With advances in generative AI, game designers can delegate some of a game’s procedural logic to an AI model, potentially enabling new dynamic play experiences. But the inability for a designer to consistently control a model’s behavior means the biases of these models can manifest in the game in unexpected ways. In this paper, we analyze how LLMs and game rules intersect to produce gameplay that is shaped by the biases of the language model. We draw on data studies to trace the ways that language models take on specific ideological commitments through their training, including biases along lines of race, gender, language, religion, geography, and more. Using Soraya Murray’s framing of games as playable representations and Stuart Hall’s encoding/decoding model, we analyze how the play dynamics of language model interactions are shaped by these commitments. Through case studies of Infinite Craft and 1001 Nights, we argue that LLM-driven gameplay can produce implicit rhetoric encouraging players to accept the model’s underlying beliefs by rewarding the player for adopting them. We conclude by reflecting on the implications for future generative AI-based games, arguing for greater criticality in their design.

In this paper, we perform a postmortem of Paradise, a prototype game we built to explore the potential of using a social simulation to structure language model-driven gameplay. We describe our experience using the Ensemble social physics engine and extending its architecture with GPT-3. We detail our resultant hybrid simulation system, in which the simulation state informs generated dialogue, and generated dialogue informs the simulation state. We note that authoring behavior for this system was unmanageable, entailing an unstable balancing act between prompt engineering, the Ensemble social model, and the mapping between the two. We conclude by sharing our reflections on what we learned from the project; we hope this postmortem will be valuable to others researching social physics and language models in games.

In creative roleplaying games, game masters take on many roles, including keeping track of the game’s story, remembering past actions, and improvising new content based on player choices and desires. This dynamic storytelling can be quite difficult, particularly in real-time play, and this paper explores how digital tools might be able to assist GMs with elements of this process. In particular, we explore issues that game masters experience in running the GUMSHOE One-2-One system by analyzing posts from the online role playing game community, and discuss the design of a digital tool that helps to address common problems. The resulting tool, Shoelace, helps game masters keep track of story events with a graph-based game world visualization, while also providing creative suggestions using Prolog queries over a database of game information.

Tabletop role-playing games require game masters to mediate between the desires of a group of players and the affordances of a mechanically rich game world. In this paper, we explore the potential for large language models like GPT-3 to assist with the improvisational needs of a game master. We build upon Shoelace, a computational assistant for the GUMSHOE One-2-One role-playing framework; using GPT-3, we extend Shoelace to provide dialogue suggestions for non-player characters as well as to highlight relevant game module information. We emphasize the potential for language models to facilitate new role-playing tools and mechanics and suggest that the user experience problems presented by language models and role-playing are still underexplored and unresolved.