Trust in health AI · Informed by my work at Ontario Health, Health811 · 2022 to present

Designing the conditions for trust in AI supported health navigation

An illustrative model for how I design trust into AI supported health navigation: a safety check that runs the whole time, a nurse who is always reachable, and handoffs that carry context.

User Experience Design Lead, Ontario Health · Leads UX for Health811 · Works with program and partner teams

Context and stakes

Health811 is Ontario’s public health navigation service, and I lead UX for it. In health navigation, a confident wrong answer has real costs, and so does a person who cannot reach help at the moment they need it. No amount of model accuracy makes up for either. As AI moves into services like this, the model on this page is how I approach it.

Illustrative model: designing trust into AI health navigation. A five step escalation path: 1, describe the concern in your own words; 2, AI assessment, where follow-up questions narrow down what is happening; 3, a safety check that runs the whole time and stops the assessment on any emergency sign; 4, a recommended level of care, from self-care to 911; 5, a next step. Four handoff points: an emergency sign hands off to 911 and ends the assessment; when the AI is unsure or the person asks, a registered nurse takes over with the answers passed along; recommended care hands off to a service directory; and a Talk to a nurse option is always visible. Example safety message: What you’ve described could be serious. Call 911 now, or go to the nearest emergency department. This tool can’t help in an emergency. — Illustrative model. My own conceptual model for trust in health AI, not an Ontario Health design.

My role

I lead UX for Health811, Ontario’s public health navigation service. I mentor senior and intermediate designers across product areas, and I started a cross-program UX and CX community of practice, where leads work through projects in flight and open questions, and where I have demoed AI-assisted prototyping, from a visual to working multi-page flows in Claude Code.

I partnered with program and partner teams to embed WCAG standards into shared design system patterns, holding consistency of application across teams as the quality bar.

The question is not whether the tool can work. It is what conditions have to be in place for it to be trusted, understandable, and safe.

The model

I evaluate health AI by the experience around the model, not the model alone. The test is not only whether the AI returned the right answer, but what happens when it does not: whether the safety message is in plain language, whether the escalation path is obvious, and whether a person is reachable at the moment it matters.

The safety check runs the whole time. It is not a step at the end. An emergency sign at any point stops the assessment and sends the person to 911, rather than waiting for the AI to finish its questions.

People stay reachable. A “Talk to a nurse” option is always visible, and when the AI is unsure, it hands off rather than guesses. Every handoff carries context: the nurse or service that picks up gets what the person already said, so nobody has to start again.

Accessibility sits in the same frame. A safety message that a screen reader user cannot reach, or that assumes a reading level the person does not have, is not safe, so trust is assessed as a property of the whole experience.

Where this comes from

This model draws on my work leading UX for a public health service: embedding accessibility into shared design system patterns, running research with residents, and holding one quality bar across many teams. It is presented as my own approach, and it is the frame I bring to any team putting AI in front of people at a vulnerable moment.

Team

Program and partner teams, who carry the design standards into their products. Senior and intermediate designers I mentor.