Students' Agency in GenAI-Mediated Group Assessment: An Ecological-Emergent Perspective

Synthesis: Chen and Zou (2026) ask how students exercise Agency when GenAI enters an authentic group-based assessment. Drawing on 15 focus group interviews with 52 pre-service teachers in a multicultural education course, plus an interview with the instructor, they identify three patterns rather than one enthusiasm-to-avoidance spectrum: cooperation-oriented agency (five groups intensified GenAI use to hold the group's work together and protect its performance), normative agency (seven groups deliberately restrained use to protect authenticity, fairness, originality, and diversity of perspectives), and non-enacted agency (three groups whose GenAI practice never changed from individual work). Because all three patterns were negotiated, not simply caused by the tool, the authors argue the negotiation of acceptable GenAI use should itself become an explicit, assessable learning outcome.

Why group assessment is a distinct case

Most GenAI-in-assessment research examines individual work, and the group setting changes the problem in ways that matter. Group-based assessment is valued for building collaboration, communication, and responsibility, and it is pedagogically central in professionally oriented programmes such as teacher education — but it also forces members to negotiate whose and what kind of GenAI engagement counts as acceptable. That negotiation is an Agency question, not a tool question, and it sits inside a web of prior group experiences, peer relations, assessment design, and the instructor's framing.

The study's setting is deliberately authentic: an undergraduate multicultural education course in a five-year teacher education programme, in which groups of three to five designed and delivered a presentation based on case studies drawn from real teaching scenarios, worth 30% of the final grade. The instructor used the institutional "Use only with explicit acknowledgement" policy and explicitly permitted GenAI for idea generation, language polishing, visual layout, material development, and presentation scripting. The authenticity was load-bearing: the rubric required integration and coherence under a multicultural theme, each member's reflective insight tied to their own classroom contribution, and demonstrated group collaboration. Theoretically, the authors treat agency as ecological-emergent — achieved through the interplay of personal capacity, social relations, and structural setting (Biesta and Tedder; Priestley et al.) — and add a capability-based lens (what learners are genuinely able to be and do) and an epistemic lens (what students treat as legitimate knowledge and knowing). Their definition is worth keeping: agency as the emergent capacity to navigate and negotiate socio-material structures and convert available resources into capabilities for valued individual and collective intellectual work over time.

Three patterns of agency

Cooperation-oriented agency: GenAI as coordination infrastructure

For five groups, intensified use was a deliberate conversion of GenAI's affordances into collective achievement. The dominant rationale was coherence: with limited shared knowledge of each other's sections, students fed peers' contributions into a chatbot to decode them and to align their own part ("I have to put it into AI, and then help me understand what their parts are about," RG5). GenAI here worked as an artefact-in-transaction that restructured communication conditions, making fragmented distributed knowledge mutually intelligible. One group rebuilt its workflow entirely around the tool — "discussion → externalisation to GenAI → collective review → re-discussion" (RG13) — a situated achievement under time pressure. Two further rationales appeared: performance, where students read assessment criteria and peer benchmarks and intensified use to protect their part of a shared grade, and perceived safety through norms, where a permissive collective climate ("It is more comfortable because besides me, everyone in my group is using GenAI," RG1) lowered the felt risk of misuse.

Both halves of this pattern are double-edged. On one side, the authors argue it goes beyond cognitive offloading: students retained judgement while the tool absorbed parts of the coordination work, and heavier use did not mean less intellectual engagement — complicating the common assumption that more GenAI equals less thinking. On the other, the fluency of the workflow may bypass the disagreements through which genuine cooperation is built, and the group setting can invert the accountability that collaborative assessment is designed to create — a normative micro-ecology in which collective adoption replaces commitment to shared learning with shared risk management.

Normative agency: students as gatekeepers

Seven groups reduced their GenAI use — the counterintuitive direction, since group projects are more demanding and more time-constrained than individual work, conditions that usually increase reliance. Four rationales emerged.

• Authenticity. Students judged that the task demanded situated, relational knowledge built from shared classroom experience, which GenAI could not access: "we are classmates, we have the same problem, we have the same learning process… But AI only knows that moment when you type" (RG6). GenAI's lack of longitudinal situational awareness and personal touch made it a poor fit for an assessment that spanned a term and valued coherence between members' contributions.
• Fairness and integrity. Group membership created relational obligations: "If I generate my part in AI… it's just like I'm not involved in the project… [It would be] not fair to other groupmates" (RG2). Students who accepted risk in individual work — where consequences were self-contained — refused it in group work, where risk was distributed. The authors read this as peer accountability extending from contribution norms into technology use, and note the tension with the instructor's intention: students equated GenAI use in the group with social loafing.
• Originality. Students anticipated that similarly prompted groups would converge: "if we rely too much on it… this will make us, maybe, similar to different groups" (RG7). Originality was reframed as faithfulness to understandings developed by human classmates together — a projective, future-oriented act of agency.
• Diversity of perspectives. Groups already contained diverse viewpoints, making GenAI redundant, and several feared the tool would flatten the diversity they valued.

The authors are careful to keep this pattern at some distance: restraint may also reflect unfamiliarity with the tool, assessment-driven risk avoidance, or an intuition that caution is the more defensible position — and the equation of group GenAI use with loafing is debatable. They distinguish agency as compliance with norms from agency as normative self-regulation, placing these students in the latter: boundary-drawing driven by values.

Non-enacted agency: capabilities that never reach the group

In three groups, the group setting changed nothing. Work was partitioned into discrete subtasks ("We usually divide into different parts… it's the same [as individual work]… because we work on individual platforms," RG10), and while individual students described sophisticated GenAI use, they never contributed those capabilities to the group. Because the instructor had made coherence an explicit criterion and it still did not mobilise collective practice, the authors argue that individual capability does not produce collective agency on its own — and that how students perceive and have experienced prior collaboration is a prior condition that the GenAI literature has largely skipped.

What the study adds

• The enthusiasm–avoidance axis is too coarse. Both intensified use and restraint contained negotiation: students agentically evaluated their resources and the assessment ecology before deciding what to do. Treating either as evidence of engagement or disengagement misses the reasoning inside it.
• Assessment design is a constitutive force, not a backdrop. The authenticity criterion and the coherence criterion shaped what students did with GenAI — but the coherence criterion alone did not produce coherence, and some groups read the instruction to engage critically with GenAI as evidence that GenAI was ineffective.
• Instructors' intentions and students' practices diverge in both directions. Students used GenAI uncritically where collective adoption lowered perceived risk, and avoided it where they read the task as demanding situated knowledge. The study joins a growing line of work arguing that both perspectives must be examined together in group assessment.
• The relational question is new. Whether GenAI-mediated collaboration weakens or eases the relational labour through which group cohesion is conventionally built had not previously been posed in group-work research.

Implications for practice

• Make the negotiation an assessable outcome. Units with group work should require teams to collectively justify and document how GenAI will and will not be used, rather than leaving the norm to emerge from peer pressure or risk perception.
• Build peer interaction into the process. Peer review and Feedback as smaller tasks leading to the final product create the interactions through which norms are actually negotiated — aligning group assessment with process-oriented learning.
• Do not assume capability becomes collective. If the assessment permits a divide-and-conquer structure, individually capable students may never pool what they know; the collaboration has to be designed for, not merely required in the rubric.
• Read restraint carefully. Students who avoid GenAI may be exercising normative self-Regulation — or guarding against risk and unfamiliarity. The two call for different instructor responses.

Connected Concepts

• Agency
• Group Work
• Assessment
• Authentic Assessment
• Academic Integrity
• Collaborative Learning
• Generative AI
• Higher Ed
• Teacher Education
• AI Use Disclosure
• AI Education
• Student Experience
• Learning Design

Connected Articles

• Jin Emergent Learner Agency Implicit Hai 2026 — Emergent learner agency when AI joins a group: supportive vs. contrarian personas
• Xu GenAI Collaborative Space 2026 — How team access to GenAI (shared vs. private) shapes collaboration
• GenAI Counter Learner Groupthink 2025 — GenAI countering learner groupthink in collaborative tasks
• Chang GenAI Peer Feedback Collaborative Argumentation 2026 — GenAI-supported peer feedback in collaborative argumentation
• Hao Human AI Collaborative Problem Solving Cognition — Collaboration modes and the task-performance vs. self-regulation trade-off
• Polished Artifacts Fragile Engagement 2026 — Polished artifacts, fragile epistemic engagement
• Preservice Teacher Agency GenAI Design Learning 2026 — Pre-service teacher agency during GenAI interactions in design for learning
• GenAI Chinese Higher Education Integrity 2026 — GenAI in Chinese higher education: integrity and engagement
• Dollinger Equitable Assessment AI 2026 — Equitable assessment in the AI era
• Beyond Detection Authentic Assessment AI 2025 — Beyond detection: designing authentic assessment
• Chirikov Regulate AI Syllabi 2026 — Task-level regulation of AI in course syllabi
• AI Assessment Scale Reform — Rethinking assessment in the generative AI era

Citation

Chen, J. X. J., & Zou, T. X. P. (2026). Students' agency in GenAI-mediated group assessment: An ecological-emergent perspective. Assessment & Evaluation in Higher Education, advance online publication.