What Counts as Green AI? Mapping Efficiency, Sustainability, and Critical-Ecological Strands in a Fragmented Discourse

The concept of “Green AI” is emerging as stakeholders confront the environmental costs of artificial intelligence. However, the field remains formative and contested, marked by competing definitions, inconsistent methods, and limited cross-disciplinary collaboration. Efficiency-driven approaches frame Green AI narrowly as reducing computational and carbon costs through technical optimizations such as model pruning and energy reporting. Sustainability-driven perspectives view AI as a tool for ecological problem-solving—ranging from climate modeling to biodiversity monitoring—often linked to global policy agendas like the SDGs. In contrast, critical-ecological critiques warn that both efficiency and sustainability narratives risk obscuring exploitative infrastructures, from cobalt mining to water-intensive data centers, and reinforcing global inequalities. These perspectives rarely converge, producing conceptual ambiguity, fragmented methodologies, and a persistent policy–practice gap. To address this, the paper develops a typology that distinguishes efficiency, sustainability, and critical-ecological strands, clarifies their assumptions, and highlights their blind spots. By framing Green AI as a contested boundary project, the typology provides a foundation for methodological standardization, interdisciplinary integration, and more accountable research. Future work should build on this typology to establish shared metrics and justice-oriented practices that align AI innovation with planetary limits.