Business Case September 2026 UiT The Arctic University of Norway
AI Assisted Feedback: Findings from a pilot in two master’s courses at UiT
What it takes to trust AI feedback on quantitative, model-based work
This is relevant to you because:
- You want to know how AI-assisted feedback performs on closed-ended, quantitative coursework, not just open-ended essays.
- You need a documented account of real hallucinations and their root causes before trusting first-pass AI output.
- You want proof that human-in-the-loop review can catch every issue before a student ever sees it.
- You're planning a phased rollout and need to know what training and prompt calibration it actually takes.
Abstract
This pilot tested AI-assisted feedback on technical, model-based master's coursework at UiT - EOQ and MRP calculations in Manufacturing Logistics, and mixed-integer network-design optimisation in Supply Chain Management. Two course coordinators, Associate Professor Xu Sun and Professor Hao Yu, generated AI-assisted feedback for ten submissions and evaluated the process jointly, alongside two students who completed a detailed evaluation survey.
Both coordinators encountered real hallucinations in the AI's first-pass output - invented problems, missed issues, and a misread calculation - and disagreed that the tool was reliable on unedited output. Every hallucination was caught before release: both evaluators screened every response, and no unreliable feedback reached a student. Despite the rocky start, they cut grading time by roughly a third, would use the tool again, and would recommend it to colleagues, provided rollout is gradual and paired with training. Their testing data fed directly into a UNIwise root cause analysis that traces the errors to six specific causes, several of which are already resolved or in progress ahead of the October 2026 release.