The Cognitive Trade-off: A Correlational Analysis of Artificial Intelligence Usage and Adolescent Academic Performance

Abstract

Recent international evaluations show a decline in adolescent literacy and complex cognitive skills. The decline is clearest in capacities needed in the digital era, such as evaluating information, integrating multiple sources, and critical reading (OECD, 2026). This drop has been most pronounced among socio-economically advantaged students. That pattern raises questions about the role of technological infrastructure and access in shaping current educational trajectories. This study draws on the theoretical framework proposed by Guerra et al. (2025) and on related work on cognitive offloading (Gerlich, 2025; Risko & Gilbert, 2016). It investigates the cognitive trade-offs associated with Large Language Models (LLMs) and generative Artificial Intelligence (AI) in education. LLMs offer immediate efficiency and assistance. However, they also work as a cognitive shortcut that reduces direct, effortful engagement with written text and complex problem-solving. Over time, overreliance on these tools may weaken core literacy and analytical capacities. Socio-economically advantaged countries and individuals have better technological infrastructure and greater purchasing power, which could be linked to higher rates of AI adoption. These populations may therefore be more exposed to the cognitive trade-offs. The 2025 PISA cycle offers a unique opportunity to study this question. The first widely accessible LLM chatbot was released to the public in November 2022, shortly after the 2022 PISA assessments were administered between July and September of that year. In a preliminary exploratory analysis, we correlated a country-level measure of labor-market exposure to AI (Murugan et al., 2026) with PISA 2025 outcomes. We consider that measure a distal proxy for the construct of interest, which is the actual use of generative AI tools. The exploratory result motivated the present confirmatory study, which tests the same question with four direct measures of generative AI use. The Knowledge of Data section gives full details. The purpose of this study is to conduct a macro-level, confirmatory correlational analysis of the relationship between country-level generative AI use and adolescent academic performance across PISA participating countries and economies. The study separates static cross-sectional performance in 2025 from performance change between the 2022 and 2025 cycles. Cross-sectional performance is strongly related to national wealth and socio-economic development. Performance change, by contrast, covers the period after generative AI tools became publicly available. This study uses secondary, macro-level data. The predictors are four independent measures of actual country-level generative AI use, each tested separately and not combined into a composite. They are Microsoft AI User Share for the second half of 2025, OpenAI Signals country rank on ChatGPT messages for the fourth quarter of 2025, population-adjusted Claude.ai usage from the Anthropic Economic Index for November 2025, and the Eurostat measure of use of generative AI tools by individuals in 2025, which covers European countries only. The dependent variables are six country-level measures of adolescent academic performance taken from the PISA 2025 Results report (OECD, 2026). Three are the mean performance scores in Mathematics, Reading, and Science in 2025. The other three are the change in performance (Δ) in Mathematics, Reading, and Science between 2022 and 2025. We expect two patterns of results. First, each country-level measure of generative AI use will correlate positively with mean 2025 performance in Mathematics, Reading, and Science (Hypothesis 1, cross-sectional baseline). Higher-income, well-resourced countries historically show higher academic achievement, and the same countries have more technological infrastructure and higher AI use. Second, and this is the main focal hypothesis, each country-level measure of generative AI use will correlate negatively with performance change between 2022 and 2025 in Mathematics, Reading, and Science (Hypothesis 2, temporal trajectory). In line with Guerra et al. (2025), countries with higher generative AI use are expected to show steeper performance declines, or smaller gains, over the 2022–2025 period. This would reflect the cognitive disengagement associated with frequent AI tool use.