This study examines the shifting research on social media and psychological well-being. Prior work is split between a connection discourse (social support, belonging) and an anxiety discourse (social comparison, FoMO, problematic use), but remains largely cross-sectional. Building upon this context, this study employs BERTopic dynamic topic modeling on 7,254 Web of Science articles (2010–2025), identifying 110 topics and revealing three thematic clusters: anxiety, connection, and contextual/methodological themes. The findings indicate that anxiety-related topics are more semantically cohesive, whereas connection-related topics are more dispersed. Notably, the field experienced a turning point around 2016–2017, marking the rise and sustained dominance of anxiety-related research. Taken together, these results provide a longitudinal, computational perspective on the field and demonstrate the value of BERTopic for tracking knowledge evolution.