The sustainability paradox: A systematic review of carbon footprints and decarbonization strategies for artificial intelligence in climate science
Парадокс устойчивого развития: систематический обзор углеродного следа и стратегий декарбонизации искусственного интеллекта в климатологии
2026-07-21
SCID: 54.1/my8a86n7
Discuss with AI
AI carbon footprintscarbon accounting protocolclimate sciencedecarbonization strategiessystematic review
Figures from the paper
Abstract (AI)
Artificial intelligence has become essential infrastructure for climate science yet operates within a sustainability paradox—deploying carbon-intensive technologies to address climate challenges. Despite 3,000% publication growth (2021–2025), methodological fragmentation prevents systematic understanding of AI's environmental costs and effective decarbonization strategies. This PRISMA 2020-compliant systematic review analyzed 59 peer-reviewed studies from Scopus (2021–2025) examining carbon footprints and mitigation strategies for AI in climate applications. Data synthesis employed narrative synthesis for carbon quantification (Theme 1, n = 31, 52.5%), random-effects meta-analysis for technical mitigation (Theme 2, n = 42, 71.2%), and content analysis for governance frameworks (Theme 3, n = 30, 50.8%). Bibliometric analysis using Bibliometrix identified temporal patterns and keyword co-occurrence networks. Carbon footprint quantification reveals systematic 300–500% underestimation through exclusion of inference phases (69% of studies), hardware embodied carbon (10–50% lifecycle impact), and geographic variance (5–15 × across regions). Meta-analysis demonstrates technical mitigation achieves pooled carbon reduction of −32.4% (95% CI: −27.1 to −37.8%, I² = 78%, p < 0.001), with algorithmic optimization superior (−38.2%) to hardware upgrades (−24.8%) and operational strategies (−29.7%). Critical governance gaps include inconsistent system boundaries (77% of studies, 200–400% measurement variance), incomplete disclosure (only 32%), and zero universal protocols. Network centrality analysis reveals carbon quantification terms occupy peripheral positions (betweenness = 0.4) despite 52.5% coverage, while policy vocabulary dominates centrality (betweenness = 91.5). We propose the Sustainable AI Carbon Accounting Protocol (SAICAP) addressing identified gaps through five mandatory disclosure categories, three-tier compliance structure, and multi-stakeholder coordination. Integrated technical mitigation can achieve 30–45% carbon reduction while maintaining model performance, contingent upon adopting standardized measurement protocols and governance mechanisms.
Key Findings
1
A PRISMA 2020-compliant review of 59 studies found fragmented methods for assessing AI’s environmental impacts in climate science.
2
Carbon footprints are systematically underestimated by 300–500%, largely because studies omit inference emissions, embodied hardware carbon, and geographic variation.
3
Governance practices are inconsistent: 77% of studies use differing system boundaries, only 32% provide complete disclosures, and no universal accounting protocols exist.
4
Technical mitigation reduces AI-related carbon emissions by 32.4% overall; algorithmic optimization performs best (−38.2%), ahead of operational strategies (−29.7%) and hardware upgrades (−24.8%).
5
The review proposes SAICAP, a carbon-accounting protocol with five mandatory disclosure categories, three compliance tiers, and multi-stakeholder coordination.
Research Object
Artificial intelligence systems used in climate science and climate applications
Research Subject
Their carbon footprints, environmental impacts, carbon quantification, and decarbonization and governance strategies
Publication Details
Publication Date
2026-07-21
Journal
Publisher
ISSN
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest