Gymnasiefællesskabet, Ordbogen, and teachers across the country are starting to find answers to how AI can enhance teaching. It's no longer just a matter of rules, but of direction.
Here is this week’s newsletter with the news and articles we have selected about artificial intelligence in the education system.
When we select content for the newsletter, we always try to be balanced and avoid bias. It is always a balancing act, because the debate on AI in education is often presented in black-and-white terms, either for or against. But whether we like it or not, AI has already moved into young people’s lives. Director Marianne Skjold from Psykiatrifonden (the Danish Mental Health Fund) says that four out of ten in the 18 to 29 age group have used a chatbot for conversations about mental wellbeing.
AI has not only moved into those sensitive conversations. It is also taking up more and more space in the classroom, and it works differently from student to student.
Last week we talked quite a bit about the study from Aarhus Universitet (Aarhus University), which showed that the academically weakest students did better with AI, while the academically strongest did worse. Interestingly, we ourselves described a similar trend back in 2023 in the article: “Different types of pupils’ use of artificial intelligence in teaching”.
That is why it is important to equip teachers to understand how the technology affects different students in different ways.
Gymnasiefællesskabet (a consortium of upper secondary schools), which consists of 29 upper secondary schools (gymnasier), has just received nine million kroner from Villum Fonden (the Villum Foundation) to provide CPD for 2,500 teachers and to develop a shared language model called gym-GPT that can be used responsibly in teaching. Ordbogen has also launched its own solution, educhat.dk, where all data are processed in a Danish data centre.
But what, then, should students learn?
Principal Malene Grandjean from Nordfyns Gymnasium (upper secondary school) argues that exam formats must quickly be rethought, and that general education, judgement and ethics should carry more weight when the technology can formulate on students’ behalf.
Associate Professor Jacob Sherson from Aarhus Universitet, who leads the Center for Hybrid Intelligence, describes two possible futures towards 2050. In one, society is AI-first, and people act as overseers. In the other, companies compete on trust and human insight, and the quality of decisions becomes more important than the speed of production. Martin Ågerup from Skattebetalerne (the Danish Taxpayers Association) sees significant economic potential in the education sector but emphasises that reaping the gains requires entirely new ways of organising the sector.
These are two very different scenarios, and the choice between them is not made in 2050. It is made in the tasks teachers set today.
In this week’s commentary I look more closely at the finding from Aarhus Universitet. For if AI helps the academically weak and weakens the academically strong, it is not enough to teach students to write better prompts.
Happy reading.
In a new press release from Psykiatrifonden, director Marianne Skjold says that more young people are using chatbots such as ChatGPT for conversations about their mental wellbeing. A survey from the foundation shows that four out of ten in the 18 to 29 age group have used a chatbot one or more times for conversations about mental health problems. At the same time, Psykiatrifonden is launching the online hub “Chatbots as shadow therapists”, which highlights the benefits, pitfalls and ethical dilemmas of using AI as a conversational partner in connection with mental illness.
Marianne Skjold raises the question of regulation and suggests that chatbots might need to be labelled like cigarette packets, or that AI should be regulated at international level. She emphasises that the development can neither be nor should be stopped, but that digital tools used in connection with mental illness must be safe and well-founded.
Therefore, Psykiatrifonden calls for clear rules, stronger safety nets and extensive public information efforts in the form of education and digital literacy across age groups.
Psykiatrifonden
A new study from Aarhus Universitet shows that the use of AI in an exam-like situation can have very different consequences for students depending on their academic level.
The result was striking: the academically weaker students performed better, while the academically stronger performed worse. For some, GenAI provided structure, wording and a useful place to start. For others, it created overload by adding too much plausible material that had to be sifted and processed under pressure. The point is that good use of GenAI in teaching is not only about better prompts, but about designing tasks where AI supports thinking rather than disrupting it. Sometimes the most important GenAI competence is simply knowing when not to use it.
The study was conducted by, among others, Associate Professor Carsten Bergenholtz from Institut for Virksomhedsledelse (Department of Management) at Aarhus BSS and published in the journal Academy of Management Learning & Education. Around 140 students took part in the experiment, where they first had to solve a written case assignment without aids. They then solved a similar task, where some were randomly given access to a chatbot while others were not. The researchers used the first task to assess the students’ academic level and then compared their performances.
The short answer from the study is: Not all students perform better with help from GenAI.
Gymnasiefællesskabet (a consortium of upper secondary schools), which consists of 29 upper secondary schools (gymnasier), has received nine million kroner from Villum Fonden (the Villum Foundation) for a project to provide AI CPD to 2,500 upper secondary teachers and to develop a shared “gym-GPT” for students. The grant follows a report from Danmarks Evalueringsinstitut (the Danish Evaluation Institute) earlier this year, which showed that 90 per cent of upper secondary students use AI for schoolwork, and that 70 per cent do so weekly.
Mette Vedel, who chairs the project’s steering group and is principal at Nørre Gymnasium (upper secondary school), explains that the project aims to close a competence gap between teachers, students and the technology itself. She also emphasises that a shared language model that meets requirements for data security, GDPR and AI regulation can counteract the social inequality that arises when some students gain access to paid AI tools through their parents while others do not.
Ordbogen has launched the platform educhat.dk, an AI assistant developed for the Danish education system. At the same time, Ordbogen has been nominated for Dataetikprisen 2026.
The approach to educhat is that all data are processed and stored in Ordbogen’s own data centre in Odense, while access takes place via Unilogin. This avoids pupils or parents having to create private accounts.
CEO Peter Revsbech says that the ambition is to remove the legal uncertainty that many school leaders have faced when pupils have used foreign AI services without clear frameworks for data processing.
The platform is designed to adapt to each pupil’s year group and academic level. It also gives the teacher the option to upload their own materials so that answers are anchored in the current syllabus rather than in generic internet answers.
Ordbogen is also working on options for speech input and read-aloud as special educational support for pupils with dyslexia and reading difficulties.
Malene Grandjean, principal at Nordfyns Gymnasium, argues in an opinion piece on A4 Uddannelse that AI not only changes the labour market, but also fundamentally challenges the education system. She points out that technological development has previously mainly affected certain job types and routine tasks, but that AI goes further because the technology can also analyse, express itself and support decision-making. This means that schools and upper secondary schools can no longer focus only on what pupils know, but must to a greater extent work on how they use their knowledge.
She calls for a change in assessment and examination formats. If school continues primarily to test the reproduction of knowledge, it risks, according to Malene Grandjean, educating pupils for a reality that has already changed. She therefore believes it is urgent to rethink examination formats, because they have a major influence on what is prioritised in teaching. At the same time, she emphasises that AI makes general education even more important. Relationships, collaboration, ethics and judgement are areas the technology cannot replace, and which should therefore play a larger role in education.
Malene Grandjean, principal, Nordfyns Gymnasium
Jacob Sherson, associate professor at Aarhus Universitet (Aarhus University) and head of Center for Hybrid Intelligence, tells Science Report that AI is in the process of shifting role from tool to active collaborator that can plan, coordinate and carry out complex work processes. This means that society must not only discuss how AI is used, but also how work should be distributed between humans and machines.
His research field, hybrid intelligence, is about designing systems, organisations and markets where AI and humans collaborate in ways that both increase productivity and preserve human judgement.
He describes two possible future scenarios towards 2050. In one, society becomes “AI-first”, where large parts of knowledge work are automated and humans primarily function as controllers. In the other, a hybrid intelligence economy is developed, where AI is used to create and analyse content, while humans continue to formulate problems, evaluate solutions and take responsibility. In such a model, the quality of decisions becomes more important than the speed of production, and companies may end up competing on trust, responsibility and human insight rather than on efficiency alone.
Rasmus Stochflet Berg
Martin Ågerup, director of Skattebetalerne (the Danish Taxpayers’ Association), argues in Ræson that AI can change the public sector significantly, and he highlights education as one of the areas where the potential may be greatest. In education, technological support may influence how time and resources are prioritised. If AI can take over or support parts of documentation, casework and routine tasks, this could mean less time on administration and more time for pupils, teaching, academic guidance and pedagogical development in schools and educational institutions.
Martin Ågerup assesses that the education sector may possibly achieve roughly the same outcomes with significantly lower expenditure than today. As a cautious and illustrative estimate, he suggests that costs might be reduced to 20–30 percent of the current level, but he emphasises that this would require completely new ways of organising the sector. The central question therefore becomes not only which AI tools schools can use, but also how teaching, organisation, task execution and staff roles should change if the technology is to create real value.
We often say that pupils need to learn to prompt better. I’m not convinced that’s the right focus. To use an AI’s answer sensibly, you need to understand the subject matter well enough to judge whether it holds up.
But the most important AI competence, which outweighs everything else, is not knowing how to use AI. It is knowing when not to use the technology. That is a form of digital literacy that cannot be trained with more tools. It has to be trained through tasks where the pupil feels the difference between her own voice and a borrowed formulation.
That distinction is not equally easy for all pupils. The study from Aarhus Universitet shows that access to generative AI in a time-limited open case assignment does not affect everyone in the same way. Academically weaker students performed better, while academically stronger students performed worse. The researchers describe this as a levelling effect rather than a genuine democratisation. If we build teaching and assessment formats on an idea of the average pupil, we risk designing tasks that primarily help the academically weaker. We raise the floor, and that is good. But at the same time we may end up lowering the ceiling, and that is a form of cognitive debt we talk too little about. The academically strong pupil not only delivers a weaker performance this time. She risks ending up with an academic life in which her own thinking is gradually replaced by editing something she did not create herself.
I do not believe that AI should be banned for the academically able and permitted for the academically weaker. The same assignment, with the same access to AI, does not mean the same thing for all students. For some students, AI can open a path into the subject. For others, the technology can become a frictionless shortcut. And for the most academically able, it can either serve as a counterpoint or quietly erode their own voice. Therefore, we should not content ourselves with asking whether AI is permitted. We must ask which parts of the academic work the student must still own.
This is not a question of dialling the technology up or down. It is a question of seeing the student as they are before we hand them the tool.
