Statactivism Up and Down the Stack: Dilemmas in the Estimation of AI’s Environmental Footprint

Abstract With mounting concerns regarding the environmental footprint of AI systems, the number and diversity of actors engaged in evaluating the sustainability of machine learning (ML) and artificial intelligence (AI) more broadly is growing. Based on nine semi-structured interviews with key experts, this paper investigates the dilemmas active participants in this global project face and the strategies they employ to overcome them. Our analysis shows that experts question the extent to which quantification fosters a radical enough version of change. Some evaluators want to make AI systems more “efficient” by reducing the amount of resources needed to develop AI models and infrastructures. Others critique this approach for failing to limit overall carbon emissions. Instead they insist on making AI “frugal”, an approach which expands the range of actions to mitigate AI’s indirect environmental impacts and includes consideration of whether AI is needed at all in each specific context. This tension creates situations of emotional discomfort and divided loyalties at two key moments: access to funding and expertise. In some cases, it leads some contributors to build coalitions of workers inside and across companies in order to challenge management, or to leave their organizations altogether.