A team of computer scientists and sociologists, together with Google DeepMind, trained an artificial intelligence-based program to create a digital simulation of a person after a two-hour interview
The results of this study were published on the arXiv platform.
The basis for the development was the large language model ChatGPT. The program analyzes a person's answers during an interview and generates a digital copy of them, capable of reproducing answers with an accuracy of about 85%. To test the accuracy of the simulation, the researchers compared the responses of a person and their digital counterpart to the same questions.
“This technology is not aimed at replacing people, but at optimizing sociological research, which usually requires large resources. The new system can significantly reduce the cost of conducting and analyzing surveys,” the report says.
To train the model, the scientists recruited a thousand volunteers who were interviewed by the system. The created simulations, called “agents,” differ from the usual AI assistants. They model a personality and answer questions based on the information received.
The researchers suggest that such models can increase the efficiency of AI assistants in the future and help personal robots better understand people.