AI-Supported Problem-Based Learning for Enhancing Computational Thinking

Synthesis: This quasi-experimental study (N = 103 first-year undergraduates in Nigeria) tested whether AI-supported Problem-Based Learning (AI-PBL) outperforms conventional instruction in building computational thinking (CT) and academic achievement within computer robotics programming. Grounded in Vygotsky's Social Constructivism, the intervention cast AI tools (ChatGPT and Teachable Machine) as "more capable peers" providing adaptive Scaffolding inside students' Zone of Proximal Development across a 14-week, project-based robotics curriculum. The AI-PBL group significantly outperformed the control group on posttest CT and academic achievement after controlling for pretest scores, while gender did not significantly moderate the gains — evidence that AI-enhanced PBL can be both effective and equitable in under-resourced STEM contexts.

Key Findings

• AI-PBL students significantly outperformed the conventional-PBL control group on posttest computational thinking, controlling for baseline scores — confirming AI-enhanced scaffolding meaningfully strengthens CT acquisition.
• AI-PBL students also achieved significantly higher academic achievement in computer robotics programming, controlling for pretest performance.
• Gender did not significantly moderate the instructional effects: both male and female students benefited roughly equally from the AI-PBL approach, countering persistent concerns about gendered outcomes in STEM.
• The CT instrument captured five dimensions — abstraction/pattern recognition, algorithm design, decomposition, debugging, and critical reasoning/Metacognition — all of which were targeted by AI-scaffolded robotic tasks (e.g., line-following, assistive, recycling, and search-and-rescue robots).
• Grounding AI tools in learners' Zone of Proximal Development (as digital "more capable peers") enabled adaptive Feedback, personalized scaffolding, and iterative Problem Solving that traditional lecture-based teaching could not supply.

Study Design & Method

• Design: Quasi-experimental, pretest–posttest non-equivalent groups (intact classes; coin toss assigned group); 2 × 2 factorial with instructional strategy and age group as factors, gender treated as covariate. Analyzed with ANCOVA (JAMOVI) plus N-Gain descriptive analysis.
• Participants: 103 first-year undergraduates (59 female, 44 male; aged 16–25) across two purposively selected public universities in Southeastern Nigeria, in a compulsory Foundations of Robotics Programming course.
• Groups: Experimental (n = 51) received a 14-week AI-PBL intervention (3 contact hours/week, 42 total hours); control (n = 52) received traditional lecture-based instruction on identical content.
• AI tools: ChatGPT (idea generation, code explanation, iterative debugging, decision-logic formulation, reflective writing) and Teachable Machine (training image/color classifiers), embedded with Arduino robotics kits.
• Authentic robotic tasks: design a line-following robot navigating a dynamic maze; a service robot assisting visually impaired users (HRI, voice recognition, obstacle avoidance); a recycling-sorting robot (supervised ML classification); and a search-and-rescue robot (computer vision + decision trees).
• Instruments: 20-item Computational Thinking in Robotics Scale (CT-RS; α = 0.89) capturing abstraction, algorithm design, decomposition, debugging, and critical reasoning/metacognition; 50-item Computer Robotics Programming Skills Test (CRPST; α = 0.91) across motion control, ML for robotics, human–robot interaction, and computer vision.
• Limitation noted: the raw source file truncates before the full statistical results tables, so exact effect sizes and N-Gain values could not be reproduced here.

What this means for practice

• Instructors. Cast the AI as a "more capable peer" inside the learner's Zone of Proximal Development rather than an answer engine, so that explanations and hints carry the abstraction, decomposition, and debugging the task requires.
• Put the AI to work on specific cognitive moves — idea generation, code explanation, iterative debugging, decision-logic formulation, reflective writing — and leave assembly, testing, and revision of the robot with the students.
• Curriculum designers. Budget a full semester, not a few sessions: the intervention ran 14 weeks at three contact hours a week (42 hours) on authentic tasks such as line-following, assistive, recycling-sorting, and search-and-rescue robots.
• Choose AI tools that work in low-infrastructure settings — a general chatbot plus a browser-based classifier trainer alongside Arduino kits — because the equity case for AI-PBL rests on contexts where expert facilitators and robotics hardware are scarce.
• Do not read the null gender effect as automatic equity: the authors condition it on equal access to AI tools and call for investment in teacher training and context-aware, localized AI-PBL frameworks, with infrastructure and teacher readiness named as the remaining barriers.

Limitations

• Groups were intact classes assigned by a coin toss, so no participants were randomized and the pretest–posttest comparison is a non-equivalent-groups design rather than a trial.
• The sample is 103 first-year undergraduates at two purposively selected public universities in Southeastern Nigeria, all enrolled in one compulsory Foundations of Robotics Programming course, so the finding is bound to a single course and country.
• Both measures are achievement tests given at the end of the 14-week intervention — the 20-item CT-RS, adapted by the authors from prior CT frameworks, and the 50-item CRPST — with no delayed retention or transfer measure.
• The conditions differ by more than AI: the experimental group shared robotics kits and worked in groups while the control group received lectures, so the AI tools are not the only difference between them.

Citation

Ayanwale, M. A., & Omeh, C. B. (2026). AI-supported problem-based learning for enhancing computational thinking. Computers in Human Behavior: Artificial Humans, 7, 100263.