The Lens Is Not the Problem: Understanding Potential Privacy Risks and User Concerns in AI Glasses

Abstract

AI glasses are emerging as one of the fastest-growing categories in consumer wearables, embedding always-on cameras, microphones, and cloud-connected multimodal assistants. Yet these same sensors can continuously capture audio, video, and biometric signals from both wearers and nearby people, routing much of that data to cloud servers and human reviewers through a pipeline that can expose intimate scenes, faces, and conversations without consent. In this paper, we analyze the privacy risks of Meta's AI glasses through two complementary lenses: an audit of the company's published documentation and a large-scale study of how users discuss these glasses online. First, we organize Meta's privacy policies and data practices into six data channels that trace how captured signals move from the device through cloud and we use the LINDDUN privacy framework to classify the resulting threats across wearers, bystanders, and external adversaries. We then analyze 4,900 public discussions from Reddit about users' privacy concerns of Meta's AI glasses, and we find that public concern fixates on the visible act of recording while rarely naming that captured data is stored in the cloud, used to train AI models, or reviewed by humans. The privacy story of AI glasses, we argue, is less about the lens than about everything that happens after capture.