Self-experimentation, or using tracked data to systematically answer health and wellbeing questions via hypothesis testing, has significant potential to support personal health. However, technological support for self-experimentation has focused on expert-designed self-experiments for specific health conditions, limiting people's ability to design their own rigorous experiments. To address this gap, we developed CASEbot (Conversation Agent for Self-Experimentation), an LLM-powered chatbot using a theory-driven approach to guide users through designing well-structured, personalized, and safe self-experiments. We conducted a within-subjects, mixed-methods study with 42 participants comparing CASEbot to a traditional worksheet-based approach. When formally comparing the experiment rigor and specificity, most participants designed better experiments using CASEbot. They appreciated CASEbot's conversational approach, which prompted them to surface everyday constraints and proactively raised safety concerns, but some found the platform too rigid in its recommendations. We discuss opportunities for future generative AI self-experimentation systems for health to balance structured guidance with user autonomy.
ACM CHI Conference on Human Factors in Computing Systems