A Critical Window To Align Ml With Climate Change Mitigation
Ph.D. student Priya Donti of Engineering and Public Policy and the Computer Science Department, Assistant Professor Emma Strubell of the Language Technology Institute, and Assistant Professor Lynn Kaack of the Hertie School in Berlin have co-authored a perspective piece in Nature Climate Change outlining a systematic framework for describing the effects of AI on Greenhouse Gas (GHG) emissions. Their article also identifies priorities for impact assessment and scenario analysis, and suggests policy levers for better understanding and shaping the effects of machine learning (ML) on climate change mitigation.
ML is a powerful technology that is rapidly being applied in a number of domains to speed scientific innovation and advance social good. However, in evaluating the impact of ML on GHGs, researchers must also consider the many ways the technology can directly or indirectly contribute to emissions. Donti, Strubell, and Kaack, who received her PhD from EPP and masters in ML at CMU, co-authored this article to unravel this complex web of interactions and provide guidance on how to best align the impacts of ML with climate action.
“What this paper does is give an overview of the different mechanisms through which AI impacts greenhouse gas emissions,” said Donti. “And then it examines the numbers we have and the levers we have to shape that impact on emissions.”
The team’s framework divides the impacts of ML into three categories: computing-related impacts, immediate impacts of applying ML, and system-level impacts.
Computing-related impacts encompass contributions by ML to GHG emissions from its energy usage, the energy’s source, and the materials extracted and processed into hardware. Growth in renewable energy and the rapid transition to large cloud and hyperscale data centers have so far limited the impact of energy usage by ML on GHG emissions to less than .05 percent, based off current estimates—however, given the recent rapid growth in the size of the largest ML models, it is worth noting that these numbers may well change.
Regarding the immediate impacts of applying ML, Donti and Kaack helped co-author a previous paper that outlines how ML can enable or accelerate climate change mitigation and adaptation strategies in different areas such as energy, transportation, and land use. However, this newest publication also examines the potential for ML to accelerate applications in emissions intensive industries, and discusses the difficulties in predicting and measuring the impact of ML on these sources of emissions.
System-level impacts encompass the potential of ML to have broader societal implications on GHG emissions, which are difficult to quantify but may outweigh immediate application impacts. The team lays out several potential pathways this can occur: ML could enable ostensibly GHG reducing changes to technology that inadvertently promote behaviors that increase emissions; it could power developments that increase our dependency on existing technologies to the detriment of newer, potentially low-carbon alternatives; and it could influence broader lifestyle changes (e.g., via targeted advertising) that impact society’s demand for materials and energy.
The framework the authors have created could provide the basis for estimating the emissions impact of ML from a company, product, or policy. However, they note that this requires collection of data on ML’s impacts, as well as an understanding of alternatives to the implementation of ML for comparison. This will require significantly more research and information sharing to not only define the carbon footprint of ML, but to understand its role in relation to trends like digitalization and in systems like climate change modeling and energy modeling.
The authors close with their own recommendations for aligning ML with climate change mitigation. Though general climate policy may help drive the development and use of ML for climate change mitigation, they assert that climate change must be a major consideration in innovation and deployment of AI in general, stressing 3 factors in particular:
1) promoting the research, development and deployment of ML applications that are beneficial to the climate
(2) requiring transparency and accountability for those use cases that could increase emissions or otherwise counteract climate change goals, as well as on computational energy use
(3) employing climate-cognizant technology assessment for ML use cases that are not traditionally within the realm of climate policy, but where decisions today may have important implications for future climate impacts.
Regarding information, mandatory measurement and reporting of emissions data, when valuable, would enable ML emissions to be regulated by climate policy and help shape its design. However, with ML expertise often confined to a select group of academics and professionals, there’s a risk that the technology could exacerbate technological, economic, and social inequities. They explore how research, education, funding, and civic engagement initiatives could help ensure a broad group of stakeholders interested in aligning ML with the fight against climate change.
They close by concluding that the future of ML is not set in stone; however, the decisions that will define tomorrow are being made today. The world currently has, as the authors write, “a critical window of opportunity to shape the impacts of ML for decades to come.”
This work was performed in collaboration with researchers at the International Energy Agency in Paris, France, the Mercator Research Institute on Global Commons and Climate Change in Berlin, Germany, and McGill University in Montreal, Canada.