Learning with AI, and yet more with each other

Esther van der Stappen (Avans) and Marieke Meeuwisse (Erasmus) published an opinion piece this week on ScienceGuide with a striking title, "Learning with AI, but no longer with each other." Their concern: AI helps students find answers more quickly, but the process of searching together, doubting, and the discomfort that comes with it, is increasingly overlooked. Yet it is precisely that discomfort that constitutes a large part of learning.

I recognise the danger they describe. But in my own lessons at Utrecht University of Applied Sciences and in conversations with lecturers working with EduGPT, I often see the opposite happening. When a student uses AI as a sparring partner rather than an answer machine, that same student actually comes to a fellow student or to me more often afterwards, not less.

The same technology, opposite outcome

The clearest evidence comes from a field experiment involving 839 Turkish secondary school students (Bastani et al., 2024). One group was given unlimited access to GPT-4 during maths exercises, a second group had a version that did not provide ready-made answers but gave hints, and a third group had no AI. In the exam without AI present, the students with unlimited access performed 17 percent worse than the group that had never used AI.

The group with the hint version did not have this problem. Same model, same students, same subject matter. The only difference was that this AI withheld answers and bounced questions back. So whether AI undermines or actually strengthens learning does not depend on the technology itself, but on how you instruct it.

Scaffolding that naturally becomes redundant

Vygotsky's zone of proximal development explains why this works. A good scaffold, whether it is a teacher, a fellow student or an AI, exists to become redundant after a while. An AI that asks questions instead of filling in answers functions as such an interim step and then pushes you back into the classroom to check whether you were right.

Research into AI as a Socratic conversation partner confirms this view. A study among physics students (Education Sciences, 2026) found that a ChatGPT which asks questions instead of giving answers leads to deeper engagement and a shift from superficial to conceptual questions. A later intervention study on the same approach shows that the effect is greatest when this AI guidance is provided alongside human guidance, not as a replacement for it.

Hans Visser expressed it aptly at the opening of the ZAAM Education Academy. A pupil learning to write an essay already knows the structure; weighing up arguments is the step he or she must take themselves. The AI is then instructed to ask questions about those arguments and to point out weaknesses, without writing sentences itself. Exactly the combination of knowledge about learning and knowledge about AI that also emerges in Van Damme's scoping paper on AI in education.

People will continue to need each other

Recent qualitative research among master's students at Radboud University (Cerpentier, Van Asselt and Ketelaar, 2026) shows that social dynamics do not disappear as soon as AI comes into play. Group norms and what fellow students do remain among the strongest explanations for how someone does or does not use a tool. Students still prefer to justify their choices to each other.

Research into shared regulation among young pupils points in the same direction. Gaining insight into how a system regulates for you is something different from being allowed to co-decide on the level of difficulty and approach. Agency keeps a pupil at the table, engaged in the lesson and with each other.

Van der Stappen and Meeuwisse are right about the risk. An AI that short-circuits the process between question and answer causes the conversation between people to evaporate. But that is a design choice, not an inherent property of the technology. Build an AI that returns questions instead of answering them, and students come to each other and their teacher with more questions, not fewer. That principle is also at the core of EduGPT.