As generative AI becomes more prevalent in video game development, the use of LLM-driven NPCs is expected to grow. This paper examines how human players interact with GPT-4-driven NPCs in a Minecraft minigame designed for this purpose. Through a user study involving 28 players, we identify key patterns in collaborative behavior and discuss the limitations of language-only models without game-state or visual understanding. Our findings aim to guide future game developers in leveraging these advancing generative AI models for improved collaboration in games.
This paper discusses a study on emergent behaviors in game narratives through player interaction with large language models (LLMs). The game is a text-adventure where players solve a mystery, interacting with characters generated by GPT-4. The study involved 28 gamers whose gameplay was converted into a node-graph narrative. It was observed that players' interactions with the LLM led to the discovery of new, engaging narrative elements not originally planned. Players who enjoyed exploration and experimentation in games tended to discover more of these emergent narrative nodes.
This paper introduces GENEVA, a prototype tool for generating and visualizing branching narratives. It leverages the GPT-4 language model to generate detailed narrative graphs with branching and reconverging storylines based on high-level descriptions. GENEVA’s capabilities are demonstrated through its application to four well-known stories, showcasing its potential for aiding in game development and other game-like applications.
This paper shows how to use OpenAI Codex, a model that can generate code and natural language, to create NPCs that can talk and act in Minecraft. The authors use a few examples of how to ask Codex to make code, chat, and answer questions. They test their approach with real gamers and find some limitations and possible solutions.