Abstract

When a chatbot becomes a companion, learning in later life changes shape. This study examines how older adults converse with a large language model (LLM) in lifelong learning sessions across 4 Colombian municipalities with unequal digital infrastructures, asking which domains emerge and whether response depth varies by topic and territory. We analyzed 1,135 interactions from 273 older adults using semantic embeddings and unsupervised clustering (UMAP, HDBSCAN, and K-Means), topic modelling (BERTopic), and statistics (Wilcoxon with Bonferroni correction, chi-square, and multiple correspondence analysis). Results show: (a) six domains, dominated by socio-emotional bonding and everyday enquiries; (b) a substantial memory/local-heritage domain; (c) smaller health, gastronomy, and local-history domains; and (d) response depth patterned by topic and municipality. Overall, conversational AI use in later life is socio-affective, culturally situated, and territorially mediated, informing human-centered agent design and inclusive digital ageing and lifelong learning initiatives.