Assessing the impact of generative AI on undergraduate thesis quality: A comparative study of students and teachers.

Abstract

Generative AI (GenAI) is increasingly embedded in undergraduate thesis work, intensifying concerns about thesis quality. However, limited evidence is available on how GenAI engagement relates to teachers' and students' evaluations of thesis quality and whether these associations differ across institutional and disciplinary contexts. Using a stratified random sampling method across institutional tiers and disciplines, 934 participants were recruited (684 graduating students and 250 thesis teachers). Key variables (Extent of GenAI Involvement (EX), Perceived Effect on Thesis Quality (EF), Perceived Problems/Risks (PR), Attitudes Toward GenAI Use (AT), Perceived Writing Ability Development (AB)) were measured using structured scales, then five-step hierarchical regression analysis was employed to estimate main effects and test interactions. Results showed that EX (B = 0.200, p < .001) and AB (B = 0.185, p < .001) were positively associated with EF; AT showed a marginal association (B = 0.037, p = .053) and policy presence showed a small positive association (B = 0.118, p = .001). EX/AB/PR/AT and Group interactions increased explanatory power (R2 = .445; [Formula: see text]), PR was not significant (p = .220). Policy did not moderate Group differences ([Formula: see text], p = .684). Institutional tier and Group interactions further improved fit ([Formula: see text]), strongest in World-Class Universities (B = 0.986, p < .001). Disciplinary-category and Group interactions added incremental variance ([Formula: see text]; final R2 = .510), with the largest teacher-student gap in Natural Sciences. The findings revealed that EF was most consistently linked to EX and AB, with systematic heterogeneity by group and by institutional and disciplinary context, underscoring the need for differentiated guidance on policy-compliant, capability-oriented GenAI use; however, given the cross-sectional and self-reported design, EF captures perceived thesis quality.