Skip to content

Guides

The check that goes missing in a group

By 8 min read

For anyone who sets group work in a course where AI is allowed: a study being presented this week found that students reported checking the AI's work far less once they were in teams, and a second study found that early signs in a group's conversation predicted how it went.

Ninety-six university students reported, assignment by assignment, how they used AI across one course. On their own homework, the ones using it mostly said they checked what it handed them. Then they were put into teams, and among the students still using AI, the share who said they tested its output fell from 59.5 percent to 20.5 percent (Nizamani, Ziew, Nizamani and Goyal, arXiv preprint, revised 28 July 2026).

Those are self-reports, so they measure what students said they did. The paper is on the programme at SIGCSE Virtual, the Association for Computing Machinery's online computing education conference, which runs from 12 to 15 November (SIGCSE Virtual 2026). It comes from a computing course, and it travels to any room where students get a group assignment and a chatbot in the same week. Most of us assume a group works as a second pair of eyes on whatever the AI produces. In this course, fewer eyes were on it once the group formed.

What changed when the teams formed

The study followed an upper-division data analytics course at a large US public research university through six assignments, alternating individual homework with team project milestones, and asked students after each one how they had used AI. It's one institution, and the authors say plainly that it can't establish cause.

Use dropped hardest at the first team milestone, from 82.3 percent of students on the first homework the study surveyed to 39.6 percent on the team planning milestone. Part of that is the task, since that milestone involved no coding, and the authors say so. Use came back on the later milestones. By the authors' reading, the checking never fully came back with it. The 59.5 and 20.5 figures count only students who used AI; counted across everyone who answered, including those who skipped it, the share who tested fell from 33.2 to 13.8 percent, which is the smaller drop the abstract reports (full text). Followed student by student, 18 who had used AI steadily on their own homework stopped using it once they were in a team, and only 3 went the other way.

The authors' best guess at the mechanism is one anybody who has run group work could have written on a sticky note. They call it diffusion of responsibility, "the implicit assumption that someone else will catch errors." It's the office fridge problem, honestly. Everyone uses it, and the thing at the back that has gone off belongs to nobody.

An interview study from last year shows the same drift from the help-seeking side. Hou and colleagues interviewed 17 undergraduate computing students at seven North American research universities and found that "students often redirected questions from their peers to generative AI instead of providing assistance themselves" (Hou et al., ITiCSE 2025, arXiv preprint 14 April 2025). Seventeen interviews is a small sample, and an interview study reports what students said about their habits. Still, it describes a study group where the group keeps meeting and the helping has moved somewhere else.

Everyone uses it, and the thing at the back that has gone off belongs to nobody.

What a group that works looks like

A second study, smaller and also a preprint, keeps this from turning into an argument against group work. Caitlin Morris and Pattie Maes at MIT had 36 undergraduates learn graph theory, 24 through peer collaboration and 12 with an AI assistant (Morris and Maes, arXiv preprint, 16 January 2026). High-quality peer interactions produced curiosity and engagement the AI didn't match. Low-quality peer interactions did worse than the AI across the dimensions they measured. The AI raised students' confidence in what they knew while lowering their curiosity, and students treated it as a transactional information source even though it had been designed to collaborate.

Two details in that paper matter for anyone setting group work. The authors trace the quality to patterns in the conversation rather than to the students' individual traits. And early markers in the conversation predicted the outcome, so you may not have to wait for the final report to find out which kind of group you've got. My guess, and it's only that, is that much of it shows up in the first meeting.

Put the two papers side by side and here is my reading, which neither of them states. A study group with AI in it can look productive on every visible measure, slides made and code running, while the conversation that would have caught a wrong answer never happens. That's the group that only looks like it's working, and the confidence the AI hands out makes it harder to spot from inside.

Giving the check somewhere to live

A design-based study published on 14 October offers one structural move. Felix Höpfl, at the Rosenheim Technical University of Applied Sciences in Germany, kept AI out of the Think phase of Think-Pair-Share and allowed it only from Pair onward, after noticing that with AI allowed, "the Think phase no longer appeared to function as independent reflection" (Höpfl, Frontiers in Education 11, 14 October 2026). Across four administrations in three student groups, 50 of 55 responses (90.9 percent) agreed the AI-free Think phase was appropriate. Höpfl is careful to call that an endorsement finding: students accepted the design, and whether it improves learning is still untested.

The Nizamani data puts the drop at the moment a team forms, so that's where I'd aim. Here's a checklist for the first team milestone. It's mine and I haven't tested it, so try it on one assignment and see what breaks.

  • Put a solo step before the group meets. Each person writes down their own answer or approach, without AI, before the first meeting. The group then has something of its own to hold the AI's output against, which is Höpfl's AI-free Think stretched across a project.
  • Name the checker for each milestone, in writing. One person owns verifying anything the AI produced for that milestone, and the job rotates. In an essay or a case analysis, checking means tracing each claim the AI supplied back to a source the group has actually opened.
  • Ask for one sentence on how it was checked. Put it inside the submission: what was checked and how. A group that skipped the check finds that sentence hard to write, which is the point of asking.
  • Hear the first ten minutes, somehow. Morris and Maes found early markers in the conversation predicted outcomes. In a big class, a live drop-in on every group is a lot to ask, so a two-line note from each group after its first meeting, saying what they disagreed about, is a cheaper window.

If that sentence comes back blank, that's worth a conversation with the group.

How this was made. Full disclosure, and I mean it as a courtesy, not a confession. AI tools help me research, check my claims, tidy my grammar, and stay on top of far more reading than one person can. The thinking and the judgment are mine, I edit every word into my own voice by hand, and then I put my name on it. Telling you that is the whole point of keeping it human.

One question I'd like answered, because I don't have a good answer of my own. If you run group work in a course where AI is allowed, who in the group is responsible for checking what the AI produced, and how would you find out if nobody did? I'm especially curious about the groups that looked fine right up until they weren't.

Next edition: the notes students take, now that a transcript can take them instead.

These editions run alongside Learning to Learn With AI, a longer series working the same ground in order.

Common questions

Do students check AI output less when they work in teams? In one US data analytics course, students who used an LLM reported testing its output far less on team work than on individual work: 59.5 percent fell to 20.5 percent. The figures are self-reports from one course, and the authors say the study can't establish cause.

What is diffusion of responsibility? It's the authors' best guess at the mechanism: the implicit assumption that someone else will catch errors.

What can an instructor change in a group assignment? The edition's checklist, untested, has four items: a solo step before the group meets, a named and rotating checker for each milestone, one sentence in the submission on how the AI's work was checked, and a short note after each group's first meeting.

References

  • Nizamani, S. B., Ziew, Z., Nizamani, S., & Goyal, K. (2026, 28 July; v1 29 June). Less Deliberate in Teams: Student LLM Use Across Individual and Collaborative Work [Preprint]. arXiv:2606.30860. Accepted at SIGCSE Virtual 2026. https://arxiv.org/abs/2606.30860
  • ACM SIGCSE. (2026). SIGCSE Virtual 2026: The 2nd ACM Virtual Global Computing Education Conference, 12 to 15 November 2026, online. https://sigcsevirtual2026.acm.org/
  • Hou, I., Man, O., Hamilton, K., Muthusekaran, S., Johnykutty, J., Zadeh, L., & MacNeil, S. (2025, 14 April). "All Roads Lead to ChatGPT": How Generative AI is Eroding Social Interactions and Student Learning Communities [Preprint]. arXiv:2504.09779. Published in the Proceedings of ITiCSE 2025. https://arxiv.org/abs/2504.09779
  • Morris, C., & Maes, P. (2026, 16 January). When Peers Outperform AI (and When They Don't): Interaction Quality Over Modality [Preprint]. arXiv:2601.11777. https://arxiv.org/abs/2601.11777
  • Höpfl, F. (2026, 14 October). Redesigning collaborative teaching for human–AI partnership: a phase-based AI-constraint design principle in higher education. Frontiers in Education, 11. https://doi.org/10.3389/feduc.2026.1963902