Predictive analysis of cognitive dependence level associated with generative Artificial Intelligence use in university students through supervised classification algorithms

Abstract

Introduction: The adoption and use of generative artificial intelligence (GAI) tools has become increasingly common in the academic activities of university students. Current scientific debate focuses on understanding how the intensive use of these technologies may influence intellectual autonomy, critical thinking, and the delegation of cognitive tasks.
Objective: To construct and compare three supervised classification models capable of predicting the level of perceived cognitive dependence among students at the University of La Guajira, based on sociodemographic, technological, usage pattern, and academic perception variables, incorporating a variable engineering process.
Method: A structured questionnaire was administered to 299 university students during the 2026-1 academic period. The research process followed the CRISP-DM methodology. The target variable was constructed as a summative index based on two Likert-scale items: loss of critical thinking and decreased autonomy, categorized into three levels: low, medium, and high. Eleven derived variables were generated through variable engineering, and SMOTE was applied to balance the training classes. Decision Tree, k-NN, and Random Forest models were trained using GridSearchCV for hyperparameter optimization.
Results: The Random Forest model achieved the best performance in the testing phase, with an accuracy of 0.89 and an F1-macro score of 0.84. In addition, the multiclass ROC curve showed AUC values of 0.99, 0.96, and 0.96 for the three classes. Seven of the eleven engineered variables were among the top 20 most important predictors, with the ratio of cognitive replacement tasks standing out as a particularly relevant variable.
Conclusions: Perceived cognitive dependence on generative artificial intelligence is predictable from sociotechnical variables. The results suggest that the strongest predictors are not related to the frequency of use, but rather to the type of tasks delegated to AI. These findings provide useful evidence for designing institutional policies aimed at promoting the responsible and critical use of generative AI in higher education.