AI in education requires more than just quick answers. In this article, I examine Peter Dalsgaard’s concepts of premature coherence and reflective friction—and explore how they can be used to preserve students’ critical thinking in the classroom.
This article is based on a new research paper by Peter Dalsgaard, professor and head of the Centre for Digital Creativity at Aarhus Universitet (Aarhus University). The paper can be read here: Reflective Friction in Generative AI: Designing for Slow Thinking in Fast Systems. I briefly outline the ideas behind Peter Dalsgaard’s paper and turn them into concrete ways to work with AI in teaching. The Danish translations of the terms are my own.
Right now, we are seeing teachers and students experimenting with and using tools such as Claude, ChatGPT and Copilot in teaching. These tools are becoming standard equipment for working with generative artificial intelligence. Language models are good at delivering quick, well‑phrased answers, and the newest model, GPT‑5 from OpenAI, makes it even faster.
The answers appear so polished and convincing that they are often taken at face value – even when we know the models can hallucinate. Students’ critical faculties risk being weakened.
Peter Dalsgaard calls this “premature coherence”. AI delivers results that seem finished and correct before students have even explored the question or gone through a learning process. The fast, fluent interaction removes the natural pauses in which we would normally consider, compare and adjust our thinking.
To counter this, he proposes building reflective friction into teaching – small, deliberate pauses or prompts that give the user reason to consider the answer before accepting it.
In teaching, reflective friction is a method to ensure that AI does not take over all the thinking for students. In the paper, Peter Dalsgaard emphasises that this is not about banning the technology or making it cumbersome to use, but about creating space for immersion and critical thinking along the way.
Peter Dalsgaard has identified eight concrete ways of working with reflective friction, which I have translated below into classroom practice.
By building these strategies into the tasks themselves, students are automatically prompted to reflect on how they use AI.
Fundamentally, it is not only about better and newer AI tools, but about preserving students’ capacity for critical thinking. By designing teaching that creates opportunities to reflect on how students use generative artificial intelligence, rather than simply delivering quick answers, we help students develop better judgement – not just faster output.
The goal is to shift the focus from getting faster answers to being able to ask better questions. Peter Dalsgaard describes this as moving from optimising AI’s output to optimising students’ own learning and thinking.
As a teacher, you can:
For students, this means:
AI does not need to replace learning – it should support better learning. Reflective friction gives us concrete tools to do just that in the classroom.
Peter Dalsgaard’s paper reveals a fundamental problem with the AI tools we use in teaching. They are not designed for learning and are released by tech giants that are more focused on delivering a usable answer. SkoleGPT and other learning tools could usefully incorporate the principles of reflective friction.
Here, Dalsgaard’s points align well with something Claus Scheuer-Larsen has written previously, namely the danger that students begin to see AI’s output as “the right answer”. When that happens, they lose both the sense that knowledge can be complex and the understanding that learning is about exploring and discussing, not merely finding a single correct result.
Claus Scheuer-Larsen
AI should not function as an answer key, but as a starting point for inquiries and discussions. Reflective friction can therefore help students see AI answers as a springboard for learning—something to be examined, challenged and developed further, not merely copied directly into the assignment.
