Closer and Closer Worlds: Using LLMs to Surface Personal Stories in World-building Conversation Games

Abstract

This demonstration presents Closer Worlds, a 2-player digital game designed to facilitate intimate conversation through the magic of collaborative world-building. Players take turns responding to context-aware introspective questions generated by AI (GPT-4) and collaboratively generate images (DALL-E 2) of a personalized world they can imagine sharing. In this iteration, we contribute a novel method for fostering personal storytelling within a play system by using a large language model (LLM) to create customized questions that are responsive to players’ intentions, memories, and values. We present reflections from early play tests and our design process, which suggest a role for generative AI that empowers emotion-rich human dialogue and encourages intimacy.