AI’s Water Footprint: The Environmental Price of Innovation
Source article: Li, P., Yang, J., Islam, M., & Ren, S. (2023). Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models. https://arxiv.org/pdf/2304.03271
Before and after: California lake reborn following winter storms. (2023, April 21). Reuters. https://www.reuters.com/sitemap/authors/
AI Models Are Thirstier Than You Think
The use of artificial intelligence (or AI) is becoming increasingly common, with this technology encroaching more and more into people’s daily lives. While this almost magical technology has generated a lot of public interest and praise, AI models come with some major setbacks–particularly when regarding their impact on the environment. Unfortunately, the use of AI is incredibly resource intensive and results in a large water footprint. The various stages in the production and maintenance of generative AI models and systems require huge amounts of freshwater. Everything from data processing, experimentation, training, deployment, manufacturing, and consumer usage require water for cooling systems and electricity generation. To uncover the true cost of AI’s water use, researchers from UC Riverside and UT Arlington have posed a critical question: How much water does AI actually use, and more importantly, how can we reduce it?
The Reality of Consuming AI
The training process for one of the most popular AI models, GPT-3, uses up to 5.4 million liters of clean freshwater in cooling servers and producing electricity alone. For scale, this number is roughly equivalent to the amount of drinking water used by 27,000 people in a day. It is primarily used to cool high-powered servers that generate enormous heat and to help produce the electricity they run on. In other words, every interaction with a chatbot such as GPT-3 has a hidden cost in water use that most people do not consider.
Estimating the Full Impact of AI on Water Usage
Researchers developed a method that includes three water usage scopes to measure the full impact of AI: direct/on-site water usage, indirect/off-site water usage, and supply-chain water for server manufacturing. On-site water usage refers to the water used directly at the data center facility, primarily for cooling the servers. Since almost all of the server energy is converted into heat, cooling is necessary to avoid overheating such as through water intensive cooling towers. On-site water usage is quantified by looking at on-site Water Usage Effectiveness (WUE), which is the ratio of on-site water consumption to server energy consumption. Off-site water usage accounts for the water used indirectly such as in electricity generation for the data center. This off-site usage is quantified using the Electricity Water Intensity Factor (EWIF) which is the ratio of off-site water consumption for each kWh of electricity consumed. Supply-chain water usage includes the water used in the manufacturing of AI chips and servers, as the fabrication of wafers and cooling of semiconductor plants require huge amounts of water. It is quantified by measuring the total amount amortized over the expected lifespan of the server hardware. Using these scopes provided a holistic depiction of the water usage both on and offsite the actual data centers themselves. The researchers also looked at how water efficiency varies depending on where and when AI runs. Cooler climates and cleaner energy sources are much more efficient, where places like Arizona can use over twice as much water in comparison to a cooler and wetter place like Denmark.
Where Does This Leave Us For the Future?
This study estimated that by 2027, the global AI industry could withdraw 4.2 to 6.6 billion cubic meters of freshwater per year–a number that equals about half of the United Kingdom’s total annual water withdrawal as of 2020. The water footprint created from AI is going to keep growing, and this has dire consequences for the planet which relies on fresh water as a basic human need. For the average AI user, this water usage can be seen through GPT-3, a popular AI. For GPT-3 to generate roughly 10-50 medium length responses, about 500ml of water is needed. Yet, the spatial and temporal diversity of water efficiency reveals a shocking misalignment, where optimizing for carbon efficiency does not necessarily improve, and may even worse, water efficiency. So ultimately, generative AI use has huge implications for our planet on multiple fronts, from creating large water footprints, to exacerbating them, even when trying to shift away from carbon related energy.
To address these challenges and build sustainable AI, these researchers suggest a few effective strategies focused on transparency and resource management. First, they suggest increasing transparency and reporting regarding water consumption (similar to how they must report carbon emissions), as even direct water use is often currently omitted. Indirect water use is crucial to report to reduce AI’s true water cost in electricity generation. The obscurity of manufacturing water use leads the article to recommend further research to accurately estimate this aspect of water use. Second, the study emphasizes scheduling training and usage of AI systems during cooler times of the day, or even building data centers in cooler regions or places with more sustainable water systems. The transparency suggestions in reporting in tandem with the ‘when’ and ‘where’ aspects of running AI models can significantly cut the water footprint of AI, which in turn will be able to help the people with accessing fresh water. With less fresh water being used to operate these systems, more can go back into our communities to support those in need who are struggling with water access.
Ultimately, this study reveals how harmful the growing age of AI can be on an environmental level and why change has to be made now. As AI continues to become a bigger part of everyone’s daily lives, it is time to recognize the full cost of its development and use. Understanding its water impact and advocating for more transparency in tech industries is a step toward a more sustainable future.
Reviewed by:
