Could artificial intelligence think exactly like you? Study by Google DeepMind and Stanford University creates “digital twins” after just two-hour conversations with volunteers.
(DW) The science of creating your own digital duplicate seems to have reached a new level. Researchers from artificial intelligence (AI) company Google and Stanford University have managed to replicate the personalities of more than a thousand people with an astonishing 85% accuracy, using just two hours of conversation with each participant.
In the study led by Stanford doctoral student Joon Sung Park, which involved 1,052 people from various demographic backgrounds in the United States, the participants were interviewed using an artificial intelligence system designed to “embody” the attitudes and personality of each individual.
According to the scientific journal MIT Technology Review, each volunteer received up to 100 dollars to take part in the interview, in which a friendly voice guided them through topics ranging from childhood to their opinions on politics.
The key to the study’s success, described in an article published on November 15 in the arXiv database, lies in the methodology. Instead of using simple questionnaires or demographic data, the researchers opted for qualitative interviews that allowed them to capture unique personal nuances.
“We can create an agent of a person – an AI replica – that captures many of their complexities and idiosyncratic natures,” Joon Sung Park told the scientific journal New Scientist.
Tests prove accuracy
To verify the accuracy of the digital replicas, the researchers subjected both the participants and their virtual duplicates to a series of tests two weeks later.
A revealing detail emerged when the humans repeated the same tests: they only got their own original answers right 81% of the time, which shows that people vary their answers over time.
In turn, taking this natural variability into account, the AI agents achieved an effective accuracy of 85% and, even more significantly, outperformed the traditional demographic prediction models used so far by 14 percentage points.
Park told MIT Technology Review that he came up with this interview methodology after his own experience with podcasts. “After a two-hour interview, I thought that now people know a lot about me,” he explained.
Virtual laboratory for social sciences
The main aim of this technology is not to create digital duplicates for fun, but to facilitate research in the social sciences. The researchers propose using these agents to evaluate public policies, study responses to new products or analyze reactions to significant social events that would be too expensive or ethically complex to study with real people.
The study, however, also recognizes some important limitations. The AI agents were less accurate in situations requiring economic decision-making or involving complex social dynamics. In addition, the researchers were clear about the potential risks of these technologies, particularly with regard to their potential misuse to manipulate or impersonate other people online.
In order to protect the participants, the team established some ethical safeguards. Park explained to New Scientist that any participant can remove their data from the study or restrict access to their “digital twin”, and the use of these agents is strictly limited to academic purposes.
Glimpse of the future?
Although we are still a long way from widespread adoption of these technologies, companies like Tavus are already experimenting with digital twins that require less data to replicate personalities.
Tavus CEO Hassaan Raza told MIT Technology Review that the research paves the way for more efficient methods, such as short interviews, to train personalized models.
Ultimately, this research represents a significant advance in understanding human behavior, but it also highlights the need to balance innovation with ethical responsibility. Although AI can replicate fundamental aspects of our personality, the richness and complexity of the human experience remains a major challenge for technology.
rc (DW)