Killer on Board: Addressing the Narrative Paradox by Utilizing LLM-Driven NPCs

With the growing influence of Large Language Models (LLMs) across diverse technological fields, LLMs show promise for enhancing emergent narrative systems and Non-Playable Character (NPC) behavior. However, their use in this context seems to have received limited academic attention, indicating a gap that this study seeks to address. This study investigated how LLMs could be employed to generate text-based dialog for NPCs to address the narrative paradox while preserving narrative coherence and user agency within the Interactive Digital Storytelling (IDS) prototype, Killer on Board. A mixed-methods design was selected, utilizing the Gamer User Experience Satisfaction Scale (GUESS) (\(n = 44\)) and semi-structured interviews (\(n = 10\)) to explore participants’ perception of the LLM-driven NPCs in the IDS regarding narrative coherence and user agency. To assess the reliability of the datasets, Cronbach’s alpha coefficient (\(\alpha \)) and weighted Cohen’s kappa (\(\kappa \)) were calculated accordingly. Furthermore, a hybrid recruitment strategy was applied, with tests conducted in-person and remotely, to increase participant count. The qualitative findings suggest that LLM-driven NPCs can be utilized to aid the development of emergent narratives in video game contexts but cannot claim the narrative paradox to be fully addressed. Moreover, the GUESS scores suggested that the participants had an overall positive player experience, with all subscales rated above the midpoint (4) and strong internal reliability (\(\alpha > 0.8 \)). However, the Narrative (Score: 4.32) and Creative Freedom (Score: 4.82) subscales scored near the midpoint, suggesting room for improvement.