Critical Artificial Intelligence Literacy and Educational Inequality: Dose–Response Effects in PISA 2025 Peru

Abstract

The expansion of generative artificial intelligence in education raises the need to determine whether school-based opportunities for critical AI literacy improve learning and whether their effects are equitably distributed across students from different socioeconomic backgrounds. This study aimed to estimate the average, dose–response, and heterogeneous effects of school exposure to activities focused on critically evaluating AI-generated information on Peruvian students’ science achievement. Microdata from PISA 2025 Peru were used, comprising 6,991 students from 258 schools; the main analysis included 4,966 cases with valid treatment and covariate information. Exposure was measured through the reported frequency of school activities for evaluating the quality of AI-generated information, science achievement through ten plausible values, and socioeconomic, demographic, family, digital-resource, and school-context covariates were incorporated. AIPW-ATT, Propensity Score Matching, Entropy Balancing, and a multivalued Generalized Propensity Score were applied, accounting for PISA weights and socioeconomic heterogeneity. The global ATT was −2.75 PISA points and was not statistically significant. However, frequent exposure yielded a 23.77-point difference in estimated effects between the lowest and highest socioeconomic quartiles (95% CI [5.82, 41.73], p = .009). The findings indicate that critical AI literacy opportunities do not produce generalized average gains, but their effects are distributionally heterogeneous and may be relatively more favorable for socioeconomically disadvantaged students.