Abstract: This study, drawing on Kameyama’s “Understanding Confirmation Action Model” within the framework of Action Theory, systematically evaluates the linguistic action competence of three Large Language Models — DeepSeek, ChatGPT-4, and Doubao — in assisting elderly patients with understanding confirmation during clinical encounters. Through simulated real-life doctor–patient dialogues, the study examines how the models deploy linguistic actions in response to comprehension difficulties arising from perceptual and knowledge-related barriers. The results show that, while the three models are broadly comparable in overall competence, their performance varies across specific situational contexts. A common issue observed across all models is the blurred boundary of linguistic actions, as they frequently go beyond the scope of understanding confirmation and inappropriately intervene in diagnostic processes. In addition, the models’ scores on discourse markers point to their limitations at the metapragmatic level. This study offers an initial exploration of the strengths and weaknesses of large language models in supporting elderly patients’ understanding confirmation, and provides both theoretical and practical implications for their improved application in institutional discourse settings.