Rebound Effects of AI on Sustainability: Economic and Policy Perspectives

Artificial Intelligence is reshaping economies and everyday life, but its environmental consequences are complex and often hidden. This manuscript presents a systematic review of evidence from 2010 to 2025 that examines how AI efficiency gains translate into energy use, greenhouse gas emissions, and rebound ects across sectors. A total of 1,293 records were screened across major databases and policy repositories. Findings from 101 eligible studies were synthesized, with reported metrics standardized into kilowatt houreffs and kilograms of carbon dioxide equivalent to enable consistent comparison of training and inference impacts. The synthesis shows that training large language and vision models can consume from hundreds of megawatt hours to several gigawatt hours per run, and that inference energy grows with deployment scale. Hardware and software improvements have raised performance per watt, and data center efficiency has improved, but these gains are often offset by service level rebound when lower costs and better services increase total usage. Concrete examples include machine translation, image generation, and personalized recommendation systems, all of which have driven substantial increases in user demand and aggregate compute. At the economy level, computable general equilibrium and input-output models report rebound magnitudes commonly between 30 and 60 percent under plausible scenarios, while behavioral channels such as increased comfort taking and spare time reallocation further reduce net savings. Sectoral analysis highlights elevated rebound risk in transport, buildings, industry and agriculture. Measurement is hindered by inconsistent system boundaries, limited longitudinal data and differing model assumptions, which together produce wide estimate ranges and limit precise quantification. A pragmatic policy and research agenda is recommended: Harmonized reporting standards for compute and energy should be adopted, lifecycle assessment should be paired with demand side and behavioral models, and transparent energy and carbon disclosure should be mandated for major AI systems. Technical measures such as model distillation and on device inference can be combined with market instruments, including carbon pricing and clean energy procurement. In addition, user-facing transparency tools and demand management strategies can help limit behavioral rebound effects. By bringing rebound effects into routine evaluation and governance, AI can be steered toward real sustainability gains. This review offers evidence-based guidance for policymakers, industry, and researchers who aim to align AI innovation with climate objectives.