Problem-based learning students saw AI as useful while negotiating effort, group trust, and rules

In a collaborative project, saving time and learning through the work do not always point in the same direction. Sofie Otto, Stine Ejsing-Duun, and Euan Lindsay explored how students negotiated that tension when generative AI entered a university organized around problem-based learning. The study found broadly favorable attitudes alongside concerns about effort, trust, and uncertainty. Its contribution is to place AI use within the social organization of learning, rather than viewing it only as a private exchange between a student and a tool.

The research took place at Aalborg University, where students work in groups on extended, authentic problems with supervision. An online questionnaire was distributed to first- and fifth-semester students in IT-focused programmes during the winter of 2023. After incomplete responses were excluded, the analysis included 83 students. Alongside questions about experience and attitudes, the survey asked participants to describe their most important benefit and concern. Sixty-seven respondents contributed answers to the open questions, producing 134 qualitative responses.

The researchers combined descriptive statistics and principal component analysis with a thematic analysis informed by activity theory. That perspective connects a learner’s actions with project goals, peers, expectations, tools, and divisions of work. It helped distinguish using AI to support thinking and understanding from using it to accelerate or replace task performance. Those uses can complement one another, but they can also conflict when a task is valuable precisely because students must struggle through it together.

Students described benefits including explanations, idea generation, coding support, and help getting started. Most regarded AI as a potentially useful learning aid, and many trusted their group members to use it responsibly. At the same time, some worried about dependence and the loss of learning opportunities. The qualitative accounts showed that disagreements could arise when one member treated generated output as efficient progress while another saw the same action as avoiding the reasoning the group needed to practice.

The study does not suggest that project groups were generally in conflict. Most respondents disagreed that AI had caused conflict in their own group. Rather, a subset of accounts revealed consequential tensions that a simple average satisfaction score could obscure. Students also called for clearer guidance about acceptable use and expressed concern about unintended misconduct. The authors interpret these issues as interactions among educational purposes, community expectations, and rules, rather than as a problem solved solely by choosing a more capable model.

The limitations are substantial and clearly stated. The evidence is self-reported, comes from a modest sample in one institutional setting, and captures a particular early period of adoption. Asking for the most important benefit and concern may emphasize extreme experiences. Individual responses cannot fully represent group-level dynamics, and the study did not directly observe collaboration or measure independent learning gains. It therefore maps experiences and possible design problems without demonstrating that AI caused either better learning or a decline in trust.

For AIEDHK, a practical implication is to make group expectations explicit before substantial project work begins. Students could agree which activities they want to perform themselves, where AI support is acceptable, and how they will explain generated contributions to one another. Those agreements should be revisited as the project develops. This is a suggested response to the tensions reported in the paper, not an intervention tested by it, and it treats students as participants in designing responsible collaborative practice.