BETWEEN EFFICIENCY AND JUDGMENT:

Abstract

The question of how artificial intelligence can support academic assessment without displacing the evaluator has moved from theoretical debate to operational reality faster than most institutions have had time to think carefully about it. This paper presents evidence from two empirical pilots of MeritMinds, an AI-assisted academic assessment platform developed and tested in European higher education contexts, to examine what responsible AI integration in assessment actually looks like in practice. The first pilot involved 20 bachelor's theses submitted to a Faculty of Modern Applied Languages at a Romanian university (cohort mean score: 58.5%). The second involved collaboration with an independent lecturer at a European institution. Across both pilots, the platform supported evaluator judgment through rubric formalization, criterion-level consistency checking, and cohort-level pattern detection — not to generate grades or replace evaluator decision-making. The paper draws on these pilots to address three questions institutions rarely ask before adopting AI assessment tools: what does the system actually decide, and what does the human evaluator decide? How is accountability distributed when AI and human judgment diverge? And what does a partial failure — one pilot produced results requiring diagnostic rather than operational interpretation — reveal about the conditions for responsible deployment? The paper argues that these questions are not technical but pedagogical and institutional, and that answering them honestly is a precondition for any meaningful framework for AI-assisted assessment in higher education.