Machine-Learning/Artificial-Intelligence—facilitated satellite imagery analysis for trustworthy arms control verification

Abstract With the unraveling of the arms control architecture and diminished prospects for nuclear weapons agreements between the U.S. and Russia and China, there are increased risks from worst-case assumptions, and opportunities to build mutual trust are eroding. In this context, satellite remote sensing is expected to become a vital tool for assessing nuclear forces and postures. Today’s satellite systems generate vast amounts of data that is accessible to both governments and civil society. Given the scale of data, machine learning (ML)/artificial intelligence (AI) play an increasingly central role in data fusion and analysis. However, for ML/AI-facilitated satellite imagery analysis to become a component of future arms control verification regimes that foster the requisite trust to put states on a course towards nuclear disarmament, it is necessary to conceptualize a role for such information collection and analysis mechanisms in an arms control process that is built on and promotes trust between its parties. Yet, challenges arise from the dispersed and heterogeneous nature of data sources and the diversity of actors involved, which may affect perceptions of trustworthiness. The article examines these issues through a practice-based account and trust analysis of an ML pipeline implementation of a missile silo classification task and draws lessons from the Comprehensive Nuclear-Test-Ban Treaty verification regime, where those challenges have also been encountered. The examination demonstrates that two factors have the most bearing on trust between states: the imbalance between states in technical capabilities and the susceptibility of the entailed verification arrangements to politicized claims around verification gaps.