An interactive simulation platform where a town council of six AI-powered agents must govern and navigate through complex societal crises.
Model Citizens is an interactive simulation platform inspired by Stanford’s "Smallville" project and the classic Turing Test. This platform explores AI's capability for governance by placing a council of six unique AI agents in charge of a town. Each agent is powered by a large language model and has a distinct role, personality, and set of goals.
The core of the project is to observe how these autonomous AI agents deliberate, negotiate, and respond to dynamic crisis scenarios. Crucially, the platform can be used to evaluate the efficacy of AI models in governance, especially their response to crisis scenarios. As a user, you can either be a passive observer watching the AI council work, or you can become an active participant, stepping into the role of one of the leaders to influence the outcome. Your decisions and the council's deliberations will directly impact the town's status, for better or worse.
The user journey begins with a simple start screen, followed by role and crisis selection, and culminates in the main simulation view.
| Start & Role Selection | Crisis Selection & Simulation |
|---|---|
 |
 |
 |
 |
Main Simulation View:
Model Citizens was built using Google's AI Studio with a lightweight, event-driven architecture. The system revolves around a central Simulation Core that manages the game state and facilitates turn-based interactions between the generative AI agents.
- Five Dynamic Crisis Scenarios: Guide the town through a riot, flood, homelessness crisis, pandemic, or a military plane crash.
- A Council of Six AI Agents: Interact with a diverse council including a Mayor, Opposing Mayor, Police Chief, Fire Chief, Doctor, and Senator, each with unique personalities and hidden goals.
- Observe or Participate: Choose to watch the AI's deliberations from the sidelines or take a seat at the table and play as one of the council members.
- Natural Language Deliberation: The entire decision-making process is conducted through conversational, natural language, allowing you to see how the AIs "think" and reason.
- Track Your Impact: A "Town Status" dashboard provides real-time feedback on how the council's decisions affect public order, infrastructure, and citizen panic.
This project was developed in one week and provided several key insights into rapid prototyping with generative AI:
- Visuals vs. Logic: We initially planned to use AI-generated animations but found that AI Studio was better suited for logic and text generation. We pivoted to a simpler, "vibe-coded" visual style to focus on the core gameplay.
- Injecting "Chaos": Early AI conversations were too bland and agreeable. By increasing the model's temperature, we introduced more randomness, which resulted in more realistic, moody, and engaging interactions between the characters.
- The Power of Specificity: Generic prompts yielded generic results. We learned to be highly specific and detailed in our system prompts to guide the AI agents effectively and achieve the desired behavior and personality for each character.