Navigating AI sustainability: A life cycle assessment approach toward viable future solutions

Abstract Besides ethical and privacy issues, Artificial intelligence (AI) also raises concerns from an environmental perspective. Training AI models requires very large datasets and entails substantial energy and water consumption (WC). Some projections suggest that, by 2027, the annual global energy demand attributable to AI could reach 85–134 TWh, while WC may amount to 4.2–6.6 billion cubic meters, equivalent to four to six times Denmark’s annual consumption and nearly half of the United Kingdom’s. These figures underscore the importance of assessing the sustainability of AI by considering its impacts across the entire life cycle. In this study, the environmental impacts of Generative AI (GenAI) were quantitatively assessed using the Life cycle assessment methodology, across 18 impact categories, based on secondary data. The results indicate that training a single GenAI model over q year could generate 767,814 kg CO₂ eq (comparable to the annual emissions of 167 cars), 190,145 kBq Co-60 eq, 13.283 kg 1.4-DCB eq (equivalent to the pesticide treatment of approximately 102 ha of agricultural land), 28,485 m 2 a crop eq (about 4 football fields), and 184,690 kg oil eq (equivalent to burning roughly 615 barrels of oil). Overall, the findings highlight the considerable environmental burden of AI, with potential negative consequences comparable to those of entire polluting industrial sectors. It should be noted, however, that these estimates are conservative and intended for illustrative purposes. Actual resource consumption may be higher, depending on factors such as data center efficiency, the energy mix employed, and specific operating conditions.