|
|
|||
|
|
|||
|
The
Collision of Climate Science |
|||
|
Author: Madeline Roberts, PhD, MPH Generative AI has grown and achieved widespread use in an unregulated industry where companies and developers are not required to report resource utilization or environmental impact. The increasing use of generative AI technologies, which require substantial computational power, is also producing ballooning emissions. The health impacts of air pollution and climate change are substantial and include cardiovascular and respiratory diseases such as asthma, changes in the distribution of disease vectors and related illnesses, and heat-related illness, among others. Additionally, the deleterious health effects of emissions are not distributed equally, with low-income areas bearing a greater burden of harm compared with more affluent areas. AI data centers are also consuming massive amounts of resources. In one example—after being legally compelled to report how much water it was using in The Dalles, Oregon—Google disclosed that its data centers were using almost 25% of all available water in the city. Google is struggling to realize its goal of net-zero emissions by 2030 and disclosed in July that company emissions rose by 48% compared to their 2019 baseline, mostly attributable to AI initiatives and data centers. Aiming to replenish 120% of the water it consumes by 2030, the company has currently replenished 18%. Google is not alone, Meta’s emissions are approximately 70% over their 2019 levels, and Microsoft reported emissions growth 29% over its 2020 baseline.
As the environmental cost of developing and maintaining such massive data infrastructure comes under sharper review, tech giants are collaborating with power producers to secure novel sources of clean energy, including harnessing heat below the earth’s surface and small nuclear reactors. Researchers have developed different methods for measuring and communicating the environmental impact of AI. A preprint by Lacoste et al presents an environmental Impact Calculator for Machine Learning (interactive calculator here) where companies and developers can estimate their carbon emissions. Crawford and Joler published their stunning diagram and narrative Anatomy of an AI System (currently on exhibit at the Museum of Modern Art) which details the cost of AI in terms of material resources, human labor, and data, spanning the inception of a device to when it becomes e-waste. The AI model training and inference phases exact different environmental demands, and while shorter in duration, the training phase is often the most environmentally taxing owing to the multitude of servers and graphics processing units (GPUs) required to train a model. The environmental impact of the inference phase is directly proportional to the model’s number of users, with more users yielding a higher environmental impact. For reference, generative AI searches use approximately 10 times more energy compared to standard searches using a search engine like Google. To understand just how much water and electricity data centers require, researchers at UC Irvine developed the infographics shown on the next page. There is indication that time of day as well as location can substantively reduce the amount of water used in AI data centers. Additionally, some have suggested regulatory legislation establishing energy and water benchmarks for tech companies, as well as incentivizing renewable energy use. At the very least, transparency in this area could look like implementing standardized, third-party environmental impact studies with routinely reported results. Currently, The Artificial Intelligence Environmental Impacts Act of 2024 calls for measurement and reporting standards for the environmental impact of AI, though developers reporting the results would be voluntary. It also would require an interagency study to quantify the positive and negative ecological impacts of AI. The bill’s fate in Congress can be tracked here. The environmental burden and health implications of AI energy consumption continue to increase along with unreached climate targets. A concerted effort from tech companies, clean energy power producers, researchers, and policymakers is urgently needed to move toward sustainable AI.
■ |
|||
|
|
|||
