Teaching students to work with AI
AI in group work
Group work was already the assignment students liked least and instructors trusted least. AI makes the free-rider problem worse in one specific way - and fixes it in another, if you change where the work happens.
The old complaint about group projects is that the deliverable hides who did what. One student writes it, two edit, one disappears, and the grade is the same for everyone. AI sharpens this: now the disappearing student can produce a plausible section in four minutes and nobody can tell.
The fix isn’t surveillance. It’s putting the part that matters - the disagreement - somewhere it can be seen.
A group deliverable tells you what the group agreed on. It tells you nothing about what they disagreed about, which is where all the learning was.
Structure the disagreement
1. Position first, discussion second
Before any group conversation, each student privately commits to a position and a reason. Only then are the positions revealed. This one change does more than any other: it prevents the first confident voice from setting the answer, it gives every student something of their own to defend, and it gives you an individual artifact from a group assignment.
It also makes the spread visible. Students routinely assume everyone agrees with them, and the reveal is usually a surprise.
2. Require the group to record its disagreement
One paragraph: what did you disagree about, and how did you resolve it? Groups that say “we agreed on everything” either didn’t engage or had one person decide. Both are worth knowing.
3. Assign roles with different information
Give each member material the others don’t have. The group can’t converge without actually talking, and no single member can generate the whole deliverable alone - with AI or without it.
4. The second round
After discussion, students revise their position individually and say what changed their mind or why nothing did. “I held my position because X’s objection didn’t address Y” is a better artifact than most group reports.
Where AI sits in a group
Three configurations, each teaching something different:
No AI in the room
Students talk to each other, full stop. Worth doing deliberately rather than by default - there are things you only learn by having to persuade a peer who can push back in real time and whom you’ll see next week.
AI as a resource the group consults
Available, but the group decides when to use it and has to record what they did with the answer. The interesting assessment moment is the disagreement about whether to accept an output - one member wants to use it, another thinks it’s wrong. That exchange is worth more than the deliverable.
AI as an opponent
The group argues a position against a character briefed to attack it from a specific angle. Groups defend their reasoning better against an outsider than against each other, and it surfaces the weak part of the argument fast.
The shared room matters more than the tooling
Most of this only works if the group’s work happens somewhere you can see it rather than in a private chat you’ll never read. A shared class workspace - where students ask, answer each other, and work through the problem together, with an agent available or not at all - turns group process into something observable. How Class Agents & Workspace handles this.
The free-rider problem, specifically
Four things that actually help, in rough order of effectiveness:
- Individual positions before group convergence. A student who contributed nothing has nothing to submit at step one.
- Visible participation in a shared space. Not a word count - a record of who engaged with whose argument.
- An individual follow-up question. One question per student about a part of the deliverable they didn’t write. Takes two minutes, ends most free-riding permanently once students know it’s coming.
- Peer assessment with a specific prompt. “Rate your teammates” produces noise. “Name one thing each teammate contributed that you wouldn’t have thought of” produces information.
Grading it
Split the grade three ways: the deliverable (shared), the individual positions before and after discussion (individual), and the quality of engagement with others’ arguments (individual). The middle component is the one that makes group work fair, and it’s the one most courses omit.
If AI was in the room, add the use-of-AI criteria to the group component - particularly the one about what the group decided not to accept.
Common questions
Doesn’t this add a lot of grading?
The individual positions are short - a sentence or two each, scored on a three-point scale. It’s less work than reading four versions of the same group report and trying to infer who wrote what.
What about asynchronous or online groups?
This structure suits them better than a live meeting does, because the position-then-reveal sequence doesn’t require everyone in a room at once. More on asynchronous formats.
How big should groups be?
Three to four. At five the disagreement stops being recorded because it’s easier to defer, and free-riding becomes much cheaper.
Can students just use AI to generate their individual position?
They can try. It reads as generic, it doesn’t reference the specific materials the group was given, and it falls apart at the second round when they have to say what changed their mind. Ask for specifics tied to your case and the shortcut stops paying.
See what this looks like in a real assignment
Kova turns a sentence into immersive, auto-graded coursework - AI conversations students have to argue their way through, models they build and defend, transcripts scored on the reasoning rather than the artifact.
The full guideHow Conjure builds itRequest a demo