A machine learning analysis of nearly 15,000 university students across more than 40 countries has found that the cognitive benefits of generative artificial intelligence depend far less on how often students use the tools than on the metacognitive habits that surround that use. The study, published in Discover Artificial Intelligence, applied a five-layer analytical framework to a multinational survey dataset and identified four distinct student interaction profiles, including one group that combined the highest levels of AI use with the highest self-reported cognitive skill and study efficiency. The central message is provocative: the feared trade-off between learning faster and learning deeply is real but not inevitable.
The research team, led by Oladagba Stephen Bolatimi, conducted a secondary analysis of a publicly available cross-sectional survey originally collected between October 2024 and February 2025 from 22,963 higher education students in 120 countries. After a systematic cleaning protocol that removed non-users of ChatGPT, records with excessive missing data, and statistical outliers, the final analytical sample comprised 14,949 valid responses, or 65.1 percent of the original dataset. The survey was administered in seven languages, including English, Spanish, Turkish, Japanese, Arabic, Italian, and Hebrew, and targeted students enrolled at undergraduate, postgraduate, and doctoral levels in accredited institutions.
The study was designed to resolve a persistent ambiguity in the literature. Most prior research treats generative AI interaction as a single homogeneous construct measured by frequency, which obscures the very different cognitive consequences of academic writing, coding, summarisation, and note generation. The authors also noted that no earlier study had simultaneously operationalised and predicted both study efficiency and cognitive skill as separate outcome constructs within one dataset, and that emotional states such as anxiety, confidence, and frustration are rarely incorporated as meaningful predictive pathways in quantitative models of AI-assisted learning.
Methodologically, the analysis proceeded in five stages. First, confirmatory factor analysis validated the measurement model, with all constructs exceeding thresholds for internal consistency and convergent validity, and the Fornell-Larcker criterion confirming that generative AI interaction, study habits, emotional response, and cognitive skill were statistically distinct. Fit indices including CFI, TLI, GFI, AGFI, and NFI met or closely approached the recommended 0.90 benchmark, while RMSEA of 0.064 indicated good fit. Second, multivariate ordinary least squares regression established baseline predictors of critical thinking gain. Third, four machine learning algorithms, namely Linear Regression, Random Forest, XGBoost, and Gradient Boosting, were benchmarked on a stratified 70/30 train-test split, with a training set of 9,548 cases and a held-out test set of 4,092.
The best-performing models delivered moderate but consistent predictive power. Gradient Boosting explained 28.18 percent of the variance in cognitive skill and 30.00 percent of the variance in study efficiency on the test set, with Random Forest nearly tied on the efficiency target at an R-squared of 0.2995. Five-fold cross-validation confirmed stability, yielding mean R-squared values of 0.2442 for cognitive skill and 0.2344 for study efficiency across folds. The convergence of all four algorithms on similar performance suggested that the underlying signal linking AI usage, study habits, and cognitive outcomes is robust, though inherently non-linear and only moderately explainable from survey-based features.
The explainability layer produced the study’s most striking findings. Using SHAP, or SHapley Additive exPlanations, values computed on the best-performing Gradient Boosting models, the researchers found that self-testing habits carried a SHAP value of 0.128, making them the single strongest predictor of cognitive skill development. GenAI-assisted study notes followed at 0.120, and active learning habits at 0.109, indicating a synergistic dynamic between intentional study behaviours and AI engagement. In sharp contrast, study efficiency was dominated by GenAI usage frequency, with a SHAP value of 0.184 that dwarfed the next feature, research skills, at 0.081. The asymmetry is theoretically meaningful: cognitive skill appears to ride on metacognitive structure, while efficiency tracks how frequently and broadly students engage with the tools.
Task type also mattered. In the regression analysis, academic writing through generative AI was the strongest predictor of critical thinking development, with a standardised coefficient of 0.170 and a p-value below 0.001, followed by creative writing at 0.129 and professional writing at 0.052. Confidence and motivation emerged as the strongest emotional predictor at 0.101, while anxiety and positive affect were non-significant in that model. However, in the machine learning analysis of study efficiency, anxiety and worry appeared as a significant negative predictor with a SHAP value of 0.028, highlighting an affective cost of AI integration for some student subgroups and a target for pedagogical intervention.
K-means clustering across the three-dimensional space of cognitive skill, study efficiency, and generative AI usage revealed four empirically distinct learner profiles. The Augmentor cluster, comprising 2,900 students, combined high skill, high efficiency, and the highest AI usage, with a mean usage score of 3.493 on the five-point scale. The Optimiser group, the largest at 5,500 students, showed a moderate balance across all three dimensions. Traditionalists, numbering 3,000, were efficient with lower AI use, while the Substitutor cluster of 2,000 students was disengaged and low-performing on both outcomes. Notably, the Traditionalist profile challenges the assumption that avoiding AI protects learning: non-use was not associated with superior outcomes, and the disengaged Substitutors fared worst of all.
Robustness checks revealed important disciplinary boundaries. When the Gradient Boosting model was re-estimated separately for STEM students (n = 1,301) and social science students (n = 6,025), the R-squared for cognitive skill prediction dropped to 0.144 in STEM compared with 0.241 in social science and 0.282 in the full sample. The authors attribute this to the survey’s writing-centred operationalisation of AI interaction, which captures cognitive processes more effectively in disciplines where writing and research are central, while STEM skill development depends on domain-specific factors such as mathematical reasoning and laboratory competencies that the instrument did not measure. Institutional funding type also emerged as a significant predictor of study efficiency gains, echoing the second-order digital divide: students in privately funded institutions appeared to convert AI access into efficiency benefits at higher rates, consistent with concerns that AI can amplify existing inequality when guidance and infrastructure differ.
The authors are careful about the limits of their evidence. Cognitive skill was measured through self-reported Likert items covering analysis, problem-solving, research, writing, and critical thinking, making it a proxy rather than an objective performance measure. The finding that Augmentors combine heavy AI use with high skill is consistent with augmentation theory but cannot rule out selection effects, since students with stronger prior abilities may simply be more likely to use AI strategically. The sample was also geographically imbalanced, with ten countries, including Romania, Ecuador, Turkey, Italy, Mexico, Egypt, Spain, Brazil, Colombia, and India, jointly representing just over half of the data, and no formal comparison between included and excluded participants was performed. The authors explicitly state that the findings should be interpreted as predictive associations rather than causal relationships, and they call for longitudinal panel designs and objective outcome measures such as GPA trajectories and pre-post assessments.
Even with those caveats, the convergence across regression, machine learning benchmarking, SHAP attribution, and clustering points to a coherent picture with direct implications for institutions still oscillating between bans and unstructured adoption. Effective human-AI interaction, the study concludes, is not a function of usage frequency alone but of interaction quality, task type, and the metacognitive scaffolding within which AI tools are embedded. Students are advised to treat generative AI as cognitive scaffolding within structured, self-regulated routines rather than as a passive information retriever, while educators and policymakers are urged to move beyond binary use-or-prohibit policies toward frameworks that promote generative, metacognitively embedded interaction, with discipline-sensitive designs for STEM and equity-focused provisions for under-resourced institutions.
Subject of Research: How generative AI usage patterns, metacognitive habits, and emotions predict cognitive skills and study efficiency in multinational higher education
Article Title: Multinational machine learning investigation of whether generative AI interactions augment or substitute cognitive skills and study efficiency in global higher education
Article References: Bolatimi, O. S., Ricky-Okine, C., Bolatimi, F. E., Agbesi, C. C. M., Martey, F., Bolatimi, O. R., Peter, A., & Kwaku, E. O. (2026). Multinational machine learning investigation of whether generative AI interactions augment or substitute cognitive skills and study efficiency in global higher education. Discover Artificial Intelligence, 6(1), Article 952. https://doi.org/10.1007/s44163-026-02082-6
Image Credits: AI Generated
DOI: 10.1007/s44163-026-02082-6
Keywords: generative AI, higher education, machine learning, SHAP, cognitive offloading, self-regulated learning, ChatGPT, study efficiency, critical thinking, K-means clustering, educational equity, XGBoost
News Source: Blake Davidson. (October 28, 2026). Massive Global Study Finds Self-Testing, Not AI Frequency, Drives Cognitive Gains in GenAI Learning. Scienmag.
