OPINION: AU'S GAI RECOMMENDATIONS ARE A YEAR OLD – BUT HAVE ALREADY BEEN OVERRIDDEN BY TECHNOLOGY

This is an opinion piece, the views expressed in the column are the writer’s own.

GAI now functions so well as a tutor in a given curriculum that the alignment between teaching and examination (connection between academic objectives, teaching method and examination method, ed.) may be broken. Let me elaborate.

I gave a presentation on GAI in teaching for the Department of Mathematical Sciences at the University of Copenhagen on May 7, 2025, in inspiring surroundings at the Old Carlsberg Glyptotek.

Since the end of 2022, I have integrated GAI into my teaching and even allowed students to use GAI in the written on-site exam in the course 'Introduction to Mathematics and Optimisation' in the first semester of Data Science and Computer Science.

In the subsequent discussion at the Glyptoteket, I therefore defended the sacred principle of alignment between teaching and examination, when I was met with constructive and healthy scepticism. It made me reconsider on the train ride home, especially because I hadn't had time to test the new Gemini 2.5 Pro thinking model from Google, which had been released the day before.

Unlike the common models that students and employees have access to via Copilot, thinking models reflect, and you often have to wait several minutes for an answer. Thinking models were launched around September 2024 and have since become freely available, including on Google AI Studio.

GAI DELIVERS A PERFECT RESPONSE

With a simple prompt, uploading my notes and exam assignments to Google AI Studio, I saw for the first time generative AI produce a correct answer with completely correct references to the syllabus. For example, it used to be a notorious problem to appropriately integrate PDF files into a prompt for a language model (Google doesn't say how they do it).

Or to put it another way: GAI navigated the given syllabus without any problems and produced a perfect answer. A very sceptical colleague in mathematics reproduced the same phenomenon in his course a few days later. Another colleague learned that Google AI Studio could act as a tutor based on an online version of his textbook and solve assignments with flawless references to the book.

At a subsequent presentation (held on May 27, 2025) about GAI in computer science teaching at Aarhus University, the comment "What did you expect?" was made. Yes, I expected something like that, but not so quickly. Big tech is investing billions of dollars in GAI with a focus on education. The aforementioned model, Gemini 2.5 Pro, even has learning built in (LearnLM), and as of August 2025, ChatGPT and Gemini have tutor settings that guide students through a given material. Universities are extremely challenged in that area.

URGENT HALT TO THE USE OF GAI IN MY EXAMS 

I have urgently had to cancel the use of GAI for my exams and have therefore also had to give in to the mantra of alignment between teaching and exams.

I took part in developing the guidelines for the use of GAI at AU and also the preparing the forms that students must use to declare their use of GAI in written assignments at NAT. This was among the most constructive and rewarding committee work I have participated in at AU.

In the meantime, almost two years have passed, and the development has been tremendous. For example, I’m no longer sure that written homework assignments are a credible test, and that GAI should generally be allowed for exams unless otherwise stated. GAI as a tutor can be a unique tool in teaching and often provides feedback on par with the best teachers, but the imbalance between GAI in teaching and for exams is a real problem that should be discussed in depth going forward.

This text is machine translated and post-edited by Cecillia Jensen