Acting Appropriately: A Methodological Framework for Evaluating Contextual Privacy in Embodied AI

As the development and deployment of embodied AI (EAI) models accelerates, there is a need to evaluate whether EAI models adhere to complex social and environmental norms regarding privacy. State-of-the-art privacy evaluations are skewed toward unimodal outputs (text, image, video) or software agents, leaving a gap regarding embodiment. This paper addresses this gap by presenting a methodological framework for evaluating privacy in embodied systems. We operationalize contextual integrity —a prominent framework for understanding privacy norms—for the embodied modality. To enable the evaluation of an EAI’s ability to align to privacy norms and advance evaluation science more broadly, we outline a systematic evaluation protocol highlighting the unique challenges and design choices inherent in translating theoretical frameworks into quantifiable embodied tasks. EAI models need to strike the right information-sharing balance, as they may either “over-leak” sensitive information for utility or “under-share” in favour of privacy, which could limit their ability to complete tasks. We propose a new metric which enables the quantification and configuration of this balance. Finally, we recommend how the community can further develop this methodology.