Designing Fictional Worlds for Play Through Large Language Models

Abstract This paper explores the application of Large Language Models (LLMs) in addressing the challenges of procedural generation in video game design, specifically tackling the Oatmeal Problem and the Bach Faucet Problem. These issues refer to the generation of endlessly similar, unengaging content and the devaluation of content due to its infinite producibility, respectively. By integrating design thinking, computational-cognitive heuristics, and prompt engineering with modern machine intelligence tools, we propose a novel approach termed Procedural Woodworking (PWw). This methodology aims to generate unique worlds and narratives through LLMs while enhancing player engagement by embedding a player’s character into the game’s narrative in a meaningful way. We utilize OpenAI’s ChatGPT-4 to create custom GPTs for world-building and narrative generation, evaluating their novelty and logical consistency through comparative rating scales. Our results indicate that both Top-Down and Bottom-Up world-building methods can produce logically consistent and structurally coherent outputs. The Procedural Woodworking approach, in particular, shows promise in generating novel narrative content that aligns with pre-made human-designed fictional worlds, potentially mitigating the identified procedural generation problems. This study lays the groundwork for further investigation into how LLMs can enrich game design and player experience by offering a more personalized and value-driven narrative generation process.