The hidden environmental impact of AI is a topic that demands our attention, and it's not just about the energy consumption we often hear about. Water, a precious resource, is also being consumed at an alarming rate by the infrastructure supporting AI technologies.
When we think of AI, we envision sleek interfaces and seamless interactions, but the reality is far from weightless. The machinery behind the scenes is a complex web of hardware, from chips to cooling towers, all requiring substantial resources.
The Water Footprint of AI
Writing a simple 100-word email with ChatGPT, for instance, consumes water equivalent to a standard bottle. Now, imagine the cumulative effect of millions of such requests. The global AI infrastructure is projected to use a significant portion of the UK's annual water withdrawal by 2027, and much of this water is sourced from drought-prone regions.
What's particularly concerning is the hidden nature of this cost. It's not a single, large drain on resources, but a multitude of small ones that add up over time and across the globe.
Estimating AI's Thirst
Estimates of AI's water consumption vary widely. A 2024 Washington Post analysis, in collaboration with researchers from the University of California, Riverside, estimated the water needed for ChatGPT to generate an average email at an average American data center, equating it to a bottle of water.
However, this number is not static. It depends on various factors, including the AI model, the data center's location and infrastructure, and even the local weather. The same prompt could have a different environmental impact depending on when and where it's processed.
A study by Li, Yang, Islam, and Ren estimated that a model like GPT-3 could consume 500 milliliters of water for 10 to 50 medium-length responses, highlighting the importance of considering the geographical context of water use.
The Local Impact
Water is not a universal resource. A liter drawn from a wet region is not equivalent to a liter taken from a stressed aquifer. This is why the location of AI infrastructure matters.
A 2026 Guardian analysis reported that a significant number of planned U.S. data centers are in drought-prone areas. This raises concerns about the competition for water resources between AI infrastructure and local communities, farms, and ecosystems.
Beyond Chatbots
It's tempting to think that the solution is as simple as reducing our AI email usage, but the issue is more complex. While shorter prompts and answers can reduce resource consumption, the onus should not be solely on individual users.
Most people have no control over which data center handles their queries, and they cannot determine the environmental impact of each response. The problem lies in the lack of transparency and comparable reporting on AI's infrastructure and resource use.
The Real Question
The issue is not about recklessly using AI but about the industry's rapid growth outpacing public understanding of its physical demands. AI's interface may be digital, but its infrastructure is very much physical and resource-intensive.
The bottle-of-water comparison serves as a warning, but it's not a universal metric. We need to ask more specific questions: where is the computation happening, what model is being used, what water sources are involved, and who else relies on these resources?
Until we have consistent reporting on these factors, the public perception of AI will remain detached from its environmental reality. It's time to bring these hidden costs into the light and ensure that the communities bearing the physical costs are not left behind.
