Exploring the effect of computational thinking levels on students' learning performance, cognition, and behavior when

Synthesis: Zhao, Wang, Chen, Kang, and Lan (2026) examined how students' computational thinking (CT) levels shape learning when using an AI coding assistant (AICA) in K-12 Python instruction. In a four-week AICA-assisted coding course, 52 eighth-grade students (26 high-CT, 26 low-CT, split by validated CT pre-test) were compared on learning performance (pre/post-tests), behavior (Lag Sequential Analysis of screen recordings), and cognition (Epistemic Network Analysis of reflection journals). The high-CT group significantly outperformed the low-CT group on post-test performance (72.54 vs. 61.73, p = .031). Behaviorally, high-CT students used the AICA for code understanding, debugging, and explanation; low-CT students used it for immediate answer retrieval and code copying. Cognitively, high-CT students showed strong, well-connected self-regulatory coherence across planning, execution, and self-reflection phases, whereas low-CT students showed a sparse network dominated by Help-Seeking. The authors propose differentiated instruction: open-ended AICA support for high-CT students and structured SRL guidance for low-CT students.

Core Finding

Computational thinking level is a decisive learner characteristic in AI-assisted coding education. Students with high CT transformed AICA interaction into significantly greater learning gains than low-CT students: the high-CT group outperformed the low-CT group on the post-test (p = .031) despite comparable prior coding knowledge, and did so with more stable, self-regulated, and cognitively integrated behavior. The key mechanism is not the tool itself but how each group engages it—high-CT students used the AICA as a cognitive partner for understanding, while low-CT students used it as an answer source, resulting in shallow learning.

Computational Thinking as a Differentiating Learner Characteristic

The study positions CT as an independent variable rather than an outcome, arguing that CT's abstraction, decomposition, algorithmic thinking, debugging, and generalization shape how learners collaborate with AI. Prior CT and prior coding knowledge were controlled: the two groups showed comparable pre-test scores (t(50) = 1.86, p = .069), isolating CT level as the differentiating factor. This makes CT a lever for understanding the wide variance in how students benefit from generative AI coding tools—a contribution that grounds Computational Thinking as a process-level predictor of AI-assisted learning outcomes in CS Education.

Differential Behavior: Understanding vs. Answer Retrieval

Lag Sequential Analysis revealed sharply distinct interaction trajectories. High-CT students showed typical transitions like DP→RC→AAEC→DP—debugging, running code to confirm issues, then asking the AICA to explain code, reflecting logical Problem Solving and autonomous code understanding. Low-CT students showed UT→AANQ→CPC→RC—immediately querying the AICA, copying its code, and running it without deeper reflection or verification. Across the three tasks, high-CT students showed stable, persistent independent coding and used the AICA as a cognitive partner when tasks were challenging; low-CT students showed fluctuating engagement focused on task completion rather than understanding. This distinguishes productive use of AI Scaffolding from reliance, with direct implications for Student Engagement quality.

Cognitive Regulation: Self-Regulated Learning Coherence

Epistemic Network Analysis of reflection journals showed significant group differences in task strategies (χ²=11.226, p=.001), elaboration (χ²=18.281, p<.001), and self-assessment (χ²=7.239, p=.007), all favoring high-CT students. The high-CT cognitive network was dense and balanced across all three SRL phases (planning, execution, self-reflection), evidencing continuous cognitive adjustment and self-monitoring. The low-CT network was sparse, dominated by help-seeking (SH), with weak connections among planning and self-reflection. This empirically models the reciprocal CT–SRL relationship and contributes to Self-Regulated Learning by showing that CT supports connected, reflective AI Regulation in Education in AI-augmented environments.

Differentiated Instruction for AI-Assisted Coding

The authors propose two differentiated strategies. For high-CT students: open-ended AICA access encouraging exploration of multiple code-implementation methods and metacognitive reflection. For low-CT students: structured SRL scripts ("goal setting – strategy selection – monitoring – evaluation") to internalize external prompts into an autonomous regulation cycle, transforming the AICA from an "answer source" into a shared "cognitive partner." This frames the AICA as differentiated Scaffolding whose design must adapt to learner CT to avoid widening equity gaps in AI-assisted learning.

Relevance to the knowledge base

This article directly links learner cognition (CT) to measurable Learning Gains in a generative-AI tool context, and it demonstrates that self-regulatory coherence, not tool availability, drives success. It is a strong empirical anchor for the knowledge base's Computational Thinking and Self-Regulated Learning concepts, and it offers a concrete equity argument: without differentiated scaffolds, AICAs may amplify existing CT disparities, since low-CT students benefit least and rely most. It also connects to Generative AI and Scaffolding by showing that AI coding assistants function well only when matched to learners' regulatory capacity. The K-12 focus extends the knowledge base's AI-in-education evidence beyond higher education.

What this means for practice

• Instructors. Differentiate AI coding assistant access by computational-thinking level: give high-CT students open-ended exploration and low-CT students a structured goal-setting, monitoring, and evaluation script.
• Watch for answer-retrieval sequences (immediate query, copied code, run) as an early signal that an AI coding assistant is functioning as an answer source rather than a cognitive partner.
• Instructional designers. Build process indicators into AI-assisted coding tasks; only the post-test separated the groups (72.54 vs. 61.73, p = .031), so process evidence is needed to catch differences earlier.
• Treat low-CT students' sparse self-regulatory network as a design target, since their journals showed Help-Seeking dominating with weak planning and reflection connections.

Limitations

• 52 eighth-grade students in one four-week Python course, split into two groups of 26 by a computational-thinking pre-test.
• The groups' pre-test difference was not significant but not negligible (t(50) = 1.86, p = .069), so CT level is not perfectly isolated from prior coding knowledge.
• Cognitive characteristics were coded from reflection journals, which the authors note are inherently subjective and may not capture actual cognitive states.
• No phase-by-phase performance measures (task success rates, concurrent exercise accuracy) were analyzed, so the study cannot trace when group differences emerged.

Citation

Zhao, S., Wang, Z., Chen, T., Kang, C., & Lan, Y.-J. (2026). Exploring the effect of computational thinking levels on students' learning performance, cognition, and behavior when using AI coding assistants.