Asking, Playing, Learning: Investigating Large Language Model-Based Scaffolding in Digital Game-Based Learning for Elementary Artificial Intelligence Education

Digital game-based learning (DGBL) has demonstrated notable effectiveness in general artificial intelligence (AI) literacy education, but it often falls short in addressing personalized learning needs. Traditional game-based scaffolding lacks real-time feedback, thereby limiting the students’ deep cognitive engagement. To address these limitations, this study introduces a large language model (LLM)-based scaffolding designed to create an engaging and interactive AI learning environment. Using a quasi-experimental design with fifth-grade students, we compared an experimental group using LLM-based scaffolding (LLM-DGBL) with a control group using traditional scaffolding (TS-DGBL). The results showed that the experimental group outperformed the control group in both learning achievement and cognitive load reduction. Behavioral analysis showed that students in the experimental group engaged in more diverse and deeper self-regulated learning patterns. Moreover, interaction analysis revealed a key pattern: students with a higher cognitive load sought support from the LLM-based scaffolding more frequently than their peers with a lower cognitive load. This study presents a novel method for elementary AI literacy education. It also provides a valuable reference for applying generative AI in educational practice to support the development of adaptive learning tools tailored for young learners.