In 2025, award-winning journalist Karen Hao released Empire of AI, a New York Times bestselling book that seemed to give the anti-AI movement the smoking gun it had been looking for. Buried inside its pages was a statistic so dramatic it practically sold the book on its own: a proposed data center in a small Chilean town was projected to consume more than a thousand times the amount of water needed to sustain the town itself. Imagine that for a moment. A single industrial facility arriving in a quiet community and drinking a thousand years’ worth of the town’s future in a single gulp. It was the kind of number that made people gasp in interviews, that made environmentalists book TV appearances, that made activists stand outside tech campuses with signs about stolen water. The outrage felt righteous because the number felt impossible to ignore. But then something embarrassing happened. The statistic was wrong. The figure had been built on a basic unit error—confusing liters per second with cubic meters per hour—and the data center’s projected water use was actually over a thousand times smaller than originally claimed. The correction never stood a chance. By the time anyone noticed, the story had already been shared, embraced, and weaponized. The episode was a perfect parable for our moment: a vivid, frightening number, repeated so often that it became truth, while the boring mathematical correction was left to gather dust in a footnote. Good intentions, it turns out, are not enough to prevent bad information from going around the world before the truth can put its shoes on.
Artificial intelligence has become impossible to ignore. It is in our search engines, our cars, our banking apps, our health apps, and increasingly in the way we work, write, and even talk to one another. That growing presence has provoked a growing backlash, and some of it is entirely justified. There are legitimate concerns about AI replacing human workers, about the enormous amounts of electricity it consumes, about bias and surveillance and the terrifying possibility that we are building a technology we cannot fully control. But in the strange ecosystem of online outrage, one of the loudest complaints has been about water. Scroll through social media, open a political spam email, read a celebrity post about the environment, and you will see the same alarming claim again and again: AI is drinking our world dry. Headlines warn that every little chat with a bot costs bottles and bottles of fresh water. The implication is that our daily conversations with a chatbot are draining reservoirs, depleting rivers, and turning the American West into a wasteland. The reality is far more complicated and far less dramatic. This is not to say that AI uses no water. Of course it does, just as every industry does. But when the story is inflated by orders of magnitude, it stops being a useful environmental criticism and starts being noise. And that noise does real damage. It distracts from the actual dangers of AI, sours the public conversation, and makes it harder for lawmakers, scientists, and citizens to have the kind of careful, honest discussions we actually need.
Let’s look at the numbers, because they matter. A 2026 Forbes article cited one study estimating that a single ChatGPT-4 query uses anywhere from two milliliters to 150 milliliters of water, depending on factors like the length of the prompt and the efficiency of the data center. The high end of that estimate is about 30 percent of a standard bottle of water. A 2024 Washington Post article, meanwhile, claimed that a single 100-word prompt to ChatGPT uses about 519 milliliters—just over one standard water bottle. Those estimates differ by more than an order of magnitude, which tells you how difficult it is to measure something that depends on weather, server efficiency, electricity sources, and the complexity of every question asked. But let’s take the higher number at face value. Let’s assume that every time you ask ChatGPT something, you are using a full bottle of water. Now compare that to the rest of your day. It takes about 660 gallons of water to make one hamburger. That is roughly 5,000 standard water bottles. It takes another 660 gallons to make one cotton t-shirt. So even if you asked ChatGPT a question every single day for the next twelve years, your total water footprint from those questions would still be smaller than eating a single hamburger. One hamburger. A meal that millions of Americans eat every week without pausing to think about the rivers and aquifers that went into the cattle, the feed, the processing, the packaging, and the transportation. The water-bottle comparison is appealing because it feels personal, but drinking water is actually a tiny fraction of a person’s water footprint. Most of the water we consume indirectly goes into the things we eat, wear, and use, not into the glass we fill at the sink. AI, it turns out, is a very small fish in a very large, very thirsty ocean.
Zoom out from individual queries to the industrial scale, and the same pattern holds. According to a 2026 CBS article, U.S. residential lawn irrigation alone uses about 2,900 billion gallons of water every year. Golf course irrigation uses another 531 billion gallons, and that only counts direct on-site use. In comparison, all U.S. data centers in 2023 used a total of 228 billion gallons. That includes everything data centers do—not just AI, but banking, streaming, email, cloud storage, and the endless machinery of the internet. AI specifically accounts for only about 20 percent of that data center water use. In other words, the entire American AI industry, for all its exponential growth and apocalyptic headlines, uses a fraction of the water we pour into keeping suburban lawns green and putting greens immaculate. We are talking about an industry that has generated more fear than almost any other, and it does not even come close to the water footprint of a hobby. And even within that 228 billion gallons, only about 17 billion gallons were consumed directly on-site for data center cooling. The most common method, evaporative cooling, works by letting water absorb heat from servers and evaporate into the air. Yes, that water is lost to the local system, and yes, in a drought-prone region that can be a legitimate concern. But data centers choose water-based cooling over traditional air conditioning for a reason: air conditioning consumes vastly more electricity. The trade-off is usually made deliberately, to minimize the overall environmental impact. That is not an excuse, but it is context. And context is exactly what the viral headlines leave out.
The remaining 211 billion gallons used by American data centers in 2023 were not consumed on site at all. They were used indirectly, mostly to generate the electricity that powers the servers. That water is largely non-consumptive; it is used in power plants, often for cooling, and then returned to its source. It is also not unique to data centers. Every industry that uses electricity is indirectly responsible for water usage at the power plant. But here is the flaw in so many comparisons: indirect water use is counted for AI and data centers, while it is usually ignored for other industries. A hamburger’s water footprint includes the water used to grow the crops to feed the cow, but it typically does not include the water used to generate the electricity for the feedlot, the slaughterhouse, and the refrigeration trucks. When someone compares AI’s total water including electricity to a golf course’s direct water use only, they are comparing apples to oranges. The obsession with AI’s water footprint also pulls our attention away from the risks that actually deserve it. AI is projected to eliminate or transform about 6.1 percent of American jobs by 2030. In 2024, data centers accounted for 4.4 percent of all American electricity consumption, and AI data centers accounted for only about a fifth of that, roughly 0.9 percent of the national total. And then there is the deepest fear of all: in a 2023 survey of nearly 3,000 top AI researchers, the median participant believed there is a 5 percent chance that AI causes human extinction within the next hundred years. Five percent is not a tiny number. It is a profound warning. That is where our energy should go—not into fighting phantoms inflated by unit errors.
At the end of the day, artificial intelligence uses water. There is no point pretending otherwise. But it uses orders of magnitude less water than so many other things we accept without question: corn, beef, nuts, cereal, animal fodder, cotton clothing, laundry, lawns, and golf. If your real goal is protecting the planet, you could spend your life attacking hamburgers, or suburban landscaping, or fast fashion, and never once mention a data center. The misinformation about AI and water is not harmless. Every exaggerated statistic feeds a culture of suspicion in which real environmental problems become harder to solve. It also wastes the limited attention of activists, voters, and legislators who could be making sensible, thoughtful decisions about AI’s enormous and genuinely worrying downsides. Instead of arguing with a number that was off by a factor of one thousand, we should be asking harder questions: What are we going to do about the jobs AI will displace? How do we make sure the electricity powering AI comes from clean sources? And what are we willing to do, right now, to prevent the small but real possibility of catastrophe? Those are the conversations that matter. They may be harder and less viral than a scary statistic about water bottles, but they are the only ones that will actually prepare us for the world we are already building.