Comparing AI-Assisted 2D and 3D Block-Based Games for Computational Thinking Development in Introductory Programming

Computational thinking (CT) is a foundational competency in undergraduate education, yet novice programmers frequently encounter high cognitive demands and declining motivation. Although block-based programming and AI-assisted scaffolding can reduce extraneous load by enabling natural-language interaction, it remains unclear whether embedding these capabilities within immersive three-dimensional environments further enhances CT development through reduced extraneous load and enriched spatial context for schema formation. This study employed a randomized between-subjects design to compare an AI-assisted 3D block-based game (CT Logic 3D) with an AI-assisted 2D counterpart (CT Logic 2D) over a four-week intervention involving fifty undergraduate students ( n = 50, Mage = 23.3, SD = 1.07). Both conditions shared identical AI translation pipelines and instructional sequences, with spatial dimensionality constituting the sole experimental manipulation. Results indicated that participants in the 3D condition achieved significantly greater gains in CT ( d = 0.54), stronger delayed retention ( d = 0.74), and higher near-transfer problem-solving performance ( d = 0.82). Relative to the 2D group, participants in the 3D condition completed tasks 22% faster ( d = 1.08), reported lower perceived difficulty ( d = 0.94), and experienced higher levels of flow ( d = 0.89) and intrinsic motivation ( d = 0.86). Engagement increased throughout the intervention, with the largest efficiency gains observed among absolute novices. Moreover, substantial rates of learner-initiated modification of AI-generated solutions (3D: 76%; 2D: 71%) indicated active rather than passive use of AI scaffolding. Interpreted through Cognitive Load Theory, the findings suggest that combining AI scaffolding with immersive spatial representations produces measurable advantages for CT development beyond those achievable through AI scaffolding alone. Theoretical and practical implications for the design of AI-enhanced CT instruction are discussed.