A research team led by Stanford PhD student Joon Sung Park developed AI-based simulation agents by recreating replicas of 1, 000 diverse individuals. Participants completed personality tests, social surveys, and logic games twice, with their corresponding AI agents subsequently undertaking the same tasks. The results showed an 85% similarity between human participants and their AI replicas. Joon Sung Park believes these AI agents could represent the future as they may eventually replicate human decision-making accurately. These simulation agents provide a cost-effective, practical, and ethical method for conducting social science research. By using AI models to behave like real people, researchers can test interventions on social media, traffic behaviors, and more. Current tool-based AI agents—built to perform tasks like data entry or setting appointments—differ from simulation agents, yet both have the potential to enhance AI development. John Horton from MIT Sloan suggests that combining real human data to create programmable personas can significantly advance AI research. However, the study highlights potential risks, such as creating unauthorized deepfakes and ensuring accurate representation of human behaviors.

The team used basic evaluation methods, such as the General Social Survey and tests assessing the Big Five personality traits, but these don't capture all human uniqueness. They found that the AI struggled more with behavioral tests like the "dictator game. " To overcome these challenges, the researchers conducted qualitative interviews, which they found effective for capturing unique human experiences and details. Joon argues that interviews reveal personal idiosyncrasies and insights that surveys cannot. Having experienced this firsthand through podcast interviews, he regards this as a powerful technique to learn about individuals. Hassaan Raza, CEO of Tavus, notes that producing a digital twin traditionally requires substantial data input, but the new approach suggests collecting less detailed yet vital human information via interviews. His company is exploring this efficient method to create digital replicas by suggesting short, consecutive AI interviews as a means of developing a comprehensive digital twin.