From Attributional to Consequential LCA: Which Theoretical Framework for Assessing AI’s Environmental Impacts?

The current hype surrounding generative AI is amplifying the environmental footprint of Information and Communication Technologies (ICT). In response, research into 'green AI' or 'AI for green' has surged. One area of research in this field involves assessing the environmental impacts of AI services. However, studies often lack explicit epistemological and methodological framing, which hinders the quality of their results and sometimes makes them inconsistent. In response to the absence of appropriate framing, the present study draws upon the LCA literature to propose three consistent methodological frameworks, based on their respective epistemological and methodological principles : Attributional LCA, Short-Term Consequential LCA, and Long-Term Consequential LCA. Each approach is examined in terms of its ethical foundations, mathematical logic, and practical implications for defining system boundaries, handling fixed and dynamic costs, and accounting for indirect effects. These approaches are then applied to a common use case for a generative AI service. By clarifying these epistemic and methodological principles, we aim to enhance environmental assessment practices within the sustainable AI literature and improve our understanding of the environmental impacts of this technology.