God’s Innovation Project - Empowering The Player With Generative AI

Ritvik Nair, Timothy Merino, Julian Togelius · 2025

Read God’s Innovation Project - Empowering The Player With Generative AI by Ritvik Nair, Timothy Merino, Julian Togelius (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-13874.txt.

Introduction

Ritvik Nair, Timothy Merino, and Julian Togelius present a 2025 god game in which players shape the land by typing words. In a god game, the player acts as a deity and guides computer characters from above. Classic games of this kind use worlds and rules fixed by a designer ahead of time.

In their game, players collect words and use them as prompts for a light generative model. The model turns those prompts into playable terrain in real time. A user study watched how players used the system and how they felt about it. Findings speak to engagement, to how well the made terrain worked, and to text driven play as a core game mechanic.

Catalog record

CREATORRitvik Nair, Timothy Merino, Julian Togelius
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE33 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-13874

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

God’s Innovation Project - Empowering The Player With Generative AI Ritvik Nair, Timothy Merino, Julian Togelius Abstract In this paper, we present God’s Innovation Project (GIP) , a god game where players collect words to dynamically terraform the landscape using generative AI. A god game is a genre where players take on the role of a deity, indirectly influencing Non-Player Characters (NPCs) to perform various tasks. These games typically grant players supernatural abilities, such as terrain manipulation or weather control. Traditional god games rely on predefined environments and mechanics, typically created by a human designer. In contrast, GIP allows players to shape the game world procedurally through text-based input. Using a lightweight generative AI model, we create a gamified pipeline which transforms the player’s text prompts into playable game terrain in real time. To evaluate the impact of this AI-driven mechanic, we conduct a user study analyzing how players interacted with and experienced the system. Our findings provide insights into player engagement, the effectiveness of AI-generated terrain, and the role of generative AI as an interactive game mechanic. I Introduction Procedural content generation (PCG) has been used as a core game mechanic since its first usage in video games. Whether it be for map generation in open-ended games like Minecraft, creating maps for a rogue-like, or even generating types of games [ 13 ] . On the other hand, AI in video games started off as being used to control NPC behavior, creating allies or adversaries for the player. Over time, the usage for AI in games has expanded to other aspects such as game playing and level generation. The recent rush of improvements in the field of generative AI has brought about the exploration of combining the two techniques into AIPCG. One such method is done through Neural Cellular Automata (NCA) [ 10 ] . It was demonstrated that the NCAs—neural networks that emulate cellular automata behavior—can be used to iteratively construct game levels. This method allows for the emergence of complex patterns from simple, local interactions. Another method that utilizes generative AI for level generation is Path of Destruction [ 11 ] . This approach views level generation as a repair process. Starting with a dataset of existing levels, the method generates an artificial dataset by introducing various sequences of mutations to these levels. In this paper, we aim to push AI beyond its conventional role by integrating generative AI as a core game mechanic. Our goal was to explore how players interact with and respond to AI-driven terrain generation. To achieve this, we incorporated the Five Dollar Model, a lightweight text-to-image generator, allowing players to input text that is then translated into terrain modifications in real time. As part of this study, we also investigate the sentiment a player has towards the integration of generative AI in video games. Recent sentiment towards AI has been negative, especially when it comes to creative fields such as game development and artists. From controversies involving AI art to strikes done by voice actors regarding the use of AI voices in video games and cartoons, many have voiced their opinions against generative AI. With this in mind, we investigate player reactions to AI-generated terrain as a game mechanic they can interact with. We evaluate whether the system effectively met player expectations and how diverse the resulting landscapes were. To gather insights, we logged player inputs and conducted a user survey, analyzing how players engaged with and perceived the generative AI system in the context of gameplay. II Background Procedural content generation (PCG) has been a fundamental aspect of game design for decades, enabling developers to create expansive and dynamic environments without the need for manual level design. Traditionally, PCG techniques rely on algorithms such as Perlin noise, cellular automata, and rule-based systems to generate content such as landscapes, dungeons, and enemy placements. With the rapid advancement of artificial intelligence (AI), researchers have begun exploring the integration of machine learning models into procedural generation, pushing the boundaries of adaptability and player-driven content creation. The paper Procedural Content Generation via Machine Learning (PCGML) [ 9 ] discusses the potential of AI-driven approaches in PCG. One of the key research challenges outlined in the paper is the use of machine learning models not just for passive content creation but as an active game mechanic—a concept that has remained largely unexplored outside of a handful of games such has 1001 Nights [ 12 ] . This paper aims to build upon that idea by demonstrating how generative AI can function as an interactive mechanic, allowing players to influence the game world dynamically through natural language input. A significant challenge in AI-driven procedural generation is ensuring that the generated content is both coherent and meaningful within the context of gameplay. Unlike traditional PCG, where outputs are constrained by hand-crafted rules, machine learning-based PCG often introduces an element of unpredictability. While this sense of unpredictability can be leveraged to serve the game play loop — giving players a new experience each playthrough, or a sense of exploration and discovery — it also lead to frustration. Generated content may result in unplayable or unpredictable artifacts, ruining the experience. This raises the question: How do players perceive AI-generated content, and do current PCGML algorithms align with their expectations? We explore this as a central theme of our research. One notable example of AI-driven content generation is Sentient Sketchbook [ 7 ] , an AI-assisted level design tool. The tool provides designers with real-time feedback and generates alternative map layouts based on their initial designs. More than just a visualization aid, the system actively evaluates playability and suggests improvements by analyzing structural elements within the game environment. This hybrid approach—where AI serves as a collaborative tool rather than a replacement for human creativity—has inspired our work in GIP . However, unlike Sentient Sketchbook, where AI assists the developer, our work focuses on player-driven generation, where AI assists the player in shaping their own gameplay experience. Generative AI models, particularly those using deep learning architectures, have demonstrated remarkable capabilities in producing creative outputs based on human input. The Five-Dollar Model [ 1 ] is an example of a lightweight model designed for low-dimensional image generation, capable of producing sprites and tile-based maps from textual prompts. The model operates within a 10×10 grid, using a predefined set of 16 tiles to generate visually coherent maps. Unlike high-complexity models such as GANs (Generative Adversarial Networks) or diffusion-based AI models, the Five-Dollar Model prioritizes efficiency and speed, making it well-suited for real-time player interactions within a game. By integrating the Five-Dollar Model into GIP , we aimed to explore its potential as a game mechanic rather than just a content creation tool. In our implementation, players are given a set of collectible words that serve as input for the AI model. The generated terrain is not only a visual effect but directly influences gameplay mechanics, such as movement restrictions, strategic positioning, and resource availability. This interactive AI system allows for emergent gameplay, where different players may develop unique strategies based on the words they collect and the terrain they generate. By leveraging generative AI as a core gameplay mechanic, our work builds upon existing

Read the complete text