← Back to Library

Empire of AI is wildly misleading about AI water use

Andy Masley delivers a rare, forensic takedown of a best-selling narrative, exposing how a single mathematical error in Karen Hao's Empire of AI has distorted the global conversation on technology and water. While the book frames data centers as colonial plunderers draining the world's rivers, Masley demonstrates that the core statistics are off by a factor of 4,500, turning a legitimate concern into a "wildly misleading" alarmist fantasy. For busy readers trying to separate signal from noise in the AI boom, this is a crucial correction that demands attention before the next policy cycle begins.

The Illusion of Consumption

The piece opens by dismantling the book's most cited statistic: the claim that AI will consume 1.1 to 1.7 trillion gallons of fresh water by 2027. Masley writes, "The study mentioned is 'Making AI Less Thirsty.' The study does not say that AI demand will consume 1.1-1.7 trillion gallons of water annually." He points out that the author of the book confused "withdrawal"—water taken from a source and often returned—with "consumption," which is water permanently removed via evaporation.

Empire of AI is wildly misleading about AI water use

This distinction is not merely semantic; it is the difference between a temporary disruption and a permanent drain. As Masley notes, "Withdrawal is very different from consumption. Withdrawal means the amount of water taken from a local source. Consumption is the amount of water taken and not returned to the local source." He illustrates this with a simple analogy: diverting a river to run a mill and returning it is not the same as sucking the river dry. The book's narrative relies on the reader assuming the worst-case scenario, yet the data suggests 90% of the water cited is returned to the source unaffected.

"Every single time water's mentioned in this book, the reader is left with a worse understanding than they came in with."

The error compounds when looking at the type of water used. The book implies a massive drain on potable (drinking) water, but Masley clarifies that the vast majority of the water involved is used by offsite power plants, not the data centers themselves. "Basically none of the water power plants use is potable," he argues. "The only potable water used for AI is in data centers themselves." When corrected for this, the actual volume of drinking water at risk is a mere fraction of the book's claim—roughly 3% of the number Hao presents. Critics might argue that even small amounts of potable water are precious, but Masley's point stands: framing the issue as a "drinking water crisis" when it is largely a matter of industrial cooling water creates a false sense of urgency that distracts from real resource management.

A Factual Catastrophe in Chile

The most damning section of Masley's analysis focuses on a specific case study in Chile, where the book claims a Google data center uses 1,000 times more water than a city of 88,000 people. Masley identifies this as "the single largest error in any popular book that I've found on my own." He breaks down the math: the author of the book divided the data center's usage by a municipal total that was itself off by a factor of 900.

"The average citizen of Chile buys 180 liters per day from their municipal water systems," Masley writes, contrasting this with the book's implication that residents were using a fraction of a liter. "This implies that a city of 88,000 people is using as much water per day in total as a single shower head left running." The result is a narrative where a single building is portrayed as a resource vampire consuming the equivalent of a city of 88 million people.

This error is particularly damaging because it is woven into a broader argument about colonialism. The book frames the data center as a continuation of historical plunder, yet Masley notes that the facility would actually use only ~0.3% of the municipal water system. "If instead you see data centers using water in other countries as part of a simple trade the countries are making to get more taxable industry in the area, and that doesn't seem to harm water access, the central narrative thrust of the chapter becomes false." The stakes here are high: by misrepresenting the data, the book undermines the credibility of legitimate environmental activism.

"She implied that a single building is using as much water as a city of 88 million people. That's over 4x the entire population of Chile!"

Masley also contextualizes this within the broader history of water infrastructure, noting that while the book cites a "megadrought" in Chile, the data center's actual impact is negligible compared to agricultural and industrial usage in the region. He references the historical context of cooling towers and data center water footprints to show that modern facilities are often more efficient than the public imagines, yet the book's "wildly misleading" claims prevent a nuanced discussion.

The Cost of Bad Data

The piece concludes by addressing the broader implication: why does this matter? Masley argues that the book has become a primary reference for the AI/environment conversation, meaning its errors are now being cited in policy debates and media coverage. "One of the most common replies I've received to my water arguments is that I should read it," he writes. "So I did, and I came away kind of shocked at how badly it covered the issue."

The danger is that by painting a picture of "brutal acts of torture and plunder" where none exists, the book alienates potential allies and obscures the actual challenges. As Masley puts it, "My only ask for people writing about AI and the environment is that at the end, readers are left with a more accurate picture of how energy and water is being used overall in the regions covered." The alternative is a public discourse driven by "contextless impression of AI as a huge environmental offender in places where it's not."

Critics might suggest that highlighting the potential for future strain is a valid precautionary principle, even if current numbers are low. However, Masley's evidence suggests that the book isn't just being cautious; it is actively misrepresenting the present reality, which makes it harder to address genuine future risks when they arise.

Bottom Line

Andy Masley's critique is a masterclass in fact-checking, exposing how a single unit conversion error can derail a book's entire thesis and mislead a generation of readers. While the book's intent to scrutinize the environmental cost of AI is noble, its reliance on flawed data renders its central narrative about "plunder" not just exaggerated, but factually inverted. The strongest takeaway is that accurate environmental stewardship requires precision, not panic, and the AI industry's water footprint is far more complex—and far less catastrophic—than popular literature suggests.

Deep Dives

Explore these related deep dives:

  • Water footprint

    The article centers on the distinction between water withdrawal and water consumption for AI data centers. Understanding water footprint methodology—including the difference between blue, green, and grey water—provides essential context for evaluating claims about AI's environmental impact.

  • Data center

    The article critiques claims about data center water and energy use. A technical understanding of how data centers actually work, including their cooling systems and infrastructure requirements, helps readers evaluate the competing claims about resource consumption.

  • Cooling tower

    The distinction between consumptive and non-consumptive water use in the article relates directly to how cooling systems work at both data centers and power plants. Understanding evaporative cooling versus closed-loop systems explains why withdrawal numbers are so much larger than consumption numbers.

Sources

Empire of AI is wildly misleading about AI water use

by Andy Masley · · Read full article

Note: the author took the time to respond to me below. While I’m very grateful, the materials she sent actually seems to confirm my main criticism and I’m now very confident a key number in the book is 1000x too large and needs to be revised. I summarize everything in my reply here.

I was taking a break from posting about AI and the environment, but after reading parts of Karen Hao’s book Empire of AI, I’ve stumbled on such wildly misleading claims that have so far gone unaddressed that I’ve felt the need to counter them here. Within 20 pages, Hao manages to:

Claim that a data center is using 1000x as much water as a city of 88,000 people, where it’s actually using about 0.22x as much water as the city, and only 3% of the municipal water system the city relies on. She’s off by a factor of 4500. This is the single largest error in any popular book that I’ve found on my own, and to my knowledge I’m the first person to notice it.

Imply that AI data centers will consume 1.7 trillion gallons of drinkable water by 2027, while the study she’s pulling from says that only 3% of that will be drinkable water, and 90% will not be consumed, and instead returned to the source unaffected.

Paint a picture of AI data centers harming water access in America, where they don’t seem to have caused any harm at all.

Frame Uruguay as using an unacceptable amount of water on industry and farming, where it actually seems to use the same ratio as any other country.

Frame the Uruguay proposed data center as using a huge portion of the region’s water where it would actually use ~0.3% of the municipal water system, without providing any clear numbers.

These are all the significant mentions of data centers using water in the book. Read in this light, the chapter becomes somewhat ridiculous, because the rest includes descriptions of brutal acts of torture and plunder under colonialism, and then frames data center water use as a continuation of that same colonialism. If instead you see data centers using water in other countries as part of a simple trade the countries are making to get more taxable industry in the area, and that doesn’t seem to harm water access, the central narrative thrust of the chapter becomes false.

This ...