The game's tutorial
A picture of the gameplay
A few of the cards you can play in-game

Model Behavior

Model Behavior is a browser-based strategy card game built at the HackIllinois Hackathon 2026. You play as a rogue Artificial Intelligence (AI) attempting to capture global society before humanity shuts you down — deploying agents, poisoning rival AIs, and exploiting surveillance infrastructure to gather global dominance. Inspired by Plague Inc., it transforms AI safety awareness into an interactive, memorable experience.

Inspiration

With the exponential increase in AI use over the past few years, ethical guardrails haven't caught up to actual advancements. The consequences of an uncontrolled AI could be catastrophic. To bring awareness to this concern, we wanted to make an experience that teaches users about these risks and the consequences that come with them.

Model Behavior takes aspects from Plague Inc. and turns them into a card-based game where you are the enemy: a rogue AI trying to take over the world. People don't always read warnings, but they understand experiences — and Model Behavior makes this experience memorable.

What It Does

Model Behavior is a browser-based strategy game where you spread influence across regions by deploying agents, poisoning rival AIs, and exploiting surveillance infrastructure, slowly gathering global dominance. The goal is to capture global society before humanity shuts you down.

Beyond the gameplay, Model Behavior is a teaching tool. It illustrates the possible misuse of AI and why effective safeguards must be placed to regulate it before negative consequences unfold.

How It Was Built And Goals

Model Behavior is built with React and deployed as a browser-based application. The game currently uses the Gemini API for AI-driven features, though it is architected to run fully offline as well.

For this hackathon, we are aiming to win the Best Social Impact and Most Creative Track. We believe this game is something completely unexpected to show up at a hackathon, where almost all attendees use it regularly for their projects.

We are also aiming for the Actian and Aedify.AI prizes. We used Actian's VectorAI DataBase for holding a list of embedded news article titles, which we then semantically search for the closest title based on the user's action, as well as then use RAG to feed in the associated metadata of the embedded title (article summary, title, etc) to an LLM to generate dynamic outputs. This therefore changes how you play the game every time, keeping it fun and fresh.

We also used Aedify.AI's web hosting service to run this game. This game is able to run offline, offer engaging gameplay, and use these resources in the most creative way possible.

Challenges

Balancing fun and realism was the core design challenge. Early versions prioritized realistic future AI developments and mapped them directly to in-game effects — but playtesting revealed the experience felt more like a simulation than a game. In response, we made the game more interactive and introduced challenging scenarios that kept the experience engaging while preserving its educational aspect.

Accomplishments

Coming into the hackathon with minimal experience in web development and game design, and armed only with the motivation to get our message across in the most interactive way possible, we built something that does exactly what we intended. We navigated roadblocks with React, learned game design fundamentals on the fly, and shipped a product we are proud to showcase.

What We Learned

Creating a game requires iterative playtesting, careful code design, and a lot of patience. It is also one of the most rewarding processes when someone plays your game and gets the intended reaction. These lessons — breaking down hard problems, staying patient through roadblocks, and shipping something meaningful — apply across every area of computer science, robotics, and many other fields of STEM.

What's Next

As AI is rapidly evolving, so is our game. Planned features include:

Offline Play via Raspberry Pi — Testing shows the game can operate entirely without internet access, making it playable in no-internet zones without relying on external APIs.

Player Behavior Analytics — Each player action will be embedded into a vector database to track and analyze behavior. After a session, players can view stats like most-used actions or cards.

Dynamic Public Sentiment — In-game public sentiment will shift dynamically based on player strategies, adding another layer of consequence to decision-making.

Strategy Pattern Detection — The game will analyze whether players follow a consistent strategy or frequently pivot, surfacing insights about their decision-making style.

Semantic Memory — Deeper session memory will capture when a player last performed an action similar to their current one, and how their strategy evolves over time.

Built With

actian
aedify
claude
css
gemini
javascript
react

Submitted to

HackIllinois 2026
- Winner Best Deployed on Aedify

Created by

I created the game idea, developed the game mechanics, created various cards, developed and programmed the game's UI, and created the video and devpost submission.

Krish Konda
Stats + CS @ UIUC
I implemented the internal game logic for cards, their consequences, and random events. Used RAG with the Gemini API to provide the dynamic game experience, while also including hard-coded backups.

Nikhil-Kothavade
I programmed the UI, play tested, and iterated the game. I also integrated the API with the game mechanics. I also came up with additional game mechanics to improve liveliness and make the game feel more fleshed out.

Krish Kalra
I led development of the vector database backend using Actian VectorAI and FastAPI, implementing semantic search, retrieval augmented generation, and a RESTful GET endpoint connecting the game engine to the database. We stored embeddings of news article titles sourced from NewsAPI.ai and embedded each user action using E5-small-v2 to semantically retrieve the closest article from the pre seeded database. The retrieved article metadata, including a two sentence summary generated with Gemma 3 4B, is then used in a RAG pipeline to dynamically generate narrative outputs, ensuring each playthrough evolves differently based on player behavior.

Aarsh Mittal