From Cradle to Cloud: A Life Cycle Review of AI's Environmental Footprint

In a new paper submitted to arXiv, researchers Katherine Lambert and Sasha Luccioni systematically review scientific literature and technical reports to map AI's environmental footprint from raw materials to end-of-life. They propose an eight-stage life cycle framework covering hardware manufacturing, infrastructure construction, data gathering, model experimentation, training, post-training adaptation, deployment/inference, and disposal. Their analysis reveals that despite the growing popularity of terms like 'green AI' and 'sustainable AI,' the underlying definitions remain inconsistent. Some studies narrowly focus on training and inference energy, while others attempt to capture broader impacts like embodied emissions and data center construction—but rarely in a standardized way.

The review also uncovers serious gaps in reporting practices. The vast majority of papers rely on CO₂e estimates derived from coarse proxies (e.g., GPU hours × regional carbon intensity), while water consumption, raw material extraction, and multi-impact life cycle assessment (LCA) are largely ignored. The authors argue that without comprehensive metrics, it's impossible to compare models or inform policy. They conclude with concrete recommendations for more rigorous measurement and reporting, aiming to support regulators and industry leaders in making data-driven decisions about AI's true environmental cost.