Does artificial intelligence predict scientific productivity? Machine learning evidence from university research seedbeds

Reyes Reyes C.A.; Rojas Jiménez K.I.; Muñoz Gastelo N.E.; Ramos Farroñán E.V. · Frontiers in Education (2026) · DOI: 10.3389/feduc.2026.1864949

Resumen

This study examines the relative predictive contribution of artificial intelligence (AI) learning and use, compared with academic trajectory and sociodemographic variables, in explaining scientific productivity profiles among undergraduate students participating in university research seedbeds in an emerging Latin American context. A quantitative, non-experimental, cross-sectional design with supervised predictive modeling was employed, using census data from 57 students at a Peruvian university. Six dimensions of AI learning and use, along with composite indices and academic trajectory variables, were evaluated against a binary scientific productivity profile. Four machine learning classifiers (penalized logistic regression, Random Forest, Extra Trees, and XGBoost) were compared using repeated stratified cross-validation, and model interpretability was assessed through permutation importance. Results indicate that penalized logistic regression achieved the highest predictive performance (accuracy = 0.842; balanced accuracy = 0.845; F1-macro = 0.835; AUC-ROC = 0.894). The most influential predictors were time of participation in the research seedbed, age group, and gender, whereas AI-related composite indices exhibited negligible or marginally negative contributions. These findings suggest that, in this sample, self-reported AI capabilities were not the dominant predictors of scientific productivity once formative trajectory variables were considered. Instead, sustained formative engagement and academic trajectory emerge as stronger determinants, challenging the prevailing assumption of AI as a universal accelerator of research performance. The study highlights the need to reorient institutional strategies toward strengthening research training environments and integrating AI within the full scientific production cycle.

Instituciones: Universidad César Vallejo; Universidad Señor de Sipán