Peer-reviewed research on what people understand about AI — and what they do because of it.
Applied work spanning human-computer interaction, AI design, and health technology. The through-line: where AI's promise and AI's risk both run highest, and how that tension should shape what teams actually build.
A Risk Taxonomy and Reflection Tool for Large Language Model Adoption in Public Health
Before a team ships an AI feature into healthcare, the hard question isn't "does it work?" It's "where could it hurt someone, and who would know?" This paper turns that question into something a team can use. Working with public health professionals and people with lived experience across vaccines, opioid use disorder, and intimate partner violence, we mapped the risks of putting LLMs in front of vulnerable users into four dimensions: the individual, care delivery, the information ecosystem, and accountability. The result is a reflection tool teams can sit down with during design, the kind of artifact that turns "we should be careful with AI" into concrete decisions about what to build, gate, or leave out.
My role: Co-author. I moderated focus groups and contributed to analysis and diagrams.
Exposure to Content Written by Large Language Models Can Reduce Stigma Around Opioid Use Disorder in Online Communities
Product teams tend to treat AI-generated replies as a convenience feature. This study says they're closer to a behavioral intervention, and worth designing as one. We ran randomized controlled experiments where people read either an LLM-generated response, a human-written one, or nothing at all in response to someone posting about opioid use disorder, then measured their attitudes toward medications for treatment. Across both a single-exposure test with over 2,000 people and a 14-day repeated-exposure version, the LLM responses left people least stigmatizing. For anyone building support or community products, the implication is concrete: the copy a model generates can shift how users treat each other, and you can deploy and test it on purpose rather than letting it fill space by default.
My role: Co-author. I designed the experiment, ran the entire unmoderated study end to end, and contributed to analysis and findings.
Integrating Menstrual Health Support into Hinge Health's Physical Therapy App
What happens when you put an AI chatbot in front of someone making a decision about their own body? They don't trust it. We learned this the hard way: 13 of 19 users rejected a visible chatbot for health guidance during concept testing. So we pulled the AI out of the conversation and moved it into the background. For Hinge Health's 1M+ users, exercise plans ignored menstrual cycles entirely. Our redesign reads cycle phase and quietly adjusts exercise intensity, focus, and recommendations, no chat window required. The pivot is the whole story: the most useful AI here is the AI you never see.
My role: Lead designer and researcher on a 2-person team.
