AI Assisted Feedback: Findings from a second, formative pilot at UiT
THIS IS RELEVANT TO YOU BECAUSE:
- You're under pressure to make feedback more frequent and personal without adding hours to your assessors' workload, and you need evidence that AI-assisted feedback actually holds up in a real course, not just in a vendor demo.
- You ran (or read about) UNIwise's first pilot and remember its biggest weakness: feedback landed after the exam, when it was too late to help. You want to know whether that's actually fixable.
- You're weighing whether AI-generated feedback can be trusted without a human check on every single output, and what that oversight looks like in practice, week to week, submission to submission.
- You're responsible for formative assessment strategy and want a low-stakes, low-risk way to introduce AI support before considering anything closer to grading or high-stakes use.
- You want to hear directly from students, not just satisfaction scores, about what actually helped them, what didn't, and how comfortable they are with an AI writing feedback on their work.
A FIRST LOOK AT WHAT'S INSIDE...
The first UiT pilot answered the "is the output any good" question. This second pilot asks the harder one: does AI-assisted feedback actually work as a learning tool, delivered at the moment it can still change what a student does before the exam?
At UiT's Arctic University of Norway, 23 medicine students sat a formative immunology test that didn't count toward their grade. Every submission got AI-generated feedback, reviewed and screened by Professor Tor Brynjar Stuge before release, and returned four days later, well within the teaching period. The result: 13 of 14 students said the feedback arrived at a useful point for learning, and every single respondent said it helped them understand what to improve. The dimension students criticised most sharply in the first pilot became the dimension they praised most in the second.
The report walks through what changed to make that happen, including a refined, explicitly formative prompt that forces concrete next-step suggestions and bounds the level of detail to the marking guide. It sets out Tor Brynjar Stuge's own assessment as the human in the loop: strong ratings on accuracy, consistency and tone, an estimated fourfold cut in time per submission, and his account of what happened when students questioned specific pieces of feedback (spoiler: on review, the AI was right more often than the students expected). It also doesn't shy away from what still needs work, from students wanting a clearer sense of which gaps matter most, to the honest limits of a single-course, self-selected pilot.
If you're thinking about where to introduce AI into your own assessment workflow, this is a candid look at what a low-stakes, human-supervised starting point can look like, and what it takes to make feedback that students genuinely act on.