As The Decoder reported today.Stanford University,University of Washingtonand Google DeepMind researchers have jointly developed an AI Agent(AI) AgentAI agents, also known as "AI agents", can realistically simulate human behavior in social experiments.
The study noted that this type of simulation system could serve as a virtual laboratory thatHelp validate theories in economics, sociology, organizational and political science. The research team was able to accomplish this by conducting a study of More than 1,000 representative Americans（These AI intelligences were modeled by conducting interviews covering a wide range of ages, genders, educational backgrounds, and political affiliations.
The system operates asCombining detailed interview transcripts with GPT-4o modeling. When a user asks a question to an intelligent body, the system loads the interview transcript andLet AI mimic respondents' answers. To generate these records, the research team administered each respondent'sTwo-hour-long interviewsand transcribes the conversation into text using OpenAI's Whisper technology.
The research team designed several tests to evaluate AI's ability to predict human behavior, covering the General Social Survey, the Big Five personality assessment, and several behavioral economics experiments.
The results show that AI based on interview data in theAccuracy of up to 85% in predicting social survey (GSS) questions, far beyond AI intelligences that rely solely on demographic information.
The research team tested human performance against AI in five social science experiments, four of which showed thatAI results are highly consistent with humansThe correlation coefficient is as high as 0.98. The correlation coefficient is as high as 0.98, suggesting that AI has an extremely strong ability to simulate human behavior.
Interview-driven AI shows significant advantages over traditional demographic methods, especially in handling predictions of different racial groups and political views more accurately, while also effectively reducing bias between demographic categories.
The research team has opened the dataset containing 1,000 AI models to GitHub with a two-tier access mechanism. Comprehensive data is fully open to researchers, while detailed data on individual participants is subject to special approval.
Attach the address of the dataset:
-
https://github.com/joonspk-research/genagents