When Embodied AI Infers Too Much

As embodied AI systems enter homes, hospitals, and public spaces, they introduce a new class of privacy risks—not by recording more, but by inferring more. Drawing on modalities such as thermal imaging, radio-frequency sensing, chemical analysis, and pressure mapping, these systems can construct rich behavioural and physiological models of individuals, often without explicit user awareness or consent. While each signal may appear innocuous in isolation, sensor fusion can enable powerful inferences about health, habits, and emotional states. This article examines how current legal frameworks only partially address this shift from observation to inference, and explores how sensor design, regulatory oversight, and user-facing transparency can mitigate the risks. We argue that safeguarding privacy in the age of embodied AI requires recognising inference itself as a site of regulation and design—before silent insights become structural intrusions.