How Should I Handle AI in Group and Collaborative Assignments?

The group presentation comes back polished and coherent, and not one member can explain the argument in it. Or the marks come out even, and you learn afterward that one student ran the whole thing through generative AI while three others coasted on the shared grade. Nothing in the file tells you which of those happened, and your rubric has a collaboration line that everybody met.

The answer to give a colleague: decide AI's place per task rather than per course, make the group's negotiation about AI something they submit and you grade, sequence the work so it cannot be split into parallel chatbot sessions, and keep at least one component each member answers alone. All of it is task design, and none of it has been tested against a no-AI control, so you are making a design argument rather than applying a proven fix.

The short version

1. Write the permission at the level of the task phase. Say which stages are AI-open, which are not, and why.

2. Require a short, jointly authored AI-use statement that records what the group decided not to do, and grade the reasoning.

3. Protect the stages where the group builds shared understanding: the framing, the disagreement, the reconciliation.

4. Choose the access configuration on purpose. One shared interface produces negotiated, visible use; private prompting produces flows you cannot see.

5. Keep one component each member answers for alone, and make it comprehension rather than authorship.

6. Read a group's restraint before you correct it. Declining AI can be deliberate self-regulation or plain unfamiliarity.

Write the permission per phase, not per course

Chen and Zou (2026) studied groups of three to five delivering a presentation worth 30% of the grade, under an institutional policy of "use only with explicit acknowledgment." GenAI was permitted for idea generation, language polishing, visual layout and scripting. Inside that single permissive policy, groups reached opposite conclusions with defensible reasoning on each side, and the design of the task shaped which conclusion a group reached. The study is 52 pre-service teachers in one course, and the authors ask for their patterns to be read as possible responses to design rather than as a distribution you should expect. A course-level rule does not determine what a group does.

So write the permission at the level of the artifact. Name the phases that are AI-open and the phases that are meant to be worked out without it, and tell students why the line sits there. The scoping review's design logic points the same way: if AI-mediated communication compresses negotiation and collective sensemaking, then the segments where a group builds shared understanding are the ones worth protecting (Wei and Perkins, 2026).

Be honest about the warrant in the staff meeting. Neither review claims that withholding AI improves learning outcomes. That case rests on the structure of the trade-off below, not on a measured comparison.

Make the group submit its AI negotiation

Chen and Zou's central design argument is that the negotiation of acceptable GenAI use should itself become an explicit, assessable outcome. Teams should collectively justify and document how GenAI will and will not be used, rather than leaving the norm to peer pressure or perceived risk. That artifact exists only if the group actually talked, which is what makes it worth points: it converts an unenforceable rule into reviewable student judgment.

Two reasons explain why groups intensify use, and both respond sensibly to a shared grade. Performance: students read the criteria and peer benchmarks and used AI to protect their part of the mark. Perceived safety through norms: a permissive collective climate lowered the felt risk of misuse, so groups used the tool more openly and, in the cooperation-oriented case, more dependently. If you want transparency, a stated classroom norm is more effective than a warning.

Ask for the negative as well as the positive. A statement that records what the group decided not to do, and why, tells you more than a list of tools. It also gives a group that disagrees somewhere to put the disagreement, which is usually the hard part already done.

Do not assume the document reflects a shared understanding. Korchak, Costley and Fanguy (2026) interviewed 10 postgraduate students aged 23 to 39 (mean 27.7) in a scientific writing course, where groups used GenAI through pre-planned strategies or through open, individually driven interactions coordinated in shared documents. One distinctive use was GenAI as an editorial integrator, merging parallel sections into coherent text. Their most pointed observation: students' accounts sometimes contradicted their own written reflections, and one student denied having a group strategy while describing one in writing. A strategy that is not equally visible to every member is not reliably visible to you either.

Sequence the work so it cannot be divided

Wei and Perkins (2026) ran a PRISMA-guided scoping review of 18 English-language studies published between January 2023 and March 2025, analyzed with reflexive thematic analysis and mapped across eight themes. The benefits they report are group knowledge development, idea generation, support for reflective thinking, communication efficiency, task coordination and feedback. The risks are reduced peer interaction and engagement under over-reliance, plus privacy, transparency and accuracy concerns. Their assessment is blunt: the benefits are empirically supported, while the risks around privacy, transparency, bias and accuracy are largely discussed conceptually, and much of the research is short-term or conceptual with little longitudinal work on group dynamics or cognitive development.

The trade-off governs the design. Group work became more efficient, with reduced demand for communication, negotiation and collective sensemaking, a finding the review takes from Lin et al. Efficiency gain and interaction loss can be the same mechanism, which means you cannot take the speed without the loss unless you build the negotiation back in as required work. The review concludes that group-based assessment should shift its focus from product to process.

Chen and Zou show how hard that is to arrange by accident. Three patterns of Learner Agency ran in different directions at once. Cooperation-oriented agency (five groups) intensified use to hold the work together, often feeding peers' contributions into a chatbot to decode them and align their own part; one group rebuilt its workflow as discussion, externalisation to GenAI, collective review, and re-discussion. Normative agency (seven groups) deliberately restrained use, either judging the task to demand situated knowledge AI could not reach ("AI only knows that moment when you type") or treating AI-generated sections as unfair to the groupmates who would carry them. Non-enacted agency (three groups) changed nothing, partitioning work into discrete subtasks on individual platforms while individual students used GenAI in ways they never contributed to the team.

Their rubric already required integration and coherence under a multicultural theme, each member's reflective insight tied to their own classroom contribution, and demonstrated group collaboration. Three groups still divided and conquered. A coherence criterion does not produce collaboration on its own, so change the sequence instead of the wording: put at least one stage where the whole team has to be in the same conversation, and make the output of that stage a deliverable.

Put one interface at the center, and choose it on purpose

Xu et al. (2026) observed 18 students in six groups of three working on a single shared ChatGPT-4 interface for roughly 45 minutes, and interviewed nine further students about asynchronous teamwork. With a shared, visible interface, teams co-constructed collective prompts, ran a recurring surface, evaluate and embed cycle, treated the chat as shared external memory, and openly negotiated AI's role. In asynchronous work, private prompting and output de-labeling produced opaque, privatized information flows that raised the cost of maintaining a shared cognitive model. GenAI's role shifted too, from subordinate assistant to contested teammate, and the system is cognitively involved while remaining contextually unaware of the team's shifting focus.

The practical difference is whether the prompt is a group artifact. A shared document with the prompts and outputs in it gives you something to read and gives the group something to argue with. Private prompting gives each member a different assistant and no common record.

Keep a part each member answers alone

Pair the group mark with an individual component: an oral defense of their own understanding, or a written explanation of any section they did not draft. That is also the honest answer to the free-rider problem, because it grades comprehension rather than authorship, and comprehension is what a group mark was always a proxy for.

Martínez-Peláez et al. (2025) screened 203 studies from 2023 and 2024 in the Web of Science Core Collection and included 22 under a PRISMA 2020 protocol registered in PROSPERO. They describe LLMs acting as catalysts for collaboration: idea generation, organization, peer feedback, simulated rubric-based evaluations and expert reviews, and lower participation barriers. For problem solving they report help with alternative solutions, interdisciplinary perspectives and authentic scenarios. They also note a useful inversion: LLMs often produce incomplete or incorrect responses, which pushes students to question, verify and improve the material. That exchange is what to listen for at the individual checkpoint.

Their limitations travel with the citation: a single database, journal articles only, no attention to ethics or privacy, the first two years of the technology, and no risk-of-bias instrument applied. Read "catalysts" as a description of observed practice, not a measured outcome.

Detection is the wrong instrument for sorting individual contributions: none of these six studies tests detectors in group settings, and the individual-level problems with detection are documented elsewhere in this knowledge base. Restraint needs interpreting before it needs correcting. Chen and Zou caution that a group declining AI may be avoiding something unfamiliar rather than exercising deliberate self-regulation, and the two call for different responses: one is a teaching problem, the other is a considered position you can grade.

Simulated teammates do not settle the question either. Fang (2026) fine-tuned LLM agents on real participant dialogue (LoRA/QLoRA adapters on LLaMA 3.2-3B) to model 48 participants across 3,824 turns and six thematic codes, then compared real and simulated discourse with Epistemic Network Analysis. The simulated network reached a distance of 0.17 from the empirical network, below the 0.30 threshold, with a permutation p-value of 0.65, statistically indistinguishable, though it slightly overemphasized Technical Constraints and Design links and under-represented Data and Performance Parameters. It is a framework paper with medium confidence, and it supports one claim: participant-specific agents can reproduce realistic collaborative discourse. It does not show that students learn more from simulated teammates.

Form the groups on what members believe, not on who is friends

Whether AI becomes a shared reference point or a lookup depends partly on how much your groups already agree. In Wang et al.'s (2026) study, fourteen pre-service teacher groups worked a 60-minute simulation inquiry with an on-demand metacognitive chatbot. Groups whose members held different preconceptions about the physics routed AI use through conceptual talk, with the AI sitting at the center of their interaction network; groups that already agreed ran a linear query-then-answer path and sustained negative evaluation of the output, which the authors partly attribute to proactive prompts arriving after consensus had formed. Both kinds of group asked the same kinds of questions, so what changed was uptake rather than volume.

Two moves follow, and both are cheap:

• Run a short pretest on the core concept and mix the response patterns across groups, rather than letting students self-select into comfortable ones.
• In groups that converge quickly, keep shared-regulation prompts on demand rather than pushed, because a nudge that arrives after agreement reads as interruption.

Treat it as a hypothesis rather than a rule. That is fourteen groups, one 60-minute task on a single topic, pre-service teachers in one country, and transcripts translated from Finnish before coding.

What instructors and administrators get wrong about this

A clear course policy settles what groups do. Chen and Zou watched groups reach opposite conclusions inside one permissive policy. The task, the rubric and the group's own reading of the risk decided it.

A collaboration criterion produces collaboration. Their rubric named integration, coherence and demonstrated collaboration, and three groups still partitioned the work and reassembled it. Criteria describe the product; sequencing produces the interaction.

Group work's real problem is free-riding, and AI makes it worse. No study here shows that AI has worsened free-riding, or that any intervention reduces it. The fairness complaint predates the tool: Chen and Zou's restrained students treated generating their own section through AI as unfair to the groupmates who would carry it, and the performance rationale shows students protecting their share of a shared grade. Individual accountability and visible process are defensible on design grounds, and they are not proven fixes.

More AI in group work means better collaboration. Martínez-Peláez et al. describe genuine supports, and Wei and Perkins found the benefit side empirically supported while the risk side remains largely conceptual. The sharpest caution in the review is the randomized evidence showing AI produced more innovative suggestions without significantly raising participants' overall innovativeness, while homogenizing ideas and inviting cognitive offloading. Gains measured during AI-supported group work also have not been shown to transfer to collaboration without the tool.

A shared document means a shared process. Korchak, Costley and Fanguy found accounts that contradicted the written reflections, and strategies that not every member could see. Look for the decision log, not the folder.

Detection can tell you who did what. None of these studies tests it in a group setting, and a group artifact is the worst case for it: several writers, one file, and a revision history that a shared account flattens.

The objections you will hear

"Group work is already unfair, and AI makes it worse." AI changes the shape of the problem, and the fairness complaint was already inside the group before anyone opened a chatbot. What is new is how cheaply a member can produce a plausible section alone. An individual checkpoint plus a visible process has a design rationale behind it. Neither has been shown to reduce free-riding.

"I have 300 students. I cannot run oral defenses." Then do not run 300 of them. Put the individual explanation in writing for everyone, and sample oral checks where the submitted work does not match earlier work. The point is that every member knows a conversation is possible.

"I cannot grade process." You are already grading a proxy for it and calling the result product quality. Grade the artifact that exists only if the process happened: the negotiated AI-use statement, the decision log, the short individual explanation. All three are bounded and checkable, and the rubric lines are no vaguer than the ones you already write for coherence.

"Students will prompt privately whatever I specify." Some will, which is why the shared interface and the submitted statement carry more weight than the rule. Xu et al. found that a visible shared interface changed what teams did with the tool, and Chen and Zou found that a permissive climate raised open use. What you cannot do is infer the private case from the group file, so design for the visible one.

"Isn't this a policy question for the department?" The department can set the default. It cannot sequence your assignment. Chen and Zou's finding is that the same institutional policy produced different group behavior depending on the task, which puts the decision in your hands whether or not you wanted it there.

What to do this week

1. Add a half-page, jointly authored AI-use statement to the next group brief, and give the reasoning marks. Include a line for what the group decided not to use AI for.

2. Find the stage of the task that can be cut into independent pieces and reassembled. Resequence so one stage requires the whole team at once, and make that stage produce something you collect.

3. Add one individual checkpoint to the rubric: a short written explanation of a section the student did not draft, or a two-minute oral defense.

4. Name the AI-free segments in the brief and say why they are AI-free, in one sentence each.

5. Decide where the prompts will live. If the answer is "wherever each student keeps them," you have chosen private prompting without saying so.

Where to go next

Group Work and Collaborative Learning carry the underlying evidence on group assessment and on what makes collaboration productive. How Do I Write a Course AI Policy and Communicate It to Students? covers the course-level wording students actually read, How Do I Redesign Assessment So That a Grade Still Tells Me Something Defensible About What the Student Knows or Can Do? covers the assessment redesign this sits inside, and Academic Integrity is where the misconduct side is handled.