Publications

LLM chat apps often warn that “LLMs can make mistakes.” In a user study (N = 217) on LLM-generated answers about reproductive health, infant care, cancer screening, and vaccination, we test in-line disclaimers, pop-ups, priming, and citations, and find that none significantly improve users’ ability to identify health misinformation.

A patient-facing LLM tool augments breast cancer clinical notes with simplified text, medical definitions, FAQs, key information, and to-do items. A study with 200 healthy participants found improved comprehension and self-reported confidence; interviews with seven participants with a history of breast cancer and clinician error analyses highlight patient preferences, concerns, and the need for careful evaluation before deployment.

This mixed-methods study examines how radiologists and non-expert physicians rely on AI advice when interpreting head CT scans, showing that both underreliance on correct advice and overreliance on incorrect advice can limit effective human–AI collaboration.

A controlled study with 49 participants examines how developers translate stakeholder needs into requirements and AI system designs. Requirements that preserve stakeholder roles, contextual goals, human needs and constraints, and relevant values were associated with designs closer to expert judgments; conventional requirements quality, self-reported AI knowledge, and prior AI development experience did not significantly predict design performance.

Tangent-SBM augments Schrödinger bridge learning with mechanistic sensitivity supervision, enabling stochastic transport models to match observed distributions while responding appropriately to interventions in physical conditions and parameters.

Interactive LLM agents talk with pregnant people to provide personalized prenatal-care information, plans, and resources, and to assist clinicians.

A framework for centering people throughout the AI development lifecycle, from problem formulation to deployment, so systems reflect the needs of the communities they serve.

An analysis of how participatory-AI projects recruit community participants, and what recruitment strategies mean for whose voices shape the resulting systems.

A geospatial machine-learning approach to predicting PFAS contamination in aquatic ecosystems, supporting environmental-health monitoring.