Researchers have found that a two-hour chat with an artificial intelligence (AI) model is enough to accurately replicate someone's personality. A study, released on Nov. 15 and available in the arXiv preprint database, saw researchers from Google and Stanford University develop "simulation agents. " These AI replicas were based on interviews with 1, 052 individuals, using the discussions to instruct a generative AI model intended to mimic human behavior. The study assessed the AI replicas' accuracy by having participants undertake personality tests, social surveys, and logic games twice, with a follow-up two weeks later. The AI replicas completed the same tests, aligning with human responses 85% of the time. The research paper argued that AI models that imitate human behavior could benefit various research areas, such as assessing public health policies, understanding reactions to product launches, or modeling responses to significant societal events that would otherwise be too challenging, expensive, or ethically problematic to examine with human subjects. In a related development, AI speech generators have reportedly reached "human parity, " though scientists consider them too dangerous for public release. "General-purpose simulation of human attitudes and behavior — where each simulated person can engage across a range of social, political, or informational contexts — could enable a laboratory for researchers to test a broad set of interventions and theories, " the researchers stated in their paper. Simulations could aid in piloting new public interventions, developing theories about causal and contextual interactions, and deepening our understanding of how institutions and networks affect people. To build the simulation agents, researchers conducted detailed interviews covering participants' life stories, values, and opinions on societal issues. This approach allowed the AI to grasp nuances that standard surveys or demographic data might miss.

The interview structure also permitted researchers to focus on what was of personal significance to them. The scientists used these interviews to craft personalized AI models capable of forecasting participants' answers to survey questions, social experiments, and behavioral games. This encompassed responses to the General Social Survey, the Big Five Personality Inventory, and economic games like the Dictator Game and the Trust Game. Although the AI agents closely mirrored human behavior in several areas, their accuracy varied by task. They excelled in replicating personality survey responses and determining social attitudes but were less effective in predicting behaviors in interactive games requiring economic decision-making. The researchers noted that AI typically struggles with tasks involving social dynamics and contextual subtleties. They also acknowledged the potential for misuse of the technology. Already, AI and "deepfake" technologies are employed by bad actors to deceive, impersonate, and manipulate others online. Simulation agents could similarly be exploited, the researchers cautioned. Nonetheless, they suggested that the technology offers a way to study human behavior in previously impractical ways by creating a controlled test environment devoid of the ethical, logistical, or interpersonal challenges associated with human involvement.