Introduction
While generative AI has revolutionized conversational interfaces, its inherent unpredictability, tendency for hallucinations, and privacy risks make it unsuitable for many rigorous academic and clinical research applications. In fields like public health, psychology, and bioethics, researchers require standardized, controlled environments to collect reliable qualitative data. To address this need, we present ISCA (Interview-Style Conversational Agents), an open-source, non-generative framework designed for structured data collection without the need for complex coding or high-compute resources.
System Architecture and Design
ISCA is a Django-based web server framework that prioritizes experimental control over open-domain fluency. Unlike LLM-driven bots, ISCA utilizes a rule-based dialog manager. The conversation is structured around two primary components:
- Main Questions: A set of core, standardized questions delivered to every participant to ensure consistency across the study.
- Reflections: Dynamic follow-up prompts triggered by the participant's specific input.
As shown above, the system processes user input through lexicon matching (e.g., LIWC, WN-Affect) and sentiment analysis (VADER). If a user mentions specific topics like "health" or "family," the system selects a pre-written reflection to encourage deeper elaboration. This approach allows the agent to mimic active listening and empathy while remaining strictly within the bounds defined by the researcher.
User Experience and Administration
The framework provides a comprehensive end-to-end experience for both participants and researchers. Participants move through a structured flow including study information, pre-interview surveys, the core chat interface, and post-interview feedback.
For researchers, ISCA includes a no-code administrative dashboard. This dashboard allows for the configuration of survey questions, the definition of lexicon triggers, and the execution of built-in qualitative analysis tools like Topic Modeling (LDA and BERTopic) on the collected data.
Empirical Evaluation
The effectiveness of ISCA was validated through two distinct case studies:
- COVID-19 Expressive Interviewing: In a study of over 370 participants, ISCA was used to facilitate stress reduction. The agent outperformed a popular mental health app (Woebot) in reducing self-reported high stress, with a 27% decrease (from 91% to 64%) compared to Woebot's 18% decrease. Users rated ISCA as more useful and easier to navigate.
- Neurotechnology Ethics: ISCA was used to survey 39 students on the ethics of brain organoids. The results closely replicated findings from traditional static surveys, demonstrating that the conversational format maintains data integrity while providing a more engaging user experience.
Conclusion
ISCA demonstrates that for sensitive research interventions, a highly controlled, non-generative agent is often safer and more effective than a fluent but unpredictable LLM. By providing a low-compute, no-code solution, ISCA empowers social scientists and clinicians to deploy sophisticated conversational tools that prioritize reliability and participant safety. Future work will explore human-in-the-loop integrations to add flexibility while maintaining the framework's core commitment to experimental control.
Summary updated