LLM-based chatbots for academic stress counseling: comparing Socratic and Didactic styles in fostering stress reduction and therapeutic alliance

As academic stress continues to rise in higher education, large language model (LLM)-based chatbots are gaining attention as scalable counseling tools. However, little is known about how counseling styles shape the therapeutic alliance between chatbots and students. Grounded in Bordin’s (1979) therapeutic alliance framework, this study examines how two distinct chatbot counseling styles, Socratic and Didactic, affect academic stress reduction and therapeutic alliance development. Using a mixed-methods design, 44 students’ perceived stress and alliance experiences were examined across both conditions. Quantitative outcomes were assessed using two-way repeated-measures ANOVA and paired-samples t-tests, while qualitative data from semi-structured interviews were examined through narrative analysis. Results showed that the Didactic chatbot significantly outperformed the Socratic chatbot in reducing stress (p <.05) and strengthening therapeutic alliance (p <.05), contrasting with prior human counseling literature where Socratic dialogue is often associated with deeper relational engagement. Qualitative findings revealed that the Socratic chatbot was perceived as less helpful, reducing students’ sense of clarity, insight, cooperation, and comfort. Theoretically, this study extends Bordin’s therapeutic alliance framework to LLM-based academic stress counseling. Practically, it provides empirical evidence that directive, solution-focused communication is more effective for chatbot-based counseling, underscoring the importance of aligni chatbot interaction styles with user expectations.