AI has changed what an assessment needs to prove. When a polished final product can be generated in seconds, the product alone no longer tells you much about what a student has learned, which is why so many institutions are shifting their attention from outputs to the learning process itself. That shift was the focus of our recent webinar with EDHEC Business School, and this article gives you the highlights. You can also watch the full replay to hear it from the speakers themselves.

Meet the speakers

Claudia Carrone is a Digital Learning Manager at EDHEC Business School, part of the Pedagogical Innovation Laboratory (PILab), and winner of the FeedbackFruits Learning Design Community Award 2025.

Michelle Sisto is Associate Professor, Associate Dean, and Director of the EDHEC AI Centre. She taught data and analytics for more than 20 years before making AI the central focus of her career, and now leads EDHEC's AI integration strategy.

Jack Bainbridge is a Partner Manager at FeedbackFruits, supporting institutions to get the most out of the Learning Design System.

Why this matters now

Detection tools have not solved the AI assessment problem. They are unreliable, they signal distrust, and students adapt faster than the tools do. A more durable answer is to assess the learning process and the quality of a student's judgment, and peer review is built for exactly that, because judging someone else's work forces you to understand the standard you are judging against.

The stakes are real. NACE research shows students rate their own critical thinking roughly 25 points higher than employers do, and in a YouScience survey, 86% of employers say new hires need additional training before they add value. The evidence for peer assessment is equally strong. A meta-analysis of 54 studies finds it improves performance across subjects, and on FeedbackFruits activities students rate the feedback they receive 8.87 out of 10 across 5.63 million ratings, rising to 9.17 when the feedback itself is graded. That design move, grading the quality of the feedback students give, sits at the heart of this case.

The activity: build, evaluate, improve

EDHEC asks each student to build an AI assistant, something like a custom GPT or Copilot agent, for a task they care about: a marathon coach, an interview preparation assistant, a quiz generator built on course material. The assignment runs in three phases:

1. Build together. In class, students and faculty create a first assistant side by side to learn the basics.
2. Evaluate and give peer feedback. Students build their own assistant, then test and review two peers' assistants with specific, constructive comments.
3. Revise and submit version two. Students improve their work based on the feedback, and the final grade goes to that improved version.

Feedback becomes part of the learning process rather than a verdict at the end, and faculty grade work that has already been challenged and improved.

The moment the assessment changed

Left alone, students test the happy path. They ask their assistant friendly questions it can obviously answer, and everything looks great. Real users are different: they bring ambiguous requests, push systems out of scope, and sometimes try to break them, and that is when assistants hallucinate or misbehave. Michelle grounds this with real cases, including a widely reported one where an airline was held legally responsible for a refund its chatbot wrongly promised.

So the peer review asks students to actively try to make each other's assistants fail, using four criteria: interaction and user experience, accuracy and reliability, effectiveness of guardrails, and language and tone. The comments that result are wonderfully specific. One reviewer got an assistant talking about cats in space by confusing it, another flagged that a finance assistant refused an off-topic request but never explained its actual scope, and another tested tone with slang and watched the assistant mirror it right back. For ready-to-share language for your own students, see our peer feedback examples for students.

Scaling from 25 to 1,000 students

Michelle used to run peer feedback on paper, with scissors and a stapler. That works at 40 students, not at the nearly 1,000 this course reached across four sections. EDHEC scaled the activity from about 25 students to 50 to almost 1,000, with the AI Centre defining the learning goal, PILab designing the activity, and FeedbackFruits providing the structure: rubrics, feedback rounds, evidence collection, and completion tracking. The Peer Assessment solution page shows how those pieces fit together.

Grading stays light. Michelle scores each student's feedback on a simple scale, rewarding identified risks and concrete suggestions, and no longer needs to test every assistant herself because students surface the failures for each other. To handle the volume of comments, the team even used AI to surface common themes across the qualitative feedback, a fitting case of AI supporting judgment rather than replacing it.

What students take away

In focus groups afterward, students consistently said that writing clear instructions and guardrails was harder than expected, that iteration stopped feeling like failure, and that they learned as much from reviewing a peer's assistant as from building their own. They also left with a more realistic, critical view of AI, which matters: an EDHEC survey found about one in five students felt undermined by AI, and this activity showed them where their own judgment adds value. One student said it finally felt like real learning.

The structure adapts well beyond AI too. Michelle first used it in statistics, and EDHEC has applied it to leadership and coaching courses, including fully online. Our authentic assessment learning journey walks through the wider course design.

Try it in your own course

EDHEC has shared the complete activity as a reusable template, with the learning objectives, instructions, rubric, and feedback-on-feedback settings: Peer review of student-created AI assistants. Setup takes about 2 to 3 hours in FeedbackFruits, and it can replace 8 to 12 hours of marking for a cohort of 100 students. Most of that setup time is the thinking, not the tool, so time spent clarifying your criteria pays for itself.

For everything in one place, the peer assessment get started bundle brings together templates, guidance, and support, and our complete guide to peer assessment and the ebook A guide to authentic peer assessment in higher education cover the research and the roadmap. And do watch the webinar replay, which includes the audience Q&A on scaling, online delivery, and teaching students to find edge cases.

Final thought

Peer review in an AI assistant assignment does more than assess output. It develops critical thinking, builds confidence, and moves assessment into the learning process itself. The lesson from EDHEC is clear: students do not only need to build AI solutions, they need to evaluate them, test them, challenge them, and improve them.

Watch the webinar replay now