Designing for patients’ AI use in healthcare

Patients are turning more and more to AI for health questions. It offers immediate care at minimal cost, which the current healthcare system can’t match. But the result can be a weaker relationship between patients and doctors, with AI substituting for the doctor instead of supporting the visit.

Our early research confirmed the gap: 59.25% of survey respondents don’t share AI-generated health information with their doctor, and another 29.6% do so only occasionally. Doctors often have no way of knowing where a patient’s information came from.

The landscape: existing options each leave a gap. Doctronic accepts health insurance but has full capabilities in only one US state. Amazon Health AI can access records but requires an Amazon Prime account. Consumer AI tools are familiar and widely used, but weren’t designed for health questions.

The opportunity: none of them is built around the relationship between patient and doctor. Our aim is to support that relationship as AI becomes more common, not replace it. We framed the work around three questions:

• How might we establish trust for patients using AI for health inquiries?
• How might we increase adoption of a health-specific AI over foundational models?
• How might we encourage patients to disclose AI use to medical experts?

Why these methods

A survey found patterns and screened for participants who had real experience. Interviews explained why patients behave as they do, especially around disclosure. The contextual inquiry let us watch what people actually did with AI, which mattered because early interviews suggested stated trust and real behavior can differ. In fall, interactive prototypes let participants react to an experience, not just a description of one.

My role

• Contributed to developing the interview guide and question set for patient and medical expert interviews
• Led or scribed multiple interviews, documenting each session for analysis
• Wrote recruitment materials, including the survey invitation and outreach emails to patients and medical professionals
• Served as the team’s writer and editor, keeping documents consistent as ideas changed
• Built prototypes for Concepts 1 and 3
• Researched existing pre-visit and post-visit tools, and presented the success metrics, risk plan, and next steps at our spring final

Five research insights

Hover over or tap an insight to see the evidence and what it meant for design.

Three concepts

The insights led to three concepts, each aimed at making it easier and more comfortable for patients to tell their doctors about AI use. Hover over or tap a sketch to enlarge it.

Concept 1: plug-in for consumer AI

Meet users where they already are. It detects health conversations, summarizes shared symptoms, and prompts for disclosure.

Concept 2: what changed tracker

AI as a care companion. It monitors symptoms and recovery and nudges patients to share with their doctor.

Concept 3: AI-generated summary

Reduce friction around disclosure. Patients review, edit, and approve a shareable report for their doctor.

Where we are now

Early concept testing was small. Of the first five survey responses, four preferred the post-visit tracker. Follow-up conversations showed that preferences depended on each person’s needs and on specific features, more than on the concept itself.

We chose to focus on the pre-visit concept for four reasons: our Microsoft sponsor and staff pointed to it; both concepts fit different patients’ needs; it answers all three of our how-might-we questions; and it keeps AI from playing doctor, while clearly labeling the report as AI-generated. The turning point was reframing the question from “pre or post” to how a tool prepares a patient for their visit.

How we’ll know it works

• Disclosure rates: how many participants choose to share AI summaries or conversations with their provider
• Communication quality: whether the summaries improve patient-provider communication
• Accuracy and time: whether summaries reflect patient concerns while reducing effort and preparation time

Our main risk is that user testing produces competing feature requests. Our plan is to prioritize feedback that ties back to the core research insights: trust, disclosure, continuity, and convenience.

“I feel that mentioning that I used AI doesn’t help.”

Three things have shaped how I work on this project. Recruiting takes longer than planned: by our first deadline we had only five concept-test responses, so we reached out to people personally and followed up with interviews. Testing an experience is harder than testing an interface, so we made the prototypes look alike and used fixed AI responses to keep people on the experience. And separating pre-visit from post-visit too strictly made our concepts harder to tell apart until we reframed the question around what helps a patient prepare.

Beyond the research itself, the capstone has taught me a great deal about working as a team: brainstorming ideas together, organizing a long project across many people, and reaching out to find the right participants. It has also sparked a genuine interest in healthcare. Seeing how patients and clinicians communicate, and where AI now sits between them, has made me want to keep designing for health contexts.