Researchers from Saint Petersburg State University, the AIRI Institute of Artificial Intelligence, and the RAS Institute of Psychology conducted a study on digital personality models. The team analyzed how well large language models can reconstruct individual psychological traits based on interview transcripts.
The experiment involved 645 participants. Researchers evaluated 38 indicators, including personality traits, values, and social attitudes. The AI models were tasked with predicting participants' responses to psychometric tests, which were then compared against actual data.
The study found that models accurately reproduce traits directly reflected in speech, such as extraversion and religiousness. However, they struggle with hidden motives, anxiety, and justice perceptions. Additionally, models exhibit an idealization bias, consistently overestimating kindness and honesty.
The researchers tested five different neural network architectures. They discovered that increasing the volume of input data does not always improve prediction accuracy. Instead, selecting context relevant to the specific psychological characteristic yields better results.
The findings suggest potential applications in personalized education, psychological support, and organizational communications. Experts emphasize that AI-generated digital twins should be interpreted as probabilistic forecasts rather than definitive statements about human behavior.
