Sustainable AI Research?

Three Approaches Within High-Performance Computing Research Communities and a Critique from the Margins

Abstract

High-performance computing research, particularly concerning so-called artificial intelligence, is increasingly scrutinised for its environmental impacts. Many researchers are actively debating how to respond. Surveying recent English language research community publications, this paper identifies three distinct but overlapping proposed approaches to greening compute-heavy machine learning research: 1) efficiency, 2) the net benefit of downstream impacts, and 3) virtuous competition through environmental metrics. Alternative, and much more radical, proposals can be found in a small but growing contingent of scholars who seek to 4) subordinate research questions to planetary limits. Each approach is more than merely technical, with significant considerations given to researchers’ social incentives and relational commitments to one another. Nevertheless, all approaches are constrained by limitations in enumerating compute and mobilising social incentives. This article assesses tensions across these approaches and concludes by connecting them to wider discussions on the constitutive and increasingly planetary character of method in STS.