remains that there is still no real participation, says Westermann.

An example from the USA shows what participation can look like. In the summer, almost 100 young people developed their own proposal for using AI in schools in a replica of the US Senate chamber. Your “Students First Act” contains rules on the extent to which the use of AI is permitted, as well as suggestions on the burden of proof and the defense in suspected cases.

At German universities, the reality of students' lives is still missing the debates, says Anja Westermann. “If they want to assert themselves on the job market later, they will have to have AI skills,” both general and, increasingly, subject-specific. This development can no longer be ignored - but it can be shaped.