Digital Culture & AudioVisual Challenges

SAFARI (Greece, 2024) is a 3D IF (Interactive Fiction) computer game made with the Unity game engine using Foundational AI (Artificial Intelligence) models. This paper presents firstly the methodologies applied to enhance narrative content via interactivity with AI NPC characters (Artificial Intelligence Non-Player Character) and secondly, the conclusions and observations regarding best practices that emerged whilst users interacted with Safari’s AI NPCs during the art exhibition  Do Androids Dream of Electric Sheep? The human condition in the age of AI (Athens, 2024). This paper aims to provide insights into AI NPC design and branching narrative design through LLM (Large Language Model) dialogue systems.

Objective: This study's primary focus is on developing a framework for narrative design with AI tools and pointing out pitfalls designers and writers must avoid to maintain immersion in IF video games. Specifically, it examines AI NPC design, the peculiarities and opportunities interactive narratives offer, and how to utilize LLMs for branching narratives.

Theoretical Framework: The growing research interest in AI spawned a plethora of research papers and books about AI and video games. Despite this, there is a gap in scientific literature between narrative design research and video game artificial intelligence papers. This paper aims to bridge that gap and develop a framework for interactive fiction narrative design with foundational AI models.

Research/Study Design: We conducted a case study of user interaction with NPCs in Safari to design interactive fiction narratives with AI models using qualitative research methods. First, we observed the participants from the first two days of the exhibition, collected the observations in field notes, and analysed them thematically. Next, we conducted a literature review on related papers from journals and conferences, supplemented insights from participant observation with findings from the literature review, and updated Safari according to the suggested design. Using qualitative methods, such as participant observation, allowed us to notice issues in our initial design directly, which quantitative methods (e.g., surveys) could not provide. Finally, we conducted competitive analysis on New VS Old NPCs and compared different AI Character Creation Tools.

Observations: 

To Design Intelligent NPCs:

• Provide a clear backstory, including the character’s goals, role, and personality
• Describe the NPC’s appearance and the scene around them in detail
• Ensure characters have different speaking styles
• Ensure the NPC’s state of mind changes depending on the game state
• Define the NPC’s character arc

To Train Intelligent NPCs:

• NPC training is a complex process in need of overall & case-specific methodologies (Realism -Consistency - Immersion - Adaptability -Emotional impact)
• Decide when to use Short and Long Term Memory
• Training material & data (dialogue samples, transcripts, behavioral examples, emotional responses), knowledge bank
• Fine-tune the Language Model of your choice
• Use Behavior Trees or Story Graphs, Emotional memory, and Voice training

To Create Next Gen Branching Narratives:

• Start by writing the main story and world description into your preferred tool’s knowledge bank
• Create customized branching narrative systems for each character with triggers and sections based on the main story, player choices, and consequences
• Allow users freedom in dialogue while bringing them back when they move away from the story

Conclusion:

In conclusion, this paper bridged the gap between game artificial intelligence research and narrative design by providing a framework for narrative design using foundational artificial intelligence models and explaining how designers should approach branching narrative and NPC creation. Further research in the field can accelerate the implementation of LLM dialogue systems in commercial video games. Findings from Safari can contribute to ameliorating the design practices in AI IF for an optimal user experience.

References:

[1] B. Xia, X. Ye, and A. O. M. Abuassba. 2020. “Recent Research on AI in Games.” 2020 International Wireless Communications and Mobile Computing (IWCMC), Limassol, Cyprus, pp. 505‑510  

[2] C. Phillips. 2019. All Choice No Consequence: Efficiently Branching Narrative. Game Developers Conference  

[3] N. Montfort. 2013. Riddle Machines: The History and Nature of Interactive Fiction. John Wiley & Sons, Ltd  

[4] Q. C. Gao & A. Emami. 2023. The Turing Quest: Can Transformers Make Good NPCs? Annual Meeting of the Association for Computational Linguistics  

[5] T. B. Wright & J. L. Weible. 2024. Current attitudes on digital interactive fiction and text adventure games within learning contexts: A systematic literature review. Journal of Research in Science, Mathematics and Technology Education

[6] UBCx. 2024. Interactive Narrative. EdX