The environmental impact of AI is a complex and multifaceted issue, and the water footprint of AI systems is a critical aspect that often goes unnoticed. While it might seem insignificant that a single chatbot answer uses a small amount of water, the cumulative effect of millions or billions of requests is what makes it a pressing concern. The water footprint of AI is not just about the direct use of water in data centers, but also about the indirect water use associated with electricity generation and the infrastructure that powers these systems.
The estimates of AI water use vary widely, and this is due to the different systems, time periods, assumptions, workloads, and boundaries involved in each study. For instance, a Google-authored paper reported that the median Gemini text prompt used a minuscule amount of water, equivalent to about five drops. This figure is in stark contrast to the bottle-scale estimates often cited for ChatGPT, highlighting the need for transparent and comparable reporting.
The location of AI infrastructure is crucial in understanding the local impact of water use. A significant number of planned U.S. data centers are located in regions that have experienced drought conditions, which raises concerns about the competition for water resources. The example of Microsoft's OpenAI infrastructure in West Des Moines, Iowa, where it used a substantial portion of the local water district's supply, underscores the real-world implications of AI's water footprint.
While it is tempting to focus on individual actions, such as writing fewer AI emails, it is essential to recognize that this cannot substitute for infrastructure disclosure. Most people cannot choose the data center handling their query or determine the water footprint of the electricity behind it. The industry's rapid growth has outpaced public understanding of its physical demands, and this gap needs to be addressed.
The real question is not about the recklessness of every AI use but about the industry's responsibility to provide transparent reporting. The public perceives AI as a clean digital service, but the physical cost is often borne by communities near the pipes, pumps, and cooling systems. Until consistent reporting is achieved, this disconnect will persist, and the environmental impact of AI will remain a hidden cost.
