AI Insight
This study of 12,191 Chinese 15-year-olds found that students who used generative AI chatbots for schoolwork more frequently showed higher rates of academic help-seeking behavior, which was associated with better scientific literacy scores. However, GenAI use also had a direct negative association with scientific literacy, and the purpose of use mattered significantly: using GenAI "to help me learn" was linked to higher scores, while using it to summarize texts or draft assignments was associated with lower performance. The positive effects through help-seeking offset about 40% of the negative direct effect, resulting in a small overall negative association.
Why it matters
The findings suggest that not all GenAI use in education is equal, and the purpose for which students use these tools may be more important than whether they use them at all. This has practical implications for educators and policymakers developing guidelines for AI use in schools, emphasizing the need to teach students to use GenAI as a learning aid rather than as a replacement for cognitive effort.
Understand the Science
IntroductionAdolescents increasingly use generative artificial intelligence (GenAI) chatbots for schoolwork, yet evidence on whether this use supports or undermines learning is mixed. Drawing on research on cognitive offloading and self-regulated learning, we examined whether GenAI use for schoolwork is related to scientific literacy directly and through two components of self-regulated learning, academic help-seeking and metacognitive strategy use, and whether the purpose of use matters.MethodsWe analyzed the Program for International Student Assessment 2025 (PISA 2025) data from 12,191 15-year-old students in 374 schools in Beijing, Shanghai, Jiangsu, and Zhejiang, China. Weighted two-level mediation models compared students within the same school, used all ten plausible values, and evaluated indirect effects with Monte Carlo confidence intervals. Results were checked against a design-based model with school fixed effects and balanced repeated replication weights.ResultsStudents who used GenAI more often than their schoolmates reported more help-seeking (β = 0.17) and more metacognitive strategy use (β = 0.30). Help-seeking was positively associated with scientific literacy, yielding a positive indirect effect (β = 0.013, 95% CI [0.009, 0.017]); the indirect effect through metacognitive strategies was smaller (β = 0.007) and was fully accounted for by science self-efficacy. The direct association between GenAI use and scientific literacy was negative (β = −0.048, about 4 PISA points), and the positive indirect effects offset about 40% of it, leaving a small negative total association (β = −0.028). Purpose mattered: among users, using GenAI “to help me learn” was associated with higher scientific literacy (+5.8 points per response category), whereas using it to summarize texts, research new topics, or draft assignments was associated with lower scores (−6.6 points). The results did not differ by gender or socioeconomic status and were identical under the design-based model.DiscussionGenAI use for schoolwork is associated with scientific literacy through opposing pathways: a positive one through academic help-seeking and a negative direct one, and the direction of the association differs by purpose of use. Because the data are cross-sectional, these associations may partly reflect which students use GenAI rather than the effects of its use.