Reconceptualizing LLM-Induced Hallucinations as Game Features

Abstract

This study presents a novel and systematic framework to integrating Large Language Models (LLMs) into video game design by reconceptualizing the inherent characteristic phenomenon of "hallucinations"—instances where LLMs generate plausible yet inaccurate or fictitious content—as intrinsic game features. Instead of treating hallucinations as errors, we adapt them to enrich narrative complexity and enhance player experience. We introduce two key design strategies: 1) controlling narrative boundaries to limit the disruptive impact of hallucinations and 2) establishing an irrational worldview, which seamlessly incorporates these stochasticities into the game mechanism. We demonstrate these strategies through case studies of three diverse LLM-driven games across different genres. Our work contributes to the game studies community by offering innovative design paradigms that position LLMs as core interactive mechanisms, while considering their unique generative capabilities and implications for game design and research.