AI's Cross Border Energy and Water Footprints: State Duties, No Harm Thresholds, and Paris Agreement Compliance Frameworks
DOI:
https://doi.org/10.70670/sra.v3i3.875Keywords:
AI Environmental Footprint, Transboundary No‑Harm, Virtual Water Accounting, Paris Agreement Compliance, Sustainable Data Centers.Abstract
The world’s largest artificial intelligence models now draw more electricity than some mid-sized nations and evaporate billions of liters of freshwater each year, yet neither international climate law nor the classical transboundary harm doctrine has fully absorbed their impact. This article conducts a systematic review of empirical footprint studies with a comparative legal analysis of the no harm principle, emerging corporate due diligence statutes, and transparency rules of the Paris Agreement. Lifecycle data show that training a single GPT 3 class model consumes about 1.3 GWh of power and 5.4 million L of water, while global inference loads could withdraw 22 billion L annually by 2027—concentrated in already stressed basins. Because affordable mitigation tools (carbon-aware routing, liquid cooling, and a mixture of expert architectures) can reduce these impacts by 40–60 percent, failure to deploy them breaches the due diligence standard embedded in Trail Smelter and its progeny. This study proposes a hybrid allocation framework that attributes operational footprints to host states but assigns embodied and service-based impacts to consumer states, enabling parties to integrate Scope 3 emissions and virtual water transfers into Biennial Transparency Reports without amending treaty text. Embedding dual carbon and water baselines into Article 6 crediting schemes would channel finance toward low-impact data centers and close a rapidly widening governance gap.