In modern online environments, forming opinions is a highly complex and dynamic process. Social media discourse is often fragmented and contradictory, making it challenging to understand and summarise public opinion. This study empirically evaluates the ability of multiple large language models (LLMs), such as T5, DeepSeek and GPT-4, to generate reliable opinion summaries from ‘discussion branches’ on social media. The focus is on their capacity to maintain factual accuracy, preserve diverse viewpoints, and retain minority opinions. Key factors affecting summary faithfulness are identified, including context window size, model architecture, and prompting strategies. The trade-offs between coherence and oversimplification of nuanced positions are also examined. To address the dynamic evolution of public opinion, an interpretable ‘opinion tree’ framework is employed in conjunction with dynamic topic mapping and branch identification methods to quantify potential distortion risks in opinion synthesis. Using a subset of Coronavirus (COVID-19) discussions on Twitter as the analysis corpus to demonstrate the capability of LLMs in integrating information, retaining minority opinions and representing dynamic viewpoints.