Productive Friction: An AI Prototype for Peer Dialogue

CoTinker AI@AU started in 2026 as a two year research project at the Department of Computer Science, Aarhus University. The starting point was an observation that most educators will recognise: Students already use generative AI, often in ways that work against the learning their education is meant to build. In our interviews with educators from physics, mathematics, economics and linguistics, the same picture came up again and again: a student hands a task to a chatbot and accepts the answer without engaging with the reasoning behind it.

Banning the technology is not the answer we are after. We want to know what constructive use of GenAI in higher education could look like, how it differs across disciplines, and how tools can be designed so they strengthen rather than weaken student learning. We pursue these questions through participatory design and cooperative prototyping, building on our earlier work with the CoTinker project and the Webstrates platform.

To move from interviews to design, we analysed the interview material for moments where a tool could make a real difference, and built four interactive prototypes, each taking a clear position on constructive AI use. We then brought them into a futures workshop with our four participating educators. The futures workshop (Jungk and Müllert, 1984) is a workshop method that moves from critique to utopia to realisable ideas. In our workshop, the prototypes worked as discussion objects with a point of view, giving participants something specific to push back against and anchoring a collective discussion about constructive AI use across their disciplines.

The prototype in focus here addresses the lonely chatbot exchange directly. The scenario is two students stuck on a difficult question who turn to the AI. Instead of answering, the AI asks them to discuss the question together while it listens in. It then generates a short questionnaire based on what they actually said, and only once they have worked through it does the AI respond. Its answer builds on what the students already discovered on their own.

The idea behind it is productive friction. A conventional AI tutor is optimised to remove friction, to get the student to an answer as fast as possible. We are deliberately slowing it down, inserting a moment of peer dialogue at exactly the point where a student would otherwise outsource their thinking. It makes the students articulate their own knowledge before the AI answers, and it gives the AI something to respond to rather than a single prompt.

A faster AI gives a better answer. This one tries to leave the student with a better question. Whether it does is still an open question, and one we are still experimenting our way through.