AI Interview Chatbot
An AI-enabled qualitative interview chatbot built by modifying Geiecke & Jaravel's open-source code to explore how visual media affects self-regulated learning, interest development, and engagement in financial education.
← Back to ResearchDissertation Archive
The interview platform used in the dissertation study is archived on GitHub with its code, configuration files, and documentation. The hosted version has been retired, so there is no live demo.
Talks: The Droids You're Looking For? Using AI Chatbots for Qualitative Research (With Caveats), IT Professionals Forum, University of Illinois, June 4, 2026 (recording available) · Presented at the ITEL Future Learning Technologies Showcase, University of Illinois, May 11, 2026
About the Chatbot
This chatbot is a modification of open-source code developed by Geiecke & Jaravel for AI-enabled qualitative interviewing. The original code was adapted to focus specifically on how visual media, including data visualizations, visual metaphors, and interactive graphs, influences self-regulated learning, interest development, and engagement in the context of financial education.
In the dissertation study, participants completed an online intervention on compound interest and were then immediately interviewed by the chatbot. The interview covered five areas: their learning experience with visual aids, how visuals affected engagement and interest, how visuals supported comprehension and self-regulated learning, their personal preferences for different visual formats, and practical applications and ideal design characteristics.
The chatbot was configured to ask one question at a time, follow up on participant responses, and stay on topic. During development I tested Anthropic's Claude Sonnet 3.5 against OpenAI's GPT-4o mini over several months. Both models occasionally repeated questions, but Claude followed the interview guidance more consistently. When the Claude models I had been using were deprecated in October 2025, I switched to Claude Sonnet 4 and retested without comparing other models. The IRB-approved study yielded 132 interviews, which I coded manually in MAXQDA rather than with AI.
Affordances & Limitations
Using a chatbot as a qualitative research instrument involves real tradeoffs. These shaped how the tool was designed and how the data were interpreted.
- Immediate capture: The chatbot interviewed participants right after the intervention, capturing in-the-moment reflections that would not have been possible to schedule with a human interviewer at that scale.
- Scale: AI-enabled interviewing allows qualitative data collection across a sample size not feasible with human interviewers, supporting the explanatory phase of a mixed-methods design.
- Limited in-the-moment reactivity: The chatbot was trained to stay on task and ask non-leading questions, reducing the kind of interviewer reactivity that can shape responses in human-conducted interviews.
- Consistency: Every participant received the same structured protocol in the same format.
- Researcher coding bias: The interview protocol was designed by the researcher, meaning the framing and scope of what gets asked reflects researcher decisions. This is a form of bias worth acknowledging even when in-the-moment reactivity is reduced.
- Repeated questions: Despite coded safeguards designed to prevent it, the chatbot occasionally repeated questions across the interview.
- Learning styles myth: In a small number of cases the chatbot reinforced the "learning styles" myth in its responses, an artifact of the underlying model rather than the interview design.
- Tradeoffs with depth: The tool trades some relational sensitivity and flexibility for scale and consistency. Those tradeoffs need to be weighed against the research question.