Multinational machine learning investigation of whether generative AI interactions augment or substitute cognitive skills and study efficiency in global higher education

While GenAI has been rapidly integrated into higher education, little empirical evidence explains how specific usage patterns drive cognitive augmentation versus cognitive offloading remains limited, as most studies treat usage uniformly. This study fills that gap by examining how GenAI habits, metacognitive behaviour, and emotions jointly predict cognitive skills and study efficiency as a distinct outcome across a global context. A secondary analysis was conducted on a publicly available cross-sectional multinational survey dataset of N = 14,949 higher education students, drawn from an original sample of 22,963 respondents across 120 countries. A five-layer analytical framework was applied: (1) measurement model validation using Confirmatory Factor Analysis (CFA) and Fornell-Larcker criterion; (2) multivariate OLS regression was used to predict critical thinking gain; (3) benchmarking of four machine learning algorithm (Linear Regression, Random Forest, XGBoost and Gradient Boosting) was used on a stratified 70/30 train-test spilt; (4) SHapley Additive exPlanations (SHAP) analysis was used on best performing model; and (5) K-Mean clustering (k = 4) was used across the three dimensions space of Cognitive Skill, Study Efficiency, and GenAI Usage. Robustness was assessed through five-fold cross-validation, discipline-stratified and subgroup analysis. The best-performing machine learning models demonstrated moderate predictive performance, explaining 28.18% of the variance in Cognitive Skill and 30.00% of the variance in Study Efficiency. The SHAP analysis result shows that Self-testing Habits (SHAP = 0.128) was the single strongest predictor of cognitive skill development. GenAI Usage Frequency (SHAP = 0.184) was the dominant driver of Study Efficiency. Four empirically distinct students’ interaction profiles were identified: (1) Augmentor (high skill, high efficiency, highest GenAI use; n = 2900); (2) Optimiser (moderate balance; n = 5500); (3) Traditionalist (efficient with lower GenAI use; n = 3000); and Substitutor (disengaged, low performance; n = 2000). Institutional funding type was a significant predictor of gains in study efficiency. The study concluded that effective human-GenAI interaction is not a function of usage frequency alone but of interaction quality, task type, and metacognitive structure within which AI tools are embedded. The cognitive skill efficiency trade-off is real but avoidable: Augmentor-profile students demonstrate that high efficiency and high cognitive skill co-exist when GenAI use is intensive, structured, and metacognitively scaffolded. Higher education institutions must move beyond binary use/no-use policies toward evidence-based frameworks that promote generative, metacognitively embedded interaction to maximise cognitive augmentation and mitigate passive substitution.