Gamification Powered by a Large Language Model to Enhance Flipped Classroom Learning in Undergraduate Computer Science

This study investigates whether a gamified quiz powered by a large language model can strengthen the flipped classroom model in an undergraduate software engineering course at the Colorado School of Mines. The authors built "Engin-AIR-ing," a browser-based game in which GPT-3.5 Turbo generated multiple-choice questions from flipped-content video transcripts, giving students an engaging preparatory assessment that also offered an anonymous channel for asking questions. The work addresses a recognized gap: students often skip preparatory material, and instructors struggle to build and update educational games for each module.

The experiment drew on 23 volunteers from three sections totaling 161 sophomore students, with 13 randomly assigned to play the game and 10 serving as controls across three consecutive modules. Quantitative measures included in-class quiz scores, game attempts, answer accuracy, and time per question, analyzed through Shapiro-Wilk and Levene tests and two-sample t-tests. The game group outscored controls on the first two modules, but the third reversed this, and no comparison reached statistical significance (all p-values exceeded 0.05). Participation also fell by about 30 percent by the final module, which the authors attribute to burnout and the novelty effect wearing off.

Motivation indicators were more favorable. Floater counts trended upward, completion times declined for the three students who played all modules, and some students replayed the game to improve leaderboard rankings. Survey respondents described the game as motivating, reinforcing, and enjoyable, though the timer and leaves were sometimes stressful. Only one anonymous question was submitted during the study, so the authors could not confirm the comfort hypothesis, though all survey respondents agreed that anonymity made them more comfortable. The authors acknowledge limitations including the small sample, self-selection, the ungraded nature of the quizzes, and the LLM's tendency to generate duplicate or overly concentrated questions that required manual vetting.