This report synthesizes key findings from a diverse range of sources, including academic literature, corporate sustainability initiatives, and emerging environmental tracking tools. Collectively, these documents provide a thorough overview of current methodologies for evaluating the environmental impacts of artificial intelligence (AI) systems. While several advances in methodology and tooling are evident, the review highlights substantial inconsistencies in how different lifecycle stages of AI are measured, analysed, and reported.
[.] One of the most pressing issues uncovered is the widespread reliance on indirect estimates when assessing energy consumption during the training phase of AI models. These estimates often lack real-time, empirical measurement. Furthermore, equally important lifecycle stages — such as inference (the operational use of models), Scope 3 emissions (from supply chains and hardware manufacturing), and infrastructure-level impacts (such as water consumption and cooling) — remain significantly underexplored. This reliance on proxies introduces substantial data gaps, impedes accountability, and restricts consumers’ ability to make informed, sustainable choices about AI.
To address these issues, the report uses a lifecycle-based approach, dividing the AI system's environmental impact into three stages: 1. Training, 2. Inference, 3. Supply Chain. For each stage, we examine measurement methodologies, identify current limitations, and offer recommendations for key stakeholder groups: developers (producers), users (consumers), and policy-makers. The overarching aim is to ensure that sustainability becomes a foundational element — embedded from the earliest stages of AI design to its deployment and continued use — rather than an afterthought." (Executive summary, pages v-vi)