Generative AI empowered contextual educational games as facilitators of elementary school students' computational thinking tendencies

Synthesis: Zhou, Fu, Fang and Hwang put a generative AI non-player character inside a role-playing game built with RPG Maker MZ and ran it for three weekly sessions in two intact fifth-grade classes in eastern China. After adjusting for pretest scores, the class that could question the AI teacher reported slightly stronger computational thinking tendencies than the class that played the same game with scripted dialogue, while a test of algorithmic knowledge found no reliable difference. The question logs suggest the AI worked less as a tutor than as a low-stakes interlocutor students asked to restate a problem, break it into parts, or explain a flowchart symbol. The authors are circumspect about the size of that claim.

Key Findings

1. The AI group gained modestly on self-reported CT tendencies. Adjusted posttest means favored the GenAI class, 4.325 against 4.037, a small difference.
2. Algorithmic knowledge did not separate the groups. Posttest means were 80.7 and 74.9; the gap was not statistically significant, though both classes improved from about 60.
3. Every student in the experimental group questioned the in-game teacher. Two thirds said the interaction encouraged inquiry, and several said they asked what they feared asking the real teacher.
4. Solution seeking was the largest question category. Of the logged questions, 278 sought a solution or strategy, 252 described the problem, 226 broke it into parts, 75 asked what a symbol meant, and 53 explored alternatives.
5. Both conditions were enjoyable; only the AI group prized the questioning. Both groups described strong emotional engagement, and about half emphasized motivational benefits, while 61.9% of the experimental group singled out the chance to ask.
6. The AI worked as scaffolding, not instruction. The authors read the pattern as support for inquiry, not a mechanism that taught algorithms in three weeks.

The game the class played

Five modules made up the environment: course content, story and tasks, GenAI, rewards, and records. Three tasks carried the instruction: Sending a Lamb Home, Finding Lost Treasures, and Planting and Harvesting, each requiring students to identify constraints and plan a strategy. Kolb's experiential cycle framed each round of action, feedback and reflection. The AI appeared as a teacher inside the fiction, and students could type questions at any point.

The comparison class used the same game without the GenAI module, relying on pre-programmed dialogues; the instructor and schedule were identical. The experimental group also got guidance on evaluating AI feedback critically, so the design did not hand them an oracle.

What the two classes showed

The sample was 78 students aged 10–11 in the fall 2024 semester, during a compulsory information technology course on algorithm representation: 42 in the experimental class, 36 in the control class. Pretest scores were statistically indistinguishable on both measures.

For CT tendencies, the analysis of covariance used heteroskedasticity-consistent standard errors because a Levene test flagged unequal error variances. The adjusted difference was 0.288 (95% CI [0.006, 0.570]), t(75) = 2.031, p = 0.046, partial η2 = 0.052: statistically significant and practically small, with a confidence interval whose lower bound sits close to zero. The authors call it preliminary evidence of a modest difference, not a medium or large effect, which is the honest reading.

What students asked the in-game teacher

The system logged every question, though not the AI's replies or what students did next. After off-topic entries were removed, the questions fell into five categories: problem description, decomposition, solution seeking, alternatives, and requests to clarify a concept or symbol. One student wrote, "With 135 gold coins, I have to buy seeds, but the price of each seed is different; which way to buy seeds is the most cost-effective?" Another asked the teacher to "rank the preferred planting options." A third asked, "What does the box in a flowchart mean?"

Those are recognizable CT moves: turning a narrative into variables and constraints, decomposing an optimization problem, clarifying a representation. They are evidence of help-seeking, not of solved problems: without the responses or later actions, the log shows what support students wanted, not what they learned.

Why the knowledge test may not have moved

Three weeks is short for explicit procedural knowledge, and the open-ended items used familiar contexts, such as buying a ticket, using a search engine and guiding a puppy out of a maze, which may not have separated the groups. The measures also capture different things: a six-item self-report scale records whether students approach problems systematically, while the test records whether they can describe and apply algorithms.

The authors lean on that point: dispositions shift within weeks, while understanding needs explicit instruction and practice. That is plausible but untested here: nothing separates a real dispositional change from a more positive self-perception after three weeks of play.

What this means for practice

• Instructors: treat the in-game AI as a place to ask, not a place to get answers. Students valued formulating a question, hearing it restated and trying again, so protect that cycle instead of letting the AI supply the solution.
• Instructional designers: log the questions students type. They are the cheapest process data a game like this produces, and they show which part of a task students cannot name well enough to ask about.
• Administrators: judge a GenAI game on whether it changes how students work, not on whether they enjoyed it. Both conditions here were rated fun, and the measurable difference was small.

Limitations

• Two intact classes in one school, assigned at the class level, so pre-existing class differences may not have been fully controlled.
• Three instructional weeks with no delayed posttest, so persistence and transfer are unknown.
• CT tendencies came from a six-item self-report scale and algorithmic knowledge from a locally developed test; neither is performance-based.
• The system kept student questions but not the AI's responses or later actions, and the control group produced no comparable process data.
• GenAI was bundled with narrative, rewards, experiential guidance and teacher support, so the results describe the whole environment rather than the AI alone.

Connected Concepts

• Computational Thinking: tendencies treated as a disposition, distinct from algorithmic knowledge
• Game-Based Learning: contextual role-play as the delivery format
• Help-Seeking: the in-game teacher lowered the cost of asking
• Scaffolding: motivational and cognitive support, not direct instruction
• Experiential Learning: Kolb's cycle structured the question, reflect, retry loop

Connected Articles

• AI-supported problem-based learning for enhancing computational thinking: AI-supported problem-based learning for CT in a quasi-experiment
• RoboBlockly Studio: Conversational block programming with embodied robot feedback for computational thinking: a conversational agent with block programming and robot feedback
• Effectiveness of AI-Supported Game-Based Learning: A Systematic Review of Outcomes, Challenges, and Future Directions: a review of AI-supported game-based learning
• Beyond Access: Engagement with Conversational AI-Assisted Feedback Shapes Cognitive Load in Digital Game-Based Learning: access to conversational AI in a game versus engagement

Citation

Zhou, W.-L., Fu, Q.-K., Fang, J.-W., & Hwang, G.-J. (2026). Generative AI empowered contextual educational games as facilitators of elementary school students' computational thinking tendencies. British Journal of Educational Technology, 00, 1–20. https://doi.org/10.1111/bjet.70101