Researchers from St Petersburg State University, working with the Artificial Intelligence Research Institute (AIRI) and the Institute of Psychology of the Russian Academy of Sciences, carried out a large-scale evaluation of digital personality models, AI systems that build a profile of a person from interviews. The study was published in the journal IEEE Access.
A total of 645 people took part; they were interviewed and completed psychometric tests on personality, values and attitudes, covering 38 indicators in all. Five AI models were given the interviews and tried to predict each participant’s answers, which were then compared with the real ones.
Traits most visible in speech, such as religiosity and extraversion, were predicted best, while less obvious motives and latent anxiety proved harder. The models tended to overrate kindness and honesty and underestimate traits such as narcissism, which, according to the authors, shows where these tools need calibration.
The team also found that giving the models more context does not always improve accuracy. Researcher Anastasia Panfilova stresses that such models can support research, personalised learning and psychological support, but that they produce probabilistic predictions whose reliability must be checked case by case.
