Cognitive load, critical evaluation, and translation learning during AI-assisted translation tasks

AI translation tools are ubiquitous in L2 learning, yet the mechanisms through which cognitive load and AI reliance jointly shape translation learning remain poorly specified. This study applies structural causal modeling (SCM) with directed acyclic graph (DAG) identification, bootstrapped mediation, sensitivity analysis, and counterfactual inference to cross-sectional data from 200 adult L2 learners across three text-complexity conditions. Given that the design is observational, estimates are causal conditional on the identification assumptions encoded in the DAG, which are stated and probed explicitly. Intrinsic load showed no direct association with critical evaluation once AI reliance was adjusted for (direct effect b = + 0.25, 95% CI [− 1.08, + 1.51], p = 0.718), while the indirect path through AI reliance was robust (indirect effect b = − 2.63, 95% CI [− 3.63, − 1.74]; Sobel z = − 5.53, p < 0.001) against a total effect of b = − 2.38 (95% CI [− 3.70, − 1.18]). AI reliance moderated the load–evaluation relationship, the intrinsic-load slope falling from − 7.03 at low reliance to a non-significant + 0.85 at high reliance. Counterfactuals propagated through the mediator indicate that reducing reliance to the lower end of the observed range would raise vocabulary retention by 33.94 points (95% CI [29.49, 38.75]) and translation accuracy by 9.84 points (95% CI [4.77, 14.97]). Accuracy followed an inverted U peaking at 3.78 (bootstrap 95% CI [2.95, 4.35]), so eliminating reliance was not accuracy-optimal (ATE = + 3.93, ns). Sensitivity analysis showed a confounder would need to explain 41.2% of residual variance in both reliance and evaluation to nullify their association.