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

Abstract

In AI-mediated game-based learning, generative AI introduces design tensions. This study explores these tensions in Temporium, a digital educational game where undergraduate students engage with historical consciousness through interaction with a generative narrative agent. The game invites students to explore how individuals, groups, and institutions are historically conditioned and how interpretations of the past shape understanding of the present and perspectives on the future. Unlike scripted systems, the generative AI functions as a semi-autonomous agent that reframes how historical situations are understood in response to player-system interaction. The game was deployed with 51 students, and gameplay logs were treated as the primary data source. The study adopts a design-oriented approach, focusing on how player-system interaction unfolds rather than on learning outcomes. Gameplay logs were analysed through an iterative coding process to identify failure modes and subsequently interpreted as design tensions reflecting underlying structural trade-offs. The analysis identified tensions such as instructional steering versus generative agency, narrative coherence versus exploratory freedom, and information availability versus interpretive participation. These tensions are not isolated system failures, but structural trade-offs that influence interaction and create points of friction between AI-generated outputs and pedagogical intentions. Understanding such tensions is essential before learning outcomes can be meaningfully evaluated. By identifying them, this work provides a basis for understanding the design of AI-mediated game-based learning systems and clarifies the constraints and conditions these systems introduce for supporting interpretation and historical consciousness.