Your Clothes Use More Water Than Your AI

I buy almost everything I own pre-owned. My furniture, my clothes, my kitchenware — almost everything in my home was somebody else’s first. Between the secondhand clothes and the pre-owned furniture, this choice keeps roughly a ton of CO² and more than 5,000 gallons of water out of manufacturing supply chains every year compared to buying new.¹

I’m also a GenAI director at a Fortune 100 bank (where I wear my thrifted office clothes) and an erstwhile writer about AI on Medium, where I’ve argued using AI for personal emotional support is dangerous. I have real concerns about AI as the technology develops. But AI’s environmental impact is not one of them.

Many disagree with me. Headlines decrying AI as the final blow to the integrity of the Earth’s environment are a dime a dozen. In March 2026, Bernie Sanders and Alexandria Ocasio-Cortez introduced the Artificial Intelligence Data Center Moratorium Act over these concerns.²

The numbers behind this so-called environmental crisis, however, are simply not that concerning.

How AI Actually Uses Water

The most common headlines feature concerns about AI water use. But the AI-water headlines is that nobody agrees on how much water AI actually uses.

The case rests on a few key points.

• Between 2020 and 2023, as AI workloads scaled up, Microsoft’s water consumption rose 87% to 1.7 billion gallons, Google’s rose 69% to 6.4 billion gallons, and Meta’s rose 40% to 813 million gallons.
• Researchers at UC Riverside, led by Shaolei Ren, published estimates that a single ChatGPT conversation uses about 500 milliliters of water, roughly a bottle’s worth, when you account for both direct cooling and the water consumed to generate electricity.
• Morgan Stanley projects AI data centers could consume over a trillion liters of water per year by 2028.
• More than 230 environmental organizations signed a joint statement calling for a moratorium on new data center construction, and Greenpeace warned that data center electricity demand could be 11 times higher in 2030 than it was in 2023.

Industry supporters, of course, tell a different story.

• Sam Altman says a ChatGPT query uses about a fifteenth of a teaspoon of water.
• Google says a Gemini query uses about five drops and produces 0.03 grams of CO², roughly equivalent to a standard Google search. Efficiency is improving fast: Google reported a 33-fold reduction in per-query energy between 2024 and 2025.
• Hannah Ritchie, the Oxford researcher and author of Not the End of the World, calculated that a year of daily ChatGPT use produces about 11 kilograms of CO², equivalent to driving 27 miles in an average car, or less than one-tenth of one percent of the average American’s annual carbon footprint.

Hank Green explains these differing stories well in his video “I Know Why Lying About AI Water Use Is So Easy.” Both sides are misrepresenting the data, he says. Altman’s “fifteenth of a teaspoon” excludes training, electricity generation, and chip manufacturing, which makes it misleading by omission. Critics, meanwhile, inflate their numbers by including thermoelectric water withdrawals (most of which is harmlessly returned to its source).

There’s also an engineering trade-off that most coverage ignores. Data centers can cool their servers with evaporative systems, which consume a lot of water but relatively little electricity, or with dry cooling, which uses almost no water but draws significantly more power. Choosing “waterless” cooling doesn’t eliminate the environmental cost; it shifts it from water to electricity. And where a data center sits matters at least as much as how it’s cooled. A facility drawing municipal freshwater in the Arizona desert has a fundamentally different impact than one near a river in Oregon, even if both process the same number of queries.

Water gets all the press, but electricity is a bigger concern by the numbers. The International Energy Agency projects that global data center electricity consumption will reach roughly 1,100 terawatt-hours by the end of 2026, equivalent to Japan’s entire national consumption.

That sounds alarming, and it would be, except that what determines the carbon footprint of that electricity isn’t the volume but the source. GPT-3’s training on a standard US grid produced roughly 500 metric tons of CO². BLOOM, a model of comparable size, emitted about 50 metric tons because it trained on France’s nuclear-powered grid. Same scale of computation, one-tenth the carbon. The difference was where the electricity came from.

The Environmental Costs We Don’t Question

Meanwhile, a number of largely uncontroversial things continue to savage the American environmental footprint:

• Livestock accounts for 14.5% of all human-caused greenhouse gas emissions, and producing a single pound of beef requires roughly 1,800 gallons of water, almost all of it for irrigating feed crops. The U.S. livestock sector alone consumes about 200 billion gallons of water per day, roughly 5,000 times ChatGPT’s daily water consumption.
• Transportation is the largest single source of U.S. greenhouse gas emissions at 28%, and light-duty trucks (pickups, SUVs, and minivans) account for 37% of that. They now make up nearly 73% of all new vehicle sales. A full-size pickup driven 13,500 miles a year at 20 miles per gallon produces about 6 metric tons of CO², roughly 2.5 times what a 50-mpg hybrid would emit over the same distance. All this so they can drive big trucks and hit children with them.³
• The fast fashion industry produces roughly 10% of global carbon emissions, more than international flights and maritime shipping combined, and consumes 93 billion cubic meters of water per year.

AI uses significantly less energy and water than either the beef industry or the transportation sector.⁴ If someone wants to tell me I should feel guilty about my environmental footprint, I know exactly where to start — and it isn’t with my Claude subscription.

What AI Does With That Energy

We all talk about AI like it’s all good or all bad, but that’s like asking whether “the internet” is good or bad. AI has been responsible for a lot of good things:

And at the same time, AI is also responsible for a lot of bad things. Ninety-eight percent of all deepfake videos online are non-consensual pornography, some of which depicts children. The FBI documented more than 22,000 AI-related fraud complaints in 2025, with losses totaling roughly $893 million. And then there’s the dead teenagers. But the environmental case against AI doesn’t distinguish between a data center folding proteins to find cancer treatments and a data center generating abusive material. A moratorium shuts down both.

The historian Melvin Kranzberg put it well in 1985 when he wrote that “technology is neither good nor bad; nor is it neutral.” The same technology produces radically different outcomes depending on what it’s pointed at, and evaluating it requires looking at specific applications, not aggregate electricity bills.

The Electronic Frontier Foundation made this argument explicitly in 2026, warning that regulators who fixate on use cases they don’t like without considering beneficial ones cause “enormous collateral harm.”

AI is not the world’s first dual-use technology. Another recent one was nuclear weapons. And in the end, we didn’t ban nuclear physics because it could produce weapons; we built governance structures to separate weapons from energy and medicine.

Why AI Gets the Moratorium Bill

If the environmental case is this weak, why is AI the one getting a moratorium bill? I have a theory, and the polling data backs it up.

When Pew Research asks Americans what worries them about AI, the top concerns are misinformation (66%), job displacement (56%), and loss of human connection (57%). Environmental impact doesn’t make the unprompted list. It only shows up when pollsters specifically ask about it.

And when they do, the results are revealing. An EPIC/AP-NORC poll found that 41% of Americans describe themselves as very or extremely concerned about AI’s environmental impact, compared to only 29% for meat production and 23% for air travel. People rate the environmental threat of AI as more serious than the environmental threat of beef, even though beef’s footprint is vastly larger by every available measure.

There’s also a political dimension. AI doesn’t yet have an established constituency. Nobody’s identity is wrapped up in defending data centers or puts “I love my GPU cluster” bumper stickers on their car. That makes AI a politically costless target, and I suspect the environmental framing merely exists to give deeper, harder-to-articulate anxieties about AI (job displacement, existential fear, creative threat) a more concrete and respectable-sounding wrapper.

The environmental impact of AI is not zero and we should not ignore it. Companies should site data centers in water-rich regions, run them on clean energy, and publish transparent environmental reports. But a moratorium that would freeze cancer diagnostics, protein science, drug discovery, climate modeling, and accessibility tools alongside everything else, based on a footprint that’s a fraction of what we happily spend on big cars and beef cattle? That’s another matter.

If you genuinely want to reduce your environmental impact, I have some suggestions that will go a lot further than deleting ChatGPT. Eat less red meat. Drive a smaller vehicle. Buy your stuff pre-owned. Any one of those choices will dwarf your AI usage by orders of magnitude.