Designing with Unpredictable Narrators: Tensions in AI-Mediated Game-Based Learning

This paper examines how generative AI reshapes pedagogical control in game-based learning, using Temporium, an AI-mediated educational game in which undergraduate students engage with historical consciousness through a generative narrative agent. Its central thesis is that such systems introduce structural design tensions rather than isolated technical failures. Drawing on analysis of 58 gameplay logs from 51 students at a single university course, the authors argue that generative agents destabilize instructor control in implicit learning environments, and that understanding these interactional dynamics must precede any meaningful evaluation of learning outcomes.

The study identifies five recurring tensions: instructional steering versus generative agency, narrative coherence versus exploratory freedom, information availability versus interpretive participation, adaptivity versus consistency, and engagement versus epistemic depth. Illustrated through interaction excerpts and supported by student questionnaire responses, these tensions manifested as the AI failing to answer questions, interpreting similar inputs inconsistently, and offering limited opportunities for deep discussion. A key finding is that such systems tend to favor continuation and affirmation over conceptual challenge, suggesting a systematic bias toward maintaining engagement at the expense of epistemic rigor.

The scope of the work is deliberately narrow: a single-course, small-sample prototype study with a design-oriented focus on interaction rather than measured learning gains. The authors acknowledge these limitations explicitly. Their conclusion is that examining interaction patterns and their underlying mechanisms can move research beyond descriptive accounts toward explanations of how learning occurs. They recommend testing these observations against existing learning theories to determine how well current frameworks account for the processes observed, so that future outcome-oriented studies can be interpreted in light of the interactional patterns that produced them.