More and more students use AI tools to solve homework problems, possibly resulting in worse exam performance. Returning to on-campus assessments is a discussed solution, but–to the best of our knowledge–has not been quantitatively evaluated since the wider spread of AI systems. In this paper, we present tentative results from a cohort comparison of two first-semester CS courses which changed their exam admission policy, which was i.a. motivated by increasing AI generated homework submissions. We compare the results of two subsequent cohorts, one with homework formative assignments, the other one with an on-campus test. The on-campus component is associated with better pass rates and mean grades in both courses, but also with less voluntary homework engagement. We advise putting fewer resources in giving feedback on potentially AI-generated homework answers.
