"Data Annotation as Measurement": The creation and evaluation of modern AI systems depends on annotated data. But what happens when annotated data is wrong? In this talk, I will explore how problems can arise during data annotation processes and suggest opportunities for improving data annotation that draw on measurement theory. — Emma Harvey

"When AI Becomes a Care Actor: Evidence-Based Personalization and Privacy in Voice-Based AI for Older Adults": What happens when an AI system is no longer a tool you consult, but an active participant in your care? Voice-based conversational AI systems are evolving into semi-autonomous care actors maintaining longitudinal user context and delivering personalized support alongside clinicians and care partners. But the very mechanism that makes them effective creates privacy risks that existing frameworks can't address: they were built for systems that collect data, not systems that learn, remember, and act over time. In this talk, I investigate how this tension unfolds in practice through the design and evaluation of a voice-based AI Health Assistant that delivers personalized, evidence-based multidomain lifestyle interventions for older adults, a population with distinct health and privacy needs. Drawing on a three-day remote study with 81 older adults, I show that such a system can function as a legitimate care actor and that older adults showed context-sensitive norms about what AI care actors should know, remember, and share. I argue that personalization cannot be defined by data minimization alone—it must be governed by the contextual norms that shape what AI care actors are permitted to learn, remember, and share. — Thalia Viranda