Energy and Carbon Footprint of Artificial Intelligence: Measurement, Disagreement and Mitigation

Estimates of the energy used by artificial intelligence are widely quoted and rarely agree with one another. This review examines why. We separate what is counted (training or inference, operational or embodied emissions, carbon or water) from how it is counted (meters, software trackers, calculators, retrospective estimates, life cycle assessment), and then set published figures side by side with their scope. The best-known early number, an outside estimate for one architecture search, was later put by the operator at one eighty-eighth of that value. One measured study that drew its boundary wider found idle power and manufacturing roughly doubling the footprint of a training run. Inference energy per thousand queries spans three orders of magnitude across tasks. At data-centre scale the AI-specific figures are scenarios built on analyst estimates, and their author says so. We then review mitigation through smaller models, better hardware, scheduling by place and hour, and reporting, and we look at the rebound argument. The article is a review of published work. No new measurement or experiment is reported, and each number is attributed to the source that produced it.