AI Can Now Predict What You Will Do With 85% Accuracy

New research has showcased the predictive power of an AI agent that was able to anticipate the behavior of over 1,000 people with 85% accuracy.

A group of researchers from three universities and Google's DeepMind published the results of their research on the generative agent simulation of 1,052 individuals in the U.S. The authors applied large language models to qualitative interviews about the respondents' lives and then measured how well these agents replicated the attitudes and behaviors of the participants in this research.

"The generative agents replicate participants' responses on the General Social Survey 85% as accurately as participants replicate their own answers two weeks later, and perform comparably in predicting personality traits and outcomes in experimental replications,"

the researchers found.

They claim that this paper is a "foundation for new tools that can help investigate individual and collective behavior."

Each participant completed a two-hour voice-to-voice interview with an AI interviewer that used a protocol developed by sociologists as part of the American Voices Project, a nationally representative mixed-methods research project that characterizes everyday life in the U.S. The respondents were asked to tell the stories of their lives—from childhood to education, to family and relationships, and any major life events. Other questions were related to their views on various societal issues, such as "How have you responded to the increased focus on race and/or racism and policing?"

Moreover, the AI interviewer created follow-up questions depending on each participant's responses.

The process of collecting and processing data for the research

The corresponding author of the research, Joon Sung Park from the Computer Science Department at Stanford University, said, "This work opens the door to simulating individuals."

"We believe that accurately modeling the individuals who make up our society ought to be the foundation of simulations," he noted, adding that "this work points to the beginning of an era in which generative agents can represent real people."

However, the scientist stressed that this type of work also brings some concerns, as society would need to "balance the potential benefits while safeguarding individuals' representation and agency."

The authors of the paper have suggested that this balance could be achieved by requiring systems hosting generative agents of real people to at least "support usage audits, provide withdrawal options, and respect individuals' consent and agency."