When a chatbot asks "How are you?": A cross-sectional study of AI call conversations and depressive symptom detection among older adults in rural South Korea.

backgroundRapid population aging in Korea has intensified the need for emotional and health-related support among older adults, especially in rural regions facing depopulation and limited care resources. Conversational artificial intelligence (AI) call systems have emerged as potential tools for monitoring daily well-being and detecting early signs of psychological distress. This study examined the association between conversational patterns in a conversational AI call service and risk of depressive symptoms among community-dwelling older adults, integrating quantitative and qualitative analyses to evaluate the system’s feasibility and limitations.

methodsThis cross-sectional study included 2,896 adults aged 65 years or older residing in Jeongeup City, who used the Naver Clova AI CareCall service and completed a survey between May and August 2024. We linked 37,294 call records with survey data to compare service utilization patterns between high-risk (Geriatric Depression Scale-Short Form score ≥ 2) and low-risk groups. Multivariable logistic regression was used to identify predictors of the risk of depressive symptoms. We reviewed 73 conversation cases containing depressive cues, categorized into ten subthemes based on World Health Organization’s criteria. Three researchers independently assessed AI detection accuracy as correct, misinterpreted, or missed.

resultsAmong all participants, 42.0% were classified as high-risk for depressive symptoms. They were older, more likely to live alone, received social assistance more frequently, and reported poorer health. The high-risk group had fewer positive responses, more negative and unasked responses, and longer call durations. After adjustment, average negative responses (aOR = 1.18, 95% CI = 1.04–1.35), unasked domains (aOR = 1.12, 95% CI = 1.03–1.21), and longer call time (aOR = 1.01, 95% CI = 1.00–1.02) were associated with a higher risk of depressive symptoms. Qualitative findings revealed expressions such as appetite loss, fatigue, and hopelessness, though the AI frequently misinterpreted or failed to detect them due to contextual or speech-recognition errors.

conclusionsThis study provides real-world evidence that conversational and linguistic features in AI call service are associated with the risk of depressive symptoms among older adults in rural communities. AI call service may complement mental health monitoring by enabling early detection of emotional distress. However, limited contextual understanding and variable user acceptance highlight the need for continued technical refinement and integration with professional interpretation.