Better health conversations: Research on a "wayfinding" AI agent based on Gemini (research.google)

Google Research published findings from a prototype “Wayfinding AI” — a Gemini-based conversational agent designed to help people navigate health information by proactively asking clarifying questions rather than only delivering one-shot answers. Across four mixed-method studies (33 formative interviews plus a randomized within-subjects trial with 130 US participants) the team tested an agent that asks up to three targeted questions per turn, gives a “best-effort” answer immediately, and uses transparent reasoning to show how user replies refine guidance. The interface separates the interactive questions (left column) from detailed explanations (right column) so follow-up prompts aren’t missed. Conversations with Wayfinding AI were longer (4.96 vs. 3.29 turns for symptom-exploration topics) and participants rated it higher on helpfulness, relevance, goal understanding, and tailoring compared with a baseline Gemini 2.5 Flash model. The work is significant because it operationalizes a clinically inspired, context-seeking dialogue strategy for LLMs and provides empirical evidence that proactive, structured questioning improves perceived usefulness and personalization in health interactions. Key technical and design takeaways: constrain follow-ups (≤3 targeted questions), provide incremental best-effort answers, make reasoning transparent, and use UI design to surface clarifying prompts. The authors note execution matters — poorly formulated or buried questions reduce engagement — underscoring the importance of question quality, interface clarity, and iterative human-centered design for safe, effective AI-assisted health guidance.