An AI-Assisted Design That Cuts Free-Riding in Peer Assessment

In peer assessment of group work, students mostly give one another a perfect five. It is hard to hand a low score to a close friend, and once you think about your own turn to be rated, you have little choice but to be generous. When relationships decide the score, peer assessment becomes meaningless. Everyone effectively gets full marks, so nothing is distinguished at all. AI can be put to work reinforcing the objective basis of a rating and as a safeguard that screens out throwaway feedback. Making students look at reasons rather than numbers is the starting point.

A structure that blocks score collusion

When you collect only a score, colluding is easy. Requiring written reasons raises the cost of collusion, because inventing false reasons is more trouble than simply handing out full marks. The design points are these.

• Require at least two specific reasons written beside the score.
• Have AI check whether those reasons connect to the assessment criteria. Irrelevant reasons like “was kind” or “tried hard” get sent back to be rewritten.
• Keep every rating anonymous among students, while the teacher can see who is being careless.

For peer assessment of a presentation, for instance, AI steers students toward observable behavior — “named three sources for the material and answered questions on the spot” — instead of “spoke loudly, five points.” The moment abstract praise turns into observed behavior, the assessment comes alive.

Using AI for reliability checks

1. AI flags extreme scores clustered on one student, meaning all top marks or all bottom marks.
2. It compares how consistent the reasons are across the several ratings of the same person. If only one rater saw it very differently, find out why.
3. The teacher looks only at the groups where a warning signal came up. There is no need to review all of them.

Peer assessment is not there to produce a score. It is practice in students internalizing the standard for good work.

A checklist for the classroom

• Read the rubric together before rating and do a practice scoring on a sample answer. Rate without criteria and you get impression scores.
• Do not let students rate their own group. This is the basic guard against a conflict of interest.
• Use peer scores as only part of the weighting in a grade and let the teacher handle the rest. Somewhere around 20 to 30 percent from peers usually sits comfortably.
• For a student who is careless again and again, make clear in a one-on-one conversation that “assessing is part of the assignment too.”

Teaching good peer feedback

Students have never been taught how to give good feedback. So all they leave behind is “nice job” or “not great.” Teach the feedback itself before the rating and the quality of peer assessment changes. Handing them the following frame works well.

1. One thing that worked, written specifically. Not “the opening was interesting” but “you led with a statistic in the first sentence, which caught my attention.”
2. One thing worth fixing, written as a suggestion. Not “this isn't good” but “one more line of evidence in the conclusion would make it more convincing.”
3. One thing you wondered about, left as a question. A question makes the reviewer a reader rather than a critic.

Hand out just those three boxes and the careless one-line rating disappears. The student doing the rating also has to read someone else's work closely, so in the end their eye for their own writing grows along with it.

Key takeaways

The enemies of peer assessment are relationships and carelessness. Use AI to require reasons and to flag anomalous scores, and the exercise turns from a popularity vote into a learning tool. What matters is a design that makes students look at reasons rather than scores. When it runs well, students pick up the standard for good work themselves in the act of rating a classmate, and that standard carries straight into their own next assignment.