The nonlinear link between AI and energy poverty: Empirical evidence from global data

Despite the burgeoning literature on artificial intelligence (AI) and sustainable development, cross-country panel research on AI and energy poverty has predominantly assumed a linear association, potentially obscuring stage-dependent dynamics. We examine whether this relationship is instead nonlinear, using a balanced panel of 64 countries from 2000 to 2019. Employing two-way fixed effects and system GMM estimations with Lind and Mehlum U-shape tests, we document a robust inverted U-shaped relationship: industrial-robot adoption is initially associated with higher energy poverty at low penetration levels but with alleviation beyond a critical threshold of approximately 358 industrial robots per million employees. Heterogeneity analysis shows that countries with stronger human capital and institutional quality exhibit flatter inverted-U curves; they are buffered from both the early adverse phase and the subsequent alleviation phase, while maintaining consistently lower energy poverty levels across the AI distribution. Mechanism tests provide suggestive evidence for two supply-side channels, namely energy efficiency enhancement and renewable energy transition. Results are robust to alternative measures, Bartik IV, and spatial lag models. Estimates should be interpreted as robust conditional associations rather than identified causal effects, and they primarily characterize middle- and high-income economies undergoing industrial transformation, with limited applicability to Sub-Saharan Africa. We organize the findings through an interpretive framework labelled the “conditional curse-remedy transformation”.