A student brings an AI-generated summary to a group meeting. It is clear, quick and apparently useful. Accepting it, however, does more than save time. The group must decide whether the summary represents their shared understanding, whether anyone has checked it, who carries the risk if it is wrong, and whether using it changes the work they owe one another.
That is the difficulty at the centre of Jodie X. J. Chen and Tracy X. P. Zou’s study of student agency in GenAI-mediated group assessment. The authors ask how students exercise agency when AI use is neither simply prohibited nor privately chosen. Their answer is that agency emerges through the group’s relations, assessment conditions and judgements about what the task requires.
The paper’s distinct contribution is to move beyond a scale running from enthusiastic use to cautious avoidance. In the study, using more AI, using less and making no collective change could each arise from a different organisation of group work. The amount of use does not, on its own, tell us whether students have exercised judgement together.
One permission, three patterns
The study took place in a seven-week multicultural education course within a five-year undergraduate teacher-education programme in Hong Kong. A group presentation, worth 30% of the course grade, asked teams of three to five students to use cases adapted from teaching practice, explain their multicultural design choices and reflect on what had worked.
The instructor permitted GenAI for idea generation, language polishing, visual design, material development and presentation scripts, provided its use was acknowledged. The assessment rubric valued coherence, presentation, individual reflection and collaboration. This was therefore a consequential task with an explicit permission to use AI, rather than a hidden-use scenario.
Chen and Zou conducted 15 focus-group interviews with 52 students, keeping students with their assessment groups, and a separate one-hour interview with the instructor. They used reflexive thematic analysis, informed by ecological-emergent, capability-based and epistemic accounts of agency. The researchers treated their own interpretive positions as part of the analysis and used disagreements between them to refine the themes.
Five groups reported increasing their GenAI use. The authors call this cooperation-oriented agency. Students used AI to connect separately prepared contributions, make unfamiliar material intelligible and improve performance. In one workflow, ideas moved from discussion to AI-supported synthesis, then back to collective review and further discussion.
Seven groups reported reducing their use. Here the authors identify normative agency. Students thought heavy reliance could weaken authenticity, fairness, originality or diversity of perspective. An output based on general information could not know the group’s shared classroom experience. A member who generated their contribution with little effort might also be seen as withholding intellectual labour from peers.
The remaining three groups did not substantially change their individual practices. They divided the task into separate parts and worked on their own platforms. Some students described capable individual AI use, but those capabilities did not become a collective way of working. Chen and Zou describe this as non-enacted agency at group level.
Coordination is not yet collaboration
The first pattern is easy to misread. More intensive use did not necessarily mean that students had surrendered judgement. AI sometimes helped them see how separate contributions fitted together. Yet a coherent presentation can also conceal a thin collaborative process. If a system smooths disagreement before students examine it, coordination may improve while the relational work through which a group learns is reduced.
The second pattern also resists a simple interpretation. Deliberate restraint may express a reasoned commitment to situated knowledge and shared responsibility. It may also reflect unfamiliarity, fear of being accused of misconduct or a sense that caution is the safest answer to give in an interview. The authors acknowledge these alternatives. Less use should not automatically be treated as stronger agency.
The third pattern sharpens the point. A group can contain individually skilled AI users without developing a shared capacity to judge AI-supported work. The obstacle may lie in prior experiences of group work, an established divide-and-conquer routine or the belief that discussion is too costly. A tool does not create collaboration merely by being available to every member.
This makes the study relevant to work on epistemic agency: students are not only selecting a tool but deciding what will count as knowledge, whose experience matters and how a claim becomes acceptable within the task. It also supports the case for assessment that develops evaluative judgement. In a group, that judgement has to become discussable among people who may have different standards, confidence and exposure to AI.
Permission does not travel intact
The instructor’s permission was the same across the course, but it acquired different meanings in different groups. Some students read it as an opportunity to improve coordination. Others concluded that the task’s emphasis on shared experience made AI less useful. Elsewhere, normalised use reduced the perceived risk of relying on it. The formal rule did not determine the practice.
From an Autopoietic Ecology perspective, this is a problem of translation and continuation. A policy statement enters a group already shaped by grades, prior relationships, expectations of fairness, available time and experience of earlier collaboration. Students convert it into a working arrangement. That arrangement then affects what can be questioned, whose contribution remains visible and what the group is able to do next.
This is an EdgeLab interpretation, not an additional finding reported by Chen and Zou. The study shows how students described their decisions at one point in one course. It does not follow whether a group’s arrangement changed later assessment practices, professional judgement or teaching. A longitudinal study would be needed to establish that continuation.
What the interviews cannot establish
The evidence is qualitative and situated. It comes from one teacher-education course, 52 volunteers and one instructor. The interviews record participants’ accounts after or around the assessment; the study does not independently observe every group meeting, inspect complete AI interaction logs or compare the quality of work produced with and without AI.
Interviewing students with their assessment group has a clear advantage: members can recall and contest a shared process. It also creates a possible constraint. A student may be reluctant to describe free-riding, disagreement or unacknowledged use in front of peers. The three patterns are interpretive themes across 15 groups, not prevalence estimates for higher education.
The categories may also overlap in practice. A group might use AI more for editing while refusing it for generating ideas. Restraint in one phase and intensified use in another could belong to the same group. The analysis is valuable because it makes these judgements visible, but it does not establish a stable type of student or group.
Make the decision part of the learning
Chen and Zou recommend making negotiation over acceptable AI use an explicit, assessable learning outcome. Groups could be asked to document what they will and will not delegate, and to justify those boundaries. Smaller peer-review and feedback tasks could make the process visible before the final product is assembled.
That proposal follows plausibly from the analysis, but it has not been tested by this study. Documentation can support reflection; it can also become another compliance form. Peer checkpoints can create occasions for disagreement; they can also add workload without changing who is able to speak. The practical test is whether the design improves the group’s capacity to examine a contribution and revise its arrangement when the contribution fails.
The paper therefore changes the question worth asking. “How much AI did the students use?” is too coarse. We need to ask what the group was trying to preserve, who could challenge an AI-supported contribution, how risks and credit were distributed, and whether the arrangement left the group better able to judge the next contribution. Agency becomes visible in those relations—and in the costs of sustaining them.
Paper reviewed: Chen, J. X. J. and Zou, T. X. P. (2026). “Students’ agency in GenAI-mediated group assessment: an ecological-emergent perspective”. Assessment & Evaluation in Higher Education, advance online publication, 9 September 2026. DOI: 10.1080/02602938.2026.2730644. The complete open-access article was assessed. The DOI record and author-uploaded full text were checked on 26 October 2026; no correction or retraction was identified in the records retrieved.
EdgeLab is an independent research studio and network. This critical review distinguishes the study’s evidence from EdgeLab’s interpretation and does not imply endorsement by the paper’s authors, their institution or the University of Cambridge. AI-assisted tools supported source discovery, comparison, drafting and editorial checks; source selection, verification and editorial responsibility remain with EdgeLab.
Leave a Reply