Interactive Storytelling with ChatGPT : An Explorative study

Abstract

Since the release of general-purpose generative AI tools such as OpenAI’s ChatGPT and Google’s Gemini, there has been a large, sometimes exaggerated, amount of publicity on the capabilities of these contemporary language models. Through the lens of interactive storytelling, these capabilities narrow down to a model’s ability to generate narrative content, and its ability to maintain narrative cohesion. The focus of this thesis is on ChatGPT 4o Mini’s storytelling abilities using a framework we designed, called Dungeon Master. To test this AI model’s abilities within this framework, we used a pre-written story outline as the base narrative, and created three prompt models with different levels of allowance for the generative model to modify it: Lenient, Medium, and Strict. We conducted ten tests per prompt model and analysed them based on the number of divergences that occurred in each instance of a story. Our results indicate that this model of ChatGPT manages to generate a coherent and narratively consistent storyline for the most part with some minor caveats. The three prompt models we designed and tested had a minimal impact on the way the narratives evolved.