of domestic political unrest; the technologies, once deployed, in turn suppress subsequent unrest.54 The scope of AI surveillance extends beyond physical monitoring to include analysis of online behavior, social networks, and communication patterns. At least 75 countries were using AI surveillance technologies as of 2019, and adoption was increasing in all regions, the Carnegie Endowment’s AI Global Surveillance Index reports.55 This figure is likely a lower bound given the improvements in performance of recent AI models. Many low- and middle-income countries have weak rule of law, limited checks on executive power, and constrained civic space. In such contexts, AI surveillance technologies can shift the balance between the state and citizens. In the absence of robust data protection laws, independent judiciaries, and effective oversight mechanisms, misuses of AI face few institutional constraints. The use of AI for surveillance or predictive policing can also result in mistargeting and disproportionate enforcement among marginalized groups due to preexisting biases in policing data that get encoded in AI algorithms.56 This problem extends to judicial decision-making, in which algorithmic tools used to inform bail, sentencing, and parole decisions can reproduce racial and socioeconomic disparities.57 These problems may be compounded in developing countries, where training data are scarcer and populations may be more diverse. The underlying issue is that supervised learning methods reproduce the patterns present in their training data, including discriminatory ones. Another concern is that AI surveillance could constrain civic space even without formal repression as people change their behavior because they know they are being monitored (so-called chilling effects). The extent to which this dynamic operates in low- and middle-income countries is an empirical question that warrants further research. AI surveillance systems also raise concerns about due process. Unlike traditional surveillance, which typically requires warrants and documentation, AI systems can flag individuals automatically and trigger investigations without human review. The opacity of these systems means that individuals may be unable to contest their classification, even when it is based on flawed data. An additional concern is that technologies deployed for legitimate governance objectives can be repurposed for surveillance. During the COVID-19 pandemic, Israel repurposed phone-tracking technology originally developed to combat terrorism for contact tracing to identify, notify, and monitor individuals who may have been exposed to the virus, prompting legal challenges.58 Developing countries often lack the institutional checks, such as independent courts, privacy regulators, and active civil society, needed to prevent such function creep. The global diffusion of surveillance capabilities raises questions about the responsibility of the countries and firms that export these technologies.59 Without international norms governing the export of surveillance technologies, these tools will continue to reach governments with weak accountability mechanisms. AI can also enable citizen oversight of government. In Ukraine, for instance, the nongovernmental organization Transparency International Ukraine runs DOZORRO, which applies machine learning to the government’s e-procurement system and ProZorro, the public electronic procurement system, to flag irregularities and routes citizen and civil society complaints to regulators and law enforcement bodies.60 Since 2017, the project has reviewed thousands of risky tenders and has helped save Ukraine billions of dollars in public funds.61 Whether AI strengthens accountability or surveillance depends on the institutional context and the distribution of technical capacity between the state and civil society. AI’s Political Impact: Reshaping Power Within and Across Countries 217