Abstract
The diffusion of generative
artificial intelligence (AI) into classrooms has produced two competing
narratives: that AI is the most powerful personalisation engine education has
ever had, and that reliance on machine-generated answers hollows out the
capacities education exists to build. This study tested both in a single
framework. A cross-sectional survey of 420 higher education students measured
AI usage intensity, cognitive offloading, critical thinking disposition,
metacognitive self-regulation, AI literacy, perceived learning benefit and
academic performance (α = .851 to .910), analysed through correlation, t tests,
ANOVA, chi-square, multiple regression and bootstrapped mediation. A clear
paradox emerged. Perceived benefit was high (M = 3.89, SD = 0.65), yet AI usage
correlated negatively with critical thinking (r = -.734, p < .001) and
positively with cognitive offloading (r = .685, p < .001). Heavy users
scored far lower on critical thinking than light users, F(2, 417) = 172.01, p
< .001, η² = .452. Regression explained 73.8% of variance in critical
thinking, with metacognitive self-regulation the strongest positive predictor
(β = .408) and AI literacy non-significant. Offloading carried 39.6% of the
total effect (bootstrap 95% CI [-.385, -.248]).