We live in an era where AI can replicate “everything”: from your voice, appearance, and expression style to your thoughts or decision-making patterns. Now, AI can even replicate your “personality” with remarkable accuracy.
Recently, researchers from Stanford University and Google DeepMind published a paper demonstrating how they created virtual replicas “with individual attitudes and behaviors” using AI.
The researchers call these “universal computational agents.” Most surprisingly, training these “agents” required only a two-hour interview.
▲Source: arXiv
The research was led by Stanford University doctoral student Joon Sung Park. Working with a recruitment company, researchers used stratified sampling to recruit 1,052 participants, each completing a two-hour voice interview in English.
Researchers used an interview protocol developed by sociologists, covering personal life stories and views on contemporary social issues. To quantify interview content and control experimental variables, researchers limited each subject’s responses to approximately 6,491 words.
Notably, the interviewers were also AI. Within the general framework and time constraints of the conversation, these “AI interviewers” dynamically generated follow-up questions based on each participant’s responses.
These responses were then used to train generative AI models for each individual, called “simulation agents.”
To validate these agents’ ability to mimic human personality, both participants and AI agents completed a series of tests, including personality assessments, comprehensive social surveys, five social science experiments, and five famous behavioral economics games.
The five social science experiments, published in the Proceedings of the National Academy of Sciences, are a common standard for evaluating whether generative agents can simulate human participants’ behavioral responses. The five behavioral economics games included the dictator game and prisoner’s dilemma.
Two weeks later, participants were asked to “retell” their test answers, which were then compared with the agents’ generated responses. Results showed that AI agents could simulate participants’ responses with 85% accuracy, while excelling at predicting personality traits and outcomes in the five social science experiments.
▲Source: Hong Kong Youth Association
While human “personality” might seem unquantifiable, this research shows that artificial intelligence excels at “seeing the big picture from small details” – it can accurately clone a person’s “personality” from relatively limited information.
These “agents” have clear advantages and disadvantages. The benefit is that they provide social science researchers with greater flexibility.
By creating digital copies that behave similarly to real individuals, scientists can conduct experiments or surveys directly using agents without recruiting large numbers of volunteers each time. Additionally, these agents can be used for “controversial” experiments involving moral and ethical issues.
John Horton, Associate Professor of Information Technology at MIT Sloan School of Management, noted in an interview that this research “uses real humans to generate characters, then uses these characters programmatically/simulatively in ways that real humans cannot.”
▲Source: TechRadar
The downside is that such agents will inevitably be used by malicious actors for “identity fraud” and ultimately “scams.”
More concerning is that these “agents” have, in some sense, acquired human thoughts and “consciousness,” even being able to express opinions on behalf of the original person. Dr. Park predicts that one day there will be small versions of “you” running around, actually making decisions you would make.
In fact, “AI agents” isn’t a cutting-edge technical term – similar concepts like “digital twins” existed before. However, what demands attention and vigilance is that “agents” need so little information to accurately recreate a person’s “personality.”