Resisting Surveillance AI: Mechanisms of Accountability Evasion and Community Engagement

Abstract

Computer vision methods are increasingly deployed as technosolutionism for 'safety' and as security theatre. In this extended abstract, I present an overview of my dissertation. For my dissertation, I investigate such automated surveillance systems, including age and gender estimation. These systems require training datasets, which often come from non-consensual footage of pedestrians in public spaces; the footage is low in quality and many of the annotations are speculative. Vendors then market these methods as 'smart' or 'AI-powered', which attracts investor money. The same methods are deployed for mass surveillance tasks such as re-identification, tracking, and age estimation. In Germany, police run public cameras in train stations and on public squares, and justify these systems as `highly accurate' on shoddy metrics. The UK government now demands picture uploads for age-gating. This dissertation critically examines surveillance models and systems, and what their marketing and deployment mean for society.