Context-Seeking Conversational AI for Health Information Navigation

google · Healthcare · 2025 · source Agentic: yes · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Health information navigation

Google Research wanted to move health-information AI beyond passive one-shot answering. Users often cannot articulate medically relevant context, and generic answers can feel overwhelming or impersonal, so the prototype needed to ask clarifying questions while still providing useful information safely.

Constraints:

Health information is sensitive and must be tailored without asking users to share identifying information.
Users may not know which details are medically relevant.
Clarifying questions can reduce ambiguity but hurt engagement if poorly written or hidden in long answers.
The design needed comparison against a baseline Gemini model in user studies.
The system was an early-stage research prototype rather than a clinical decision tool.

Architecture

Google built a Gemini-based Wayfinding AI that asks targeted clarifying questions, provides best-effort answers at each turn, explains how new user context changes the answer, and presents conversation/questions separately from detailed explanations in a two-column interface.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  health-user["Health user"] --> wayfinding-ai["Wayfinding AI"]
  wayfinding-ai --> clarifying-question-loop["Clarifying-question loop"]
  clarifying-question-loop --> health-user
  wayfinding-ai --> best-effort-answer-pane["Best-effort answer pane"]
  wayfinding-ai --> transparent-reasoning-explainer["Transparent reasoning explainer"]
  best-effort-answer-pane --> two-column-interface["Two-column interface"]
  transparent-reasoning-explainer --> two-column-interface
  clarifying-question-loop --> two-column-interface

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Health user`	Asks a health-related question and supplies contextual answers to clarifying questions.	`Wayfinding AI`
`Wayfinding AI`	Gemini-based prototype that guides the health conversation with clarifying questions and best-effort answers.	`Health user`, `Clarifying-question loop`, `Best-effort answer pane`, `Transparent reasoning explainer`
`Clarifying-question loop`	Asks up to three targeted questions per turn to elicit relevant context.	`Wayfinding AI`, `Health user`
`Best-effort answer pane`	Provides useful information at every turn while making clear that answers can improve with more context.	`Wayfinding AI`, `Two-column interface`
`Transparent reasoning explainer`	Explains how newly supplied user context changes or refines the answer.	`Wayfinding AI`, `Two-column interface`
`Two-column interface`	Separates interactive conversation and clarifying questions from detailed informational content.	`Clarifying-question loop`, `Best-effort answer pane`, `Transparent reasoning explainer`

Data flow

A user brings a health question. The Wayfinding AI gives a best-effort answer from currently shared context, asks up to three targeted questions to reduce ambiguity, incorporates the user's answers into the next turn, and explains how the latest context refined the previous answer.

“At each turn, the Wayfinding AI asks up to three targeted questions designed to systematically reduce ambiguity.”article, stated

“The Wayfinding AI explains how the user's latest answers have helped refine the previous answer. This makes the AI's reasoning process clear and understandable.”article, stated

Why this architecture

The design mirrors expert wayfinding: ask context-seeking questions before pretending to know the user's goals, while still giving interim value. The two-column UI addresses the practical problem that questions get missed when embedded in long answer text.

“A health professional doesn't just provide a lecture; they ask clarifying questions to understand the full picture, discover a person's goals, and guide them through the information maze.”article, stated

“To ensure clarifying questions are never missed within the longer answers in the “best-effort” answers section, we designed an interface with a two-column layout.”article, stated

Models used: Gemini · Gemini 2.5 Flash baseline

Scale

The overall research involved four mixed-method studies and 163 participants.
“In a series of four mixed-method user experience studies with a total of 163 participants, we examined how people interact with AI for their health questions”article, stated
The formative research interviewed 33 participants.
“we interviewed 33 participants about their experiences finding health information online.”article, stated
The randomized user study used 130 US-based participants.
“we conducted a randomized user study with 130 US-based participants recruited via a third party platform.”article, stated
For symptom-cause topics, Wayfinding conversations averaged 4.96 turns versus 3.29 for baseline.
“conversations with the Wayfinding AI had 4.96 turns on average, compared to 3.29 for the baseline AI.”article, stated

Alternatives considered (roads not taken)

Passive one-shot question answering

Rejected because: It gives a comprehensive answer before understanding the user's context or goals.

“many AI tools today act as passive "question-answerers" — they provide a single, comprehensive answer to an initial query.”article, stated

Deferred-answer only

Rejected because: Some health questions may not need clarification, so the system gives best-effort answers at every turn.

“Because some health-related questions may not require clarification to get a good answer, the Wayfinding AI provides a "best-effort" answer at every conversational turn”article, stated

Put clarifying questions inside long answers

Rejected because: Questions can be missed when buried in detailed answer text.

“engagement drops if questions are poorly formulated, irrelevant, or buried within long paragraphs of text where they are easily missed.”article, stated

Patterns

`prompt-chaining` (workflow)

The agent repeatedly asks clarifying questions and refines later answers based on new context.

“This approach gives the user helpful information throughout the conversation, while providing the option to further receive increasingly better answers as the conversation progresses.”article, stated

`eval-driven-development` (evaluation)

The prototype was iteratively designed and tested through formative interviews and randomized user studies.

“we iteratively designed an agent that users found to be significantly more helpful, relevant, and tailored to their needs than a baseline AI agent.”article, stated

`structured-output-contract` (output)

The two-column interface separates questions, conversation, answers, and explanations into predictable regions.

“The conversation and clarifying questions appear in the left column, while best-effort answers and more detailed explanations appear in the right.”article, stated

Tradeoffs

Ask clarifying questions before finalizing an answer

Chose context-seeking conversation over immediate comprehensive answer only.

Rationale: The approach better discovers user needs and produces more tailored information.

“by proactively asking clarifying questions, an AI agent can better discover a user's needs, guide them in articulating their concerns, and provide more helpful, tailored information.”article, stated

Provide a best-effort answer at every turn

Chose interim usefulness plus follow-up questions over withholding answers until all clarification is complete.

Rationale: Users still get immediate value while retaining the option to improve the answer through more context.

“the Wayfinding AI provides a "best-effort" answer at every conversational turn, based on the information shared so far”article, stated

Use a two-column interface

Chose separated conversation/questions and answer/explanation panes over one long mixed chat response.

Rationale: Separation prevents clarifying questions from being missed inside long informational text.

“This separates the interactive conversation from the informational content.”article, stated

Gotchas (what bit them)

Users often do not know how to formulate medically useful queries.

Root cause: Without clinical background, they may not know which details matter.

Resolution: Ask targeted clarifying questions to help users articulate their situation.

“people often struggle to articulate their health concerns. As one participant described, their process was to "...just kind of like throw all the words in there and then I'm just gonna see what comes back."”article, stated

Question-asking can hurt engagement if executed poorly.

Root cause: Poorly formulated or hidden questions are easy to ignore.

Resolution: Limit questions to targeted prompts and separate them visually in the interface.

“the effectiveness of this clarifying question–based approach depends heavily on the execution — engagement drops if questions are poorly formulated, irrelevant, or buried within long paragraphs of text”article, stated

Evaluation

Google ran formative interviews and a randomized within-subjects study comparing Wayfinding AI with a baseline Gemini 2.5 Flash model.

Metrics: Helpfulness · Relevance of questions · Tailoring · Goal understanding · Ease of use · Efficiency · Future preference · Conversation turns

“each participant interacted with both our Wayfinding AI and a baseline Gemini 2.5 Flash model to explore their health topic.”article, stated

“participants answered questions about their satisfaction with the experience along 6 dimensions: helpfulness, relevance of questions asked, tailoring to their situation, goal understanding, ease of use, and efficiency”article, stated

Outcomes

Participants preferred Wayfinding AI across helpfulness, relevance, goal understanding, and tailoring.
Wayfinding AI produced longer conversations for symptom-cause topics.
The research supports context-seeking agents for health-information navigation.

The research included 163 participants overall and a randomized study with 130 US-based participants.

“users preferred the Wayfinding AI's approach across several important dimensions, despite its less-familiar two-column interface.”article, stated

“The results from our user studies provide strong evidence that this human-centered, conversational approach is a promising direction for the future of AI in health”article, stated