Correct/Incorrect Analysis of Responses When Estimating the Cognitive Function of Spontaneous Language in the Elderly

In recent years, research utilizing conversational data from spontaneous speech to estimate cognitive functions in older adults has increased. While spontaneous speech from older adults has traditionally been directed toward physicians or licensed psychologists, conversations with AI agents are now also being considered. However, the differences between having a human or an AI agent as a conversational partner when communicating with older adults remain insufficiently examined. Furthermore, it is still unclear whether the accuracy of responses in conversations differs when the partner is a human versus an AI. This study investigates how cognitive function in older adults is influenced by responses to questions posed by either human or AI conversational partners. The participants were 14 older adults (3 males and 11 females) residing in Tokyo and utilizing a day service center. Cognitive function was assessed using the Mini-Mental State Examination (MMSE), and participants engaged in semi-structured everyday conversations with both human and AI conversational partners. This study analyzed the correctness of the responses to the questions. The results revealed that differences in conversational partners did not significantly impact the response content. Participants with lower cognitive function frequently exhibited “memory errors” or “lack of understanding” in their responses, while those with higher cognitive function displayed higher accuracy rates. Additionally, older participants demonstrated fewer instances of “socially acceptable responses” or “lack of understanding.” Furthermore, differences in question content had a substantial impact on the participants’ responses; for instance, questions related to ADL (activities of daily living) elicited more “socially acceptable responses,” whereas questions regarding orientation often resulted in “memory errors” or “lack of understanding.” These findings suggest that routine interaction with AI agents can potentially enable the early detection of cognitive decline in older adults, facilitating timely and appropriate interventions. This study aims to provide new insights into the effective utilization of AI agents in conversations with older adults, ultimately contributing to the enhancement of elderly welfare.