As large language models (LLMs) increasingly engage in emotionally sensitive interactions, their unintended therapeutic roles demand systematic investigation. We examine ChatGPT’s perceived capacity to provide emotional and mental health support by analyzing user-generated content from social networks through relevance classification and sentiment analysis. Guided by three research questions, we quantify public sentiment toward ChatGPT in therapeutic contexts. To identify posts suggesting therapeutic use of ChatGPT, we introduce two methods: SemReC, a supervised relevance classification, and PASS, an unsupervised similarity-based approach. Both methods demonstrate high accuracy and consistent performance across the dataset. We further assess the performance of existing pre-trained sentiment analysis models to benchmark their effectiveness. To capture affective sentiment propagation in multi-turn interactions, we propose two tree-structured methods—HierSent and AggSent—which model emotional dynamics within threaded conversations. Empirical results validate the effectiveness of our methods and reveal a predominance of positive sentiment toward using ChatGPT for therapeutic purposes. These findings highlight the public popularity of the emergent therapeutic use of LLMs and underscore the need to examine their broader implications for mental health.