Could an AI Design a Larp? Processes and Applications

Alba Torrebejano Osorio's study examines whether artificial intelligence can design a live-action role-playing game (larp) largely on its own. Through Project Dark.IA, a Spanish-language larp built mainly with ChatGPT (3.5) and Midjourney (4) and playtested at EntreReVs 2023, the research reaches a clear conclusion: current AI cannot produce a coherent larp unassisted, although its tools prove useful for generating ideas and supporting creative work, both within the fiction and in the non-diegetic material surrounding play. The work situates itself against a sparse body of prior work, including a 2020 article outlining potential AI roles in larps and an anecdotal online example in which ChatGPT produced a generic fantasy concept that its creator judged repetitive. As of January 2023, comprehensive AI-integrated larps had not yet appeared.

The experiment's core premise, a dystopian science-fiction setting pitting "The Insurgents" against "The Computer," along with a synopsis and character concepts, were generated readily by the AI. Difficulties emerged in character differentiation, character-sheet writing, and relationship design, where outputs were repetitive or overly similar. The human team therefore had to intervene heavily, merging and restructuring content. The authors judge that AI works best as a generator of ideas and raw material rather than as an independent designer, while noting the project's accessibility and the AI-driven NPBot extension as notable outcomes.

A further strand of the study addresses prompt engineering. Breaking design tasks into specific, concise prompts helps mitigate ChatGPT-3.5's context limit of roughly 4,096 tokens, beyond which earlier details are forgotten and inaccuracies multiply. The authors structure prompts using ludonarrative frameworks, namely Aarseth's four narrative elements and Wolf's eight storyworld infrastructures. They conclude that AI offers real creative value along with potential gains in scalability and cost, but that its full utility depends on human review, and that models with larger context windows may reduce this dependence in the future.