Generative AI (GenAI) has rapidly permeated educational settings, becoming an integral tool for students engaging in various study-related tasks. With the increasing accessibility and sophistication of AI-driven technologies such as ChatGPT, students now frequently utilize these tools for assignments, problem-solving, and comprehension support. However, while the integration of GenAI into learning environments is accelerating, its broader impact on academic achievement remains insufficiently explored.

A recent meta-analysis by Deng et al. (2025) found that the use of ChatGPT enhances academic performance, positively influences affective-motivational states, and fosters higher-order thinking propensities. Despite these promising findings, the role of GenAI in K-12 education remains understudied, particularly in terms of its mechanisms and differential effects across student populations.
One key gap in the literature is understanding how GenAI use translates into improved learning outcomes. It is unclear whether GenAI primarily enhances academic performance through increased self-regulatory behaviors—such as motivation, perseverance, and self-efficacy—or whether its influence is more direct. Additionally, it remains uncertain how GenAI use interacts with students’ cognitive abilities and grade levels, potentially moderating their academic success.

To address these research gaps, this study investigates:
• What is the direct impact of GenAI use on academic achievement?
• How does GenAI use moderate the relationship between self-regulatory factors and academic achievement?

The study utilizes a nationally representative sample of 2,609 seventh-grade students in the Czech Republic. Academic achievement is assessed through both math and Czech language scores, using item response theory (IRT) measures alongside traditional grading metrics. The independent variables include GenAI usage frequency and students’ belief that proficiency in querying GenAI surpasses the necessity of acquiring knowledge independently. Self-regulatory behaviors are measured across four dimensions: instrumental motivation, effort and perseverance, self-efficacy, and control expectations. A range of demographic characteristics is also considered. To examine the proposed relationships, we employ a Bayesian structural equation model (SEM) using the blavaan package (Merkle & Rosseel, 2018; Merkle et al., 2021) in R. The Bayesian approach provides greater flexibility in quantifying parameter uncertainty through credible intervals rather than relying solely on point estimates and p-values. Unlike frequentist methods, Bayesian models yield posterior distributions, representing the probability distribution of a parameter given the observed data and prior knowledge. These distributions allow for a nuanced interpretation of results by offering probabilistic statements about the estimated effects.

Preliminary results suggest that the frequency of GenAI use does not have a direct impact on either IRT scores or traditional grades. However, a notable finding is that students who hold the belief that querying GenAI is more important than possessing actual knowledge tend to achieve lower scores in both grading metrics and IRT-based assessments. This suggests that over-reliance on GenAI as a substitute for knowledge acquisition may hinder academic performance. Our results underscore the importance of understanding not only the extent of GenAI usage but also the attitudes surrounding its application in educational contexts. These findings contribute to the ongoing discourse on the responsible integration of AI in education, emphasizing the need for balanced pedagogical approaches that foster both digital literacy and foundational knowledge retention.

Deng, R., Jiang, M., Yu, X., Lu, Y., & Liu, S. (2025). Does ChatGPT enhance student learning? A systematic review and meta-analysis of experimental studies. Computers & Education, 227, 105224. https://doi.org/10.1016/j.compedu.2024.105224 Merkle, E. C., & Rosseel, Y. (2018). blavaan: Bayesian Structural Equation Models via Parameter Expansion. Journal of Statistical Software, 85(4), 1–30. https://doi.org/10.18637/jss.v085.i04 Merkle, E. C., Fitzsimmons, E., Uanhoro, J., & Goodrich, B. (2021). Efficient Bayesian Structural Equation Modeling in Stan. Journal of Statistical Software, 100(6), 1–22. https://doi.org/10.18637/jss.v100.i06