New estimate reveals surprising amount of water tied to typical AI prompt

That seemingly harmless request to have ChatGPT draft an email may carry a hidden cost to our water supply.

Concern over AI's environmental impact is growing after a new estimate from researchers found that one "medium-sized" query may consume about 500 milliliters of water, roughly a standard 16-ounce bottle.

Multiply that by the hundreds of millions of users requesting multiple emails to be written every hour, and that is no small issue.

The estimate was published by the Association for Computing Machinery. In the study, a "medium-sized" GPT-3 request meant roughly 800 words of input and fewer than 300 words of output, and that interaction can require 500 milliliters of water, according to BGR.

Seventy-five percent of Americans' daily water intake is about 2.5 cups, which is a little under 600 milliliters, meaning every email AI drafts consumes almost as much water as the majority of Americans do during the entire day. As BGR noted, it also means most Americans are dehydrated.

Most of that use comes from cooling. AI systems run in enormous data centers full of servers that generate heat, and preventing those machines from overheating can take large amounts of water.

The research points to striking examples of that demand: training AI in Microsoft's U.S. data centers used 185,000 gallons of onsite water and about 1.4 million gallons overall, while a Google-owned data center consumed 6.07 billion gallons of freshwater for onsite cooling in 2023.

AI's environmental toll is not only extra electricity-based pollution. Water use is often an under-reported part of the footprint, especially since about two-thirds of new AI facilities are being built in some of the driest parts of the country.

Demand for AI, and subsequently water use, will only increase. McKinsey & Company projects that companies will spend $5.2 trillion on data centers by 2030 to keep up with AI demand, per BGR.

More than 500 are planned for places that have already experienced severe drought, and in Texas, data centers could account for nearly 9% of statewide water consumption by 2040.

Data centers draw from the same regional resources that communities rely on. In drought-prone places, more industrial water use can add stress to local supplies and raise questions about who gets priority when conditions worsen.

Researchers behind the paper say one major solution is transparency: Companies should do a better job of measuring, reporting and tracking water use, including onsite cooling and the offsite water tied to electricity generation.

The researchers added that AI systems should be built and trained in places with better water efficiency, and that workloads should be shifted to hours that are more water-efficient. They also warned that more efficient future models may not solve the problem if rising AI demand erases those gains.

Get TCD's free newsletters for easy tips, smart advice, and a chance to earn $5,000 toward home upgrades. To see more stories like this one, change your Google preferences here.