Generative Agent Simulations of Human Behavior

Simulations of human behavior can empower applications ranging from immersive environments to social policy simulation. However, traditional simulations have struggled to capture the complexity and contingency of human behavior. In this talk, I demonstrate an alternative approach: constructing an agent architecture that accurately simulates individual behavior in open domains. I make my case through generative agents - computational software agents that simulate human behavior using generative artificial intelligence models. Our generative agent architecture enables individual agents to remember, reflect, and plan, and to come together to populate an interactive sandbox environment, such as one inspired by The Sims featuring a small town of twenty-five individuals. Building on this, I propose a scientific foundation for simulations by creating generative agents of real individuals, grounded in 2,000 hours of voice-to-voice interviews with 1,000 U.S. participants, and validating their attitudes and behaviors against the source individuals. I find that these agents replicate their source individuals’ attitudes and behaviors 85% as well as the individuals replicate their own behavior. Extending my line of argument, I explore how generative agent-based simulations can help us prototype social scientific theories and design more effective policies for social spaces.

Speaker: Joon Sung Park, Stanford University