AI model that copies human personality opens questions on deepfakes

Scammers are already using audio and video deepfakes to trick victims’ family, friends and business associates. What if they could also use deepfakes of their personalities?

A team of researchers from Stanford and Google DeepMind have trained AI models to replicate the personality of an individual after just two hours of conversation.

The study recruited over 1,000 participants who completed a two-hour interview, covering a wide range of topics: From personal lives to views on social issues, personality and logic tests, social science experiments and economic games. Their responses were then recorded and used to train generative AI models for each participant. The scientists named them “simulation agents.”

Two weeks later, the interviewees were asked to repeat the same tests. The test results were then compared to answers to the same questions given by the simulation agents. The AI models could replicate participants’ responses with 85 percent accuracy.

“The generative agents replicate participants’ responses on the General Social Survey 85 percent as accurately as participants replicate their own answers two weeks later, and perform comparably in predicting personality traits and outcomes in experimental replications,” the paper notes.

The paper has been published on arXiv and has not yet been peer-reviewed. Although the research does not delve into deepfakes, the scientists working on the AI models see many potential uses for the generative agents.

The simulation agents could help social scientists and other researchers to conduct studies that would otherwise be expensive, impractical, or unethical to do with real human subjects, according to MIT Technology Review.

They may also help make stronger AI agents. While an exact definition is still under debate, the most common definition of AI agents is software tools designed to autonomously complete tasks and achieve objectives. They rely on AI techniques such as natural language processing, machine learning and computer vision.

“If you can have a bunch of small ‘yous’ running around and actually making the decisions that you would have made—that, I think, is ultimately the future,” says Joon Sung Park, the Stanford PhD student who led the research.

Article Topics

AI agents  |  biometrics  |  biometrics research  |  deepfakes  |  ethics  |  generative AI  |  personality traits