Measuring the environmental impacts of artificial intelligence compute and applications

The green and digital “twin transitions” offer the promise of leveraging digital technologies for a sustainable future. As a general-purpose technology, artificial intelligence (AI) has the potential not just to promote economic growth and social well-being, but also to help achieve global sustainability goals. AIenabled products and services are creating significant efficiency gains, helping to manage energy systems and achieve the deep cuts in greenhouse gas (GHG) emissions needed to meet net-zero targets. However, training and deploying AI systems can require massive amounts of computational resources with their own environmental impacts. The computational needs of AI systems are growing, raising sustainability concerns. While AI can be perceived as an abstract, non-tangible technical system, it is enabled by physical infrastructure and hardware, together with software, collectively known as “AI compute”. In the last decade, the computing needs of AI systems have grown dramatically, entering what some call the “Large-Scale Era” of compute. At the same time, according to the International Energy Agency (IEA), data centre energy use has remained flat at around 1% of global electricity demand, despite large growth in workloads and data traffic, of which AI is estimated to represent a small fraction. While this may point to hardware efficiency gains, some researchers note that AI compute demands have grown faster than hardware performance, bringing into question whether such efficiency gains can continue. The environmental impacts of AI compute and applications should be further measured and understood. Policy makers need accurate and reliable measures of AI’s environmental impacts to inform sustainable policy decisions. The 2010 OECD Recommendation on ICTs and the Environment encourages the development of comparable measures of environmental Information ICT impacts. Further, the 2019 OECD Recommendation on Artificial Intelligence underlines that AI should support beneficial outcomes for people and the planet. The 2021 OECD Recommendation on Broadband Connectivity also stresses the need to minimise the negative environmental impacts of communication networks. Yet further efforts are needed to develop measurement approaches specifically focused on AI and its environmental impacts. The report defines AI compute as including one or more “stacks” (i.e. layers) of hardware and software used to support specialised AI workloads and applications in an efficient manner. This definition was developed by the OECD.AI Expert Group on AI Compute and Climate (the “Expert Group”) to meet the needs of both technical and policy communities. Informed by the Expert Group and experts involved in the Global Partnership on AI (GPAI), this report synthesises findings from a literature review, a public survey and expert interviews to assess how the environmental impacts of AI are currently measured. A number of indicators and measurement tools can help quantify the direct environmental impacts from AI compute, as well as the indirect environmental impacts from AI applications. The report distinguishes between direct and indirect positive and negative environmental impacts. Direct impacts stem from the AI compute resources lifecycle (i.e. the production, transport, operations and end-of-life stages). Analysis indicates that direct impacts are most often negative and stem from resource consumption, such as the use of water, energy and its associated GHG emissions, and other raw materials. Indirect impacts result from AI applications and can be either positive, such as smart grid technology or digital twin simulations, or negative, such as unsustainable changes in consumption patterns.