Reconceptualizing LLM-Induced Hallucinations as Game Features

The paper argues that large language model hallucinations, the plausible but inaccurate or fictitious outputs that are typically treated as technical defects, can instead be designed as intentional features of game experiences. Its central contention is that the "Magic Circle" logic of games, the shared understanding that play occurs within a bounded space with its own rules, can accommodate such outputs. The authors frame this as a conceptual and design-oriented contribution rather than an empirical study, and they explicitly note the absence of quantitative data on player engagement or hallucination rates.

The proposed approach rests on two design strategies. The first is controlling narrative boundaries by limiting verifiable detail, keeping dialogue brief, and maintaining thematic focus so that fabrications remain contained and atmospheric. The second is establishing an irrational or comedic worldview in which unexpected outputs read as characteristic of the game's fiction. These strategies are illustrated through three LLM-driven games: Yandere AI Girlfriend Simulator, 1001 Nights, and Suck Up!. The paper cites the Suck Up! livestream, which drew over four million YouTube views, as evidence of audience appeal, and argues that hallucinations can deepen character, enhance atmosphere, and expand player creativity.

The authors temper their enthusiasm with attention to risk. They recommend balancing emergent uncertainty against coherence safeguards such as Retrieval-Augmented Generation, and they caution that player distrust and misinformation are serious concerns, particularly in educational or other serious games. Transparency about AI-generated content is emphasized, with reference to Steam's disclosure requirements and 1001 Nights' public technical disclosure. The approach is judged best suited to single-player, narrative-driven games, with multiplayer and serious contexts posing greater hazards. The authors conclude that the proposed strategies are promising but require further empirical validation and research into how players actually receive hallucination-driven design.