Automated Interviewer or Augmented Survey? Collecting Social Data with Large Language Models

Qualitative methods like interviews produce richer data in comparison with quantitative surveys, but are difficult to scale. Switching from web-based questionnaires to interactive chatbots offers a compromise, improving user engagement and response quality. Uptake remains limited, however, because of differences in users’ expectations versus the capabilities of natural language processing methods. In this study, we evaluate the potential of large language models (LLMs) to support an information elicitation chatbot that narrows this “gulf of expectations” (Luger & Sellen 2016). We conduct a user study in which participants (N = 399) were randomly assigned to interact with a rule-based chatbot versus one of two LLM-augmented chatbots. We observe limited evidence of differences in user engagement or response richness between conditions. However, the addition of LLM-based dynamic probing skills produces significant improvements in both quantitative and qualitative measures of user experience, consistent with a narrowing of the expectations gulf.