Emma Strubell and colleagues, 2019

“These accuracy improvements depend on the availability of exceptionally large computational resources that necessitate similarly substantial energy consumption.”

Emma Strubell, Ananya Ganesh, and Andrew McCallum examined the compute, energy, and estimated emissions associated with training large natural-language processing models. Their analysis showed how architecture search and repeated experiments could multiply costs far beyond one final training run. The paper broadened evaluation by asking researchers to account for resources, access, and environmental externalities alongside incremental gains in accuracy. Its historical importance comes from the bridge it built between an abstract idea and evidence that other researchers could inspect, challenge, and extend.

A model’s published score does not reveal the complete experiment behind it. Hyperparameter sweeps, failed runs, data processing, and hardware choices all consume resources. These costs affect the climate and determine which laboratories can participate. The authors’ intervention treated efficiency as a research value and a reporting obligation, making hidden infrastructure visible enough to compare and improve. A responsible interpretation therefore asks what the system represents, what it ignores, and how its objective changes behavior once it leaves the laboratory.

AI systems have grown dramatically since 2019, while accelerators, data centers, and training methods have also become more efficient. Absolute impact depends on energy sources, utilization, cooling, location, model reuse, and inference demand. Teams should measure rather than assume. Reporting compute, electricity, carbon intensity, and useful work can guide scheduling, architecture choices, distillation, caching, and selection of smaller models when they meet the need. Current AI systems operate inside social and technical networks, so evaluation must include workflows, incentives, maintenance costs, and the behavior of people adapting to the tool.

Efficiency gains can be overwhelmed when cheaper computation drives much more use, a rebound effect that makes per-query improvements insufficient. Environmental reporting can also become selective marketing if organizations omit water, hardware production, or repeated experimentation. Credible practice uses consistent boundaries, independent review, and product-level decisions about whether an application’s value justifies its resource demand. The quotation ultimately rewards intellectual humility: state what is known, invite scrutiny, and change course when evidence from the real world contradicts the model. Strubell and her coauthors made resource use part of responsible model evaluation. Their question remains urgent: what improvement was achieved, and what complete cost produced it? The enduring question is therefore practical as well as scientific: what would count as success, and who gets to decide when the answer changes? That habit of precise comparison keeps innovation open to correction and makes it easier to distinguish a durable contribution from confidence produced mainly by scale or novelty.