Generative AI is not merely a tool that can be regulated from the outside. It is a technology that changes the very conditions for learning. Therefore, the education system must experiment. If students are to be able to navigate a world in which knowledge can be generated automatically, it is essential that they develop the ability to evaluate, make choices, and take responsibility.
Across schools and universities, students are increasingly unsure whether using artificial intelligence counts as learning or cheating. Teachers, meanwhile, face a different problem: if assignments can now be produced with the assistance of generative AI, what exactly is being assessed?
These questions are no longer theoretical. In Denmark, a recent report by the Danish Evaluation Institute (EVA) found that many upper secondary school students themselves feel they are “cheating” when using AI. The report has triggered political debate, calls for new national guidelines, and growing concern among teachers about trust, authenticity and assessment.
One teacher interviewed in the report described how trust between students and teachers is being challenged. Another questioned whether written assignments can still reveal what a student has actually learned. If one no longer knows how a text was produced, it becomes difficult to support the student’s learning process.
The debate therefore quickly revolves around cheating. But this is too narrow a framework. The primary problem is not that students are cheating. The problem is that the traditional way in which we have coupled assignment production, learning, and assessment is becoming destabilised.
To understand why, we need to take a step back and ask: what is technology, really?
The problem is that the traditional way in which we have coupled assignment production, learning, and assessment is becoming destabilised.
In a classical sociological understanding, technology is characterised by stability. Technology enables causes and effects to be tightly coupled, so that specific inputs lead to predictable outputs – such as looking something up in an encyclopaedia or starting an engine. It reduces uncertainty about what will happen. When a technology works, we can repeat a process without constantly having to reflect upon it.
The problem is that education has never been a technology in this strict sense. Schools have, of course, always relied heavily on technologies – from classroom design and textbooks to dictionaries, timetables, and teaching methods – in order to stabilise and support learning. But education itself suffers from what has been called a “technological deficit”. Teaching is communication, but learning takes place within the consciousness of the individual student. There is no stable relationship between what the teacher does and what the student learns. Unlike a machine, the same pedagogical input does not reliably produce the same outcome. Historically, schools have therefore had to manage this deficit through didactics, evaluation, and the classical idea of Bildung – education as intellectual and ethical formation.
Generative AI, however, changes the situation. It resembles a technology, but it does not behave like one.
When a student writes an assignment with the assistance of a language model, the result is not a stable output produced from a clearly defined input. Instead, the text is generated probabilistically, as a statistically likely response based on vast amounts of data. In this sense, generative AI exceeds the classical understanding of technology. It does not simply create more predictable outcomes, as technologies normally do, it relocates uncertainty. Where uncertainty previously lay in the writing process itself, it now lies in the evaluation: Is the text accurate? Are the concepts precise? Are the references genuine?
This also means that the written assignment changes character. It still expresses the student’s academic abilities, but not in quite the same way as before. Increasingly, it demonstrates the student’s ability to steer, evaluate, and revise AI-generated text within a media environment where the act of formulation itself can to a greater extent be automated.
This shift is also reflected in the growing international research on generative AI in education. Meta-analyses and review studies point to a double pattern: AI can improve performance, motivation, and confidence, while simultaneously increasing the risk of superficial learning, dependency, and a lack of transparency in the processes producing the answers.
Generative AI is neither unambiguously a benefit nor a threat. It changes the conditions for what it means to learn and to be assessed. This becomes evident in an analysis of 88 exam papers from a university course where the use of AI was permitted but had to be declared. Here, an interesting pattern emerges. On the one hand, there is no dramatic change in the grade distribution. The strongest students still perform best, but are now capable of more than before. On the other hand, there has been a shift in where the problems arise.
The weakest assignments are no longer necessarily linguistically deficient. On the contrary, they may appear surprisingly well-written. The problem instead is that they lack conceptual grounding, precise references, and coherence between theory and analysis. The texts can seem generic, as though they could fit almost any case.
Generative AI is neither unambiguously a benefit nor a threat. It changes the conditions for what it means to learn and to be assessed.
The best assignments, by contrast, distinguish themselves through something different than before. Not primarily through linguistic form alone, but through the ability to steer the AI: selecting relevant concepts, maintaining theoretical coherence, verifying references, and sustaining an analytical line of argument. This points to a crucial shift: the text is no longer primarily a sign of writing ability, but of academic judgement and the capacity to assess whether something is academically sound.
The teacher’s task therefore also changes. Where quality could previously be assessed largely through the form and coherence of the text, it now requires a much more in-depth evaluation of conceptual usage, references, and analytical necessity. Assessment shifts from form toward whether the content is actually accurate, well-founded, and academically sound.
Seen in this light, it is not surprising that students, teachers, and politicians alike experience the situation as uncertain. The norms and practices that have so far regulated teaching and assessment were developed for a media environment that is now changing. This leads to the normative question: what should we do?
The obvious answer is not to attempt to restore an earlier condition through control and prohibition. That is unlikely to be possible, and it would not address the underlying problem. Generative AI is not merely a tool that can be externally regulated. It changes the very conditions of learning itself.
Instead, the educational system must adopt a more experimental approach. We are in a situation where there are far more possible outcomes, and where it has become more difficult to predict what works and what is correct. This calls for a form of action research in which educators systematically test new forms of teaching, assignments, and evaluation.
Here, Bildung becomes a central concept – not as a nostalgic reference, but as a necessary orientation. If pupils and students are to navigate a world in which knowledge can be generated automatically, it becomes crucial that they develop the ability to assess, select, and take responsibility (sapere aude). This raises three fundamental questions: What may one know? What does this mean for the community? And who does one become in the process?
First, knowledge concerns both deep disciplinary insight and the ability to steer generative AI. One works back and forth with it – asking questions, finding texts, and verifying answers – so that AI becomes a tool, while responsibility for the correctness of the content remains one’s own. This requires both an understanding of AI and disciplinary expertise, including a critical awareness that AI systems are shaped by the interests and data on which they are built.
Second, students must learn to take the broader community into account and consider the consequences for others. Without a critical relation to what one shares, there is a risk of spreading misinformation. And if AI is used for solutions that benefit a few while harming the community or the environment, one fails to meet the responsibility that follows from its use.
(…) Bildung becomes a central concept – not as a nostalgic reference, but as a necessary orientation.
Third, on an existential level, the challenge is to find one’s own voice within the community. Not by outsourcing thinking to AI, but by taking responsibility for understanding, evaluating, and standing behind what one says and writes. This requires resisting the temptation of the quick solution and making the effort to learn something oneself. At the same time, it involves moving beyond narrow self-interest toward a responsibility for the community as well.
If we reduce the issue to cheating, we overlook these questions. But if we take them seriously, it becomes clear that we are in the midst of a more fundamental transformation. Generative AI does not merely challenge our rules. It challenges our understanding of learning, knowledge, and Bildung. The task, therefore, is not to return to what we once had. The task is to develop new ways of coupling teaching, learning, and assessment that are adapted to the conditions we actually face.
This is a difficult task. But it is also an opportunity to rethink what education should be in a time when systems such as generative AI no longer merely support our actions, but increasingly participate in shaping them.
This is an English translation of an op-ed originally published in the Danish newspaper Jyllands-Posten, May 1. 2026: https://jyllands-posten.dk/debat/kronik/ECE19188746/ai-er-ikke-snyd-men-det-aendrer-al-laering/ The op-ed is based on a paper that will be presented at the annual Media Ecology conference in Canada:
Tække, J. (2026). Generative AI in Education: The Reorganization of Contingency. Paper to “Crossing [out] Borders in Our Global Village” – The 27th Annual Convention of the Media Ecology Association. June 25–28, 2026. The University of Winnipeg, Manitoba, Canada. https://pure.au.dk/ws/portalfiles/portal/463752353/MEA_Paper_2026.pdf