内容来源：https://research.google/blog/towards-better-health-conversations-research-insights-on-a-wayfinding-ai-agent-based-on-gemini/

内容总结：

谷歌发布新型“导诊”AI研究：以主动对话提升健康信息获取体验

2025年9月25日，谷歌研究院科学家Mike Schaekermann与研究员Rory Sayres分享了基于Gemini模型研发的新型AI智能体在健康对话领域的突破性进展。该研究旨在通过主动引导、目标理解和个性化对话，帮助用户更高效地获取精准的健康信息。

当前，公众在线查询健康信息时常面临信息过载、内容泛化且缺乏针对性的困境。尽管大语言模型具备优化信息获取的潜力，但多数现有AI工具仅能被动应答，无法像专业医生那样通过主动问询来厘清用户真实需求。

为解决这一痛点，研究团队提出“导诊式AI”概念，并基于三大核心原则设计原型系统：

1. 主动对话引导：AI在每轮对话中提出最多三个针对性问题，逐步减少信息模糊性，帮助用户清晰描述健康状况；
2. 分步最佳应答：在获取完整信息前，AI会根据已有上下文提供阶段性答案，并提示补充细节可优化结果；
3. 透明推理机制：明确向用户解释其回答如何修正并细化此前的结论，增强交互过程的可信度。

为提升体验，团队采用双栏界面设计：左侧呈现对话流程与追问问题，右侧同步更新优化后的答案，确保交互性与信息呈现互不干扰。

在一项针对130名美国非医疗专业用户的随机对比实验中，该AI原型与传统问答式AI（基于Gemini 2.5 Flash模型）相比表现出显著优势。用户普遍认为“导诊式AI”在帮助性、问题相关性、目标理解度及个性化程度等六个维度上更胜一筹。特别是在症状归因类咨询中，用户与导诊AI的对话轮次平均增加50%，表明其更能激发深度交流。

研究结论指出，健康类AI的未来方向应超越被动应答，转向“主动对话伙伴”模式。通过人性化的交互设计，AI可有效赋能用户，使其在健康管理过程中获得更精准、可靠的信息支持。这一成果为AI在医疗健康领域的应用提供了重要实践参考。

中文翻译：

迈向更优质的健康对话：基于Gemini的“路径引导”AI智能体研究洞察
2025年9月25日
Mike Schaekermann（研究科学家）与Rory Sayres（研究员），Google研究院

我们分享一款创新型研究AI智能体的用户洞察，该智能体通过主动对话引导、目标理解和定制化对话，帮助人们更有效地获取健康信息。

获取清晰、相关且个性化的健康信息是患者实现自主管理健康的基石。然而，在线健康信息的世界往往令人困惑、不堪重负且缺乏个性化。我们面对的是海量通用信息，这些信息并未考虑我们独特的背景，因此很难判断哪些细节是相关的。

大型语言模型有潜力使这些信息更易获取且更具针对性。然而，当今许多AI工具仅充当被动的“问答机”——它们对初始查询提供一个单一、全面的答案。但这并非像医生这样的专家帮助他人探讨复杂话题的方式。医疗专业人员不仅仅是提供讲解；他们会提出澄清性问题以了解全貌，发现个人目标，并引导他们穿越信息迷宫。尽管这种寻求背景信息的行为至关重要，但对AI来说却是一个重大的设计挑战。

在《迈向更优质的健康对话：寻求背景信息的益处》中，我们描述了如何基于Gemini设计和测试我们的“路径引导AI”——一个探索新方法的早期研究原型。我们的基本论点是：通过主动提出澄清性问题，AI智能体可以更好地发现用户需求，引导他们清晰表达疑虑，并提供更有用、更量身定制的信息。在一系列四项混合方法的用户体验研究中（共163名参与者），我们考察了人们如何就健康问题与AI互动，并迭代设计出一款智能体。用户认为，与基线AI智能体相比，该智能体在帮助性、相关性和需求贴合度方面显著更优。

形成性用户体验洞察：在线查找健康信息面临的挑战
为了更好地理解人们面临的障碍，我们访谈了33名参与者，了解他们在线查找健康信息的体验。一个关键主题很快浮现：人们常常难以清晰表达自己的健康顾虑。正如一位参与者描述的那样，他们的过程是“……就像把所有词都扔进去，然后看看会返回什么结果。”这可能是因为没有临床背景，难以判断哪些细节在医学上是相关的。

随后，受访者能够使用不同聊天机器人的研究原型。（聊天记录未被保存。）这些参与者构成多样化的群体，并提出了涵盖广泛主题的健康问题（例如，肋骨疼痛、眩晕、持续且无法解释的体重增加、耳鸣与手术；论文中有更多细节）。我们的研究表明，当聊天机器人主动提出澄清性问题时，体验会发生显著变化。大多数参与者更喜欢“延迟回答”的方式——即AI先提问——而不是立即给出全面答案的方式。这种对话方式被认为更具个性化和安抚性。正如一位人士指出，“这感觉更像你和医生交谈时的运作方式……它确实让我更有信心，因为它想在直接给出答案之前了解更多信息。”这些澄清性问题不仅帮助AI提供更好的答案，也赋能用户，引导他们提供更相关的背景信息。我们在先前关于皮肤病学AI的研究中也发现了类似模式。

然而，这种基于澄清性问题的方法的有效性在很大程度上取决于执行效果——如果问题表述不当、不相关，或淹没在长篇文本中容易被忽略，参与度就会下降。

设计路径引导AI，通过个性化主动对话赋能用户
基于这些洞察，我们围绕三个核心原则设计了路径引导AI，以创造更具赋能效果的对话体验：

• 主动对话引导：在每一轮对话中，路径引导AI会提出最多三个有针对性的问题，旨在系统性地减少模糊性。这有助于用户更完整地讲述他们的健康情况，并直接迎合用户对更具情境化答案的渴望。
• 每轮提供尽力而为的答案：因为某些健康相关问题可能无需澄清即可获得良好答案，路径引导AI在每一轮对话中都会基于目前已分享的信息提供一个“尽力而为”的答案，同时强调如果用户能回答一个或多个后续问题，答案可以得到改进。这种方法在对话全程为用户提供有用信息，同时给予选择权，使其能够随着对话进展获得越来越好的答案。
• 透明推理：路径引导AI会解释用户的最新回答如何帮助改进了先前的答案。这使得AI的推理过程清晰易懂。

为确保在“尽力而为”答案部分较长的回答中，澄清性问题不会被遗漏，我们设计了一个双栏布局的界面。对话和澄清性问题显示在左栏，而尽力而为的答案和更详细的解释显示在右栏。这将互动对话与信息内容分离开来。

通过随机用户研究评估我们的路径引导AI
为了评估该智能体在现实世界中的潜在影响，我们进行了一项随机用户研究，通过第三方平台招募了130名美国参与者。所有参与者年龄均在21岁及以上，非医疗保健专业人员，并且有一个愿意与AI互动的健康相关问题。为确保涵盖广泛的健康主题，我们对符合研究条件的主题施加了极少的限制（论文中提供了排除查询的细节）。在随机化的受试者内设计中，每位参与者分别与我们的路径引导AI和基线Gemini 2.5 Flash模型互动，探讨他们的健康主题。在提供知情同意并回答标准人口统计问题后，参与者被指示就其问题进行至少3分钟的对话；然后继续完成调查。在与每个AI互动后，参与者回答了关于他们对体验满意度的六个维度的问题：帮助性、所提问题的相关性、对自身情况的贴合度、目标理解、易用性以及获取有用信息的效率。他们能够提供关于所学内容的开放式反馈，并可以选择上传他们与AI的对话。完成调查无需分享对话。在研究结束时，参与者被提示明确比较两个AI，并就上述六个维度中的每一个表明他们更偏好哪一个。他们还被问到：“对于未来的话题，您会更喜欢第一个还是第二个AI？” AI的呈现顺序（基线AI先于路径引导AI vs. 路径引导AI先于基线AI）在参与者间随机分配。在整个研究过程中，参与者被指示不要提供任何个人身份信息。

通过目标理解和定制化对话提供有帮助且相关的信息
如下所示，研究结果表明，尽管路径引导AI的双栏界面相对陌生，但用户在几个重要维度上更倾向于路径引导AI的方法。用户因路径引导AI的帮助性、相关性、理解其目标的能力以及根据其特定需求定制对话而青睐它。这些发现表明，路径引导AI主动提问的行为成功地为用户创造了更个性化、更有帮助的体验，同时没有给用户体验带来不必要的摩擦。

参与者不仅更偏好与路径引导AI的对话，他们的对话内容也明显不同。与路径引导AI的对话时间更长，尤其是在参与者试图理解其症状原因时。对于这些主题，与路径引导AI的对话平均有4.96个回合，而基线AI为3.29个回合。并且他们提供给每个AI的提示模式在对话中也呈现出差异：

结论
在线查找正确的健康信息可能感觉像在迷宫中穿行。虽然AI有潜力成为强大的向导，但我们的研究表明，其成功关键在于能否超越被动的问答角色，成为主动的对话伙伴。

通过将我们的路径引导AI设计得个性化和主动，我们证明了在结构良好的界面中提出有针对性的问题，能够创造出用户更偏爱的体验，优于更经典的问答式体验，从而使人们能够获得更有帮助、更相关、更量身定制的信息。我们的用户研究结果提供了有力证据，表明这种以人为本的对话方法是AI在健康领域未来的一个有前景的方向，有助于人们更好地规划自己的健康旅程。

致谢
此项研究是Google研究院、Google Health及合作团队共同努力的成果。我们要感谢Yuexing Hao、Abbi Ward、Amy Wang、Beverly Freeman、Serena Zhan、Diego Ardila、Jimmy Li、I-Ching Lee、Anna Iurchenko、Siyi Kou、Kartikeya Badola、Jimmy Hu、Bhawesh Kumar、Keith Johnson、Supriya Vijay、Justin Krogue、Avinatan Hassidim、Yossi Matias、Dale Webster、Sunny Virmani、Yun Liu、Quang Duong、Fereshteh Mahvar、Laura Vardoulakis、Tiffany Guo和Meredith Ringel Morris对本工作的贡献或审阅。我们也要感谢参与这些研究的各位参与者。

英文来源：

Towards better health conversations: Research insights on a “wayfinding” AI agent based on Gemini
September 25, 2025
Mike Schaekermann, Research Scientist, and Rory Sayres, Researcher, Google Research
We share user insights from a novel research AI agent that helps people find their way to better health information through proactive conversational guidance, goal understanding, and tailored conversations.
The ability to find clear, relevant, and personalized health information is a cornerstone of empowerment for medical patients. Yet, navigating the world of online health information is often a confusing, overwhelming, and impersonal experience. We are met with a flood of generic information that does not account for our unique context, and it can be difficult to know what details are relevant.
Large language models (LLMs) have the potential to make this information more accessible and tailored. However, many AI tools today act as passive "question-answerers" — they provide a single, comprehensive answer to an initial query. But this isn't how an expert, like a doctor, helps someone navigate a complex topic. 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. Though this context-seeking is critical, it's a significant design challenge for AI.
In “Towards Better Health Conversations: The Benefits of Context-Seeking”, we describe how we designed and tested our “Wayfinding AI”, an early-stage research prototype, based on Gemini, that explores a new approach. Our fundamental thesis is that 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. 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, and we iteratively designed an agent that users found to be significantly more helpful, relevant, and tailored to their needs than a baseline AI agent.
Formative user experience insights: Challenges in finding health information online
To better understand the hurdles people face, we interviewed 33 participants about their experiences finding health information online. A key theme quickly emerged: 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." It may be that without a clinical background, it’s difficult to know which details are medically relevant.
The people we interviewed were then able to use research prototypes of different chatbots. (The chat histories were not logged.) These participants made up a diverse group and asked health questions on a wide range of topics (e.g., rib pain, vertigo, consistent and unexplained weight gain, tinnitus and surgery; more details in the paper). Our studies revealed that when a chatbot proactively asks clarifying questions, the experience changes dramatically. The majority of participants preferred a "deferred-answer" approach — where the AI asks questions first — over one that gives a comprehensive answer immediately. This conversational style was perceived as more personal and reassuring. As one person noted, "It feels more like the way it would work if you talk to a doctor... it does make me feel a little more confident that it wants to know more before jumping right into an answer." These clarifying questions not only help the AI provide better answers, but also empower users, guiding them to provide more relevant context. We found similar patterns in prior work on AI for dermatology.
However, 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 where they are easily missed.
Designing a Wayfinding AI to empower people through personal and proactive conversations
Informed by these insights, we designed our Wayfinding AI around three core principles to create a more empowering conversational experience:

• Proactive conversational guidance: At each turn, the Wayfinding AI asks up to three targeted questions designed to systematically reduce ambiguity. This helps users articulate their health story more completely and directly incorporates users’ desire for more contextualized answers.
• Best-effort answers at each 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, based on the information shared so far, while emphasizing that the answer can be improved if the user can answer one or more of the follow-up questions. This approach gives the user helpful information throughout the conversation, while providing the option to further receive increasingly better answers as the conversation progresses.
• Transparent reasoning: 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.
  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. The conversation and clarifying questions appear in the left column, while best-effort answers and more detailed explanations appear in the right. This separates the interactive conversation from the informational content.
  Evaluating our Wayfinding AI through a randomized user study
  To evaluate the potential real-world impact of this agent, we conducted a randomized user study with 130 US-based participants recruited via a third party platform. All participants were 21 years and older, were not health care professionals, and had a health-related question for which they were willing to interact with an AI. To ensure a broad range of health topics, we imposed very few restrictions on which topic would be eligible for the study (details on excluded inquiries are provided in the paper). In a randomized within-subjects design, each participant interacted with both our Wayfinding AI and a baseline Gemini 2.5 Flash model to explore their health topic. After providing informed consent and answering standard demographic questions, participants were instructed to have a conversation spending at least 3 minutes on their question; and then to resume the survey. After interacting with each AI, 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 of getting useful information. They were able to provide open feedback about what they learned, and also had the option to upload their conversation with the AI. Sharing the conversation was not required to complete the survey. At the end of the study, participants were prompted to explicitly compare the two AIs and indicate which they would prefer in terms of each of the six dimensions above. They were also asked, "For a future topic, would you prefer the first or the second AI?" The order of AI exposure (Baseline AI first vs. Wayfinding AI first) was randomized across participants. Throughout the study, participants were instructed to not provide any identifying information about themselves.
  Helpful and relevant information through goal understanding and tailored conversations
  As shown below, the results of the study demonstrated that users preferred the Wayfinding AI's approach across several important dimensions, despite its less-familiar two-column interface. Users favored Wayfinding AI for its helpfulness, relevance, ability to understand their goal, and for tailoring the conversation to their specific needs. These findings suggest that the proactive, question-asking behavior of Wayfinding AI successfully created a more personalized and helpful experience for users without introducing undue friction in the user experience.
  Beyond simply preferring their conversations with the Wayfinding AI, participants had noticeably different conversations. Conversations were longer with the Wayfinding AI, in particular when participants were trying to understand the cause of their symptoms. For those topics, conversations with the Wayfinding AI had 4.96 turns on average, compared to 3.29 for the baseline AI. And the pattern of prompts they provided to each AI looked different across conversations:
  Conclusion
  Finding the right health information online can feel like navigating a maze. While AI has the potential to be a powerful guide, our research shows that its success hinges on its ability to move beyond being a passive question-answerer and become an active conversational partner.
  By designing our Wayfinding AI to be personal and proactive, we demonstrated how asking targeted questions in a well-structured interface can power an experience that users prefer over a more classical, question-answering experience, and thus enable people to obtain more helpful, relevant, and tailored information. 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, helping people navigate their health journeys.
  Acknowledgements
  The research described here is joint work across Google Research, Google Health, and partnering teams. We would like to thank Yuexing Hao, Abbi Ward, Amy Wang, Beverly Freeman, Serena Zhan, Diego Ardila, Jimmy Li, I-Ching Lee, Anna Iurchenko, Siyi Kou, Kartikeya Badola, Jimmy Hu, Bhawesh Kumar, Keith Johnson, Supriya Vijay, Justin Krogue, Avinatan Hassidim, Yossi Matias, Dale Webster, Sunny Virmani, Yun Liu, Quang Duong, Fereshteh Mahvar, Laura Vardoulakis, Tiffany Guo, and Meredith Ringel Morris for contributing or reviewing this work. We would also like to thank the participants who contributed to these studies.