AI Safety in Resource-Constrained Environments: Insights from an Intelligent Energy Management System in Cameroon

AI safety research has largely focused on high-resource settings: reliable power,stable internet, and unlimited compute. This paper argues that this framing ex-cludes most of the world. Using a case study of an AI-powered energy managementsystem (ISI) deployed in Yaoundé, Cameroon, we identify unique safety challengesthat emerge when infrastructure is unreliable. Drawing on the Dynamic RewardMDP (DR-MDP) framework from recent AI alignment literature, we analyze threefailure modes: (1) sensor degradation under voltage spikes, (2) the offline safetyproblem, and (3) asymmetric cost ratios between false positives and false nega-tives. We derive a robust decision criterion based on minimax regret and proposedesign principles for AI systems in low-infrastructure settings. This paper is a callto broaden the AI safety agenda to include the environments where most of theworld lives.