Skip to content

The Problems with AI: Water, Power and Who's Really Using It

Every time you type a question into a chatbot, a building full of computers answers it. That building uses electricity, water and land. The people who live next to it pay for some of that.

This post covers the problems with AI that don’t show up in the chat window: what it costs to run, who uses it, and who pays. It is not a “quit AI” post. You can use a tool and still think clearly about what it takes to run.

The building behind the chat window

In this post: the real numbers, why it needs so much, does your use matter, who uses AI, the workers, the copyright question, what happens to neighbors, FAQ.

The short version

Data centers used more electricity than all but 10 countries in 2025. Two-thirds of the data centers built or planned since 2022 sit in places that already struggle to get enough water. Now to be clear, cooling methods that use little to no water definitely exist. But setting up water cooling is cheaper and easier. And we all know that companies have never been particularly interested in the long-term effects of anything. Behind the chat windows workers in other countries are often labeling content for under $2 an hour (per a 2023 report) and millions of creative works have been used for training without asking. Now one short question costs very little. The system that answers it costs a lot and other people often get the bill.

What these words mean

A few words come up all post. Here they are once, so the rest reads easily.

Data center. A building packed with computers that answer your questions and store the internet’s data. They run all day, every day.

GPU. A chip built for the kind of math AI needs. Thousands sit in racks inside a data center, and they get very hot.

Training. The learning phase, where a model reads billions of pages of text. It’s expensive, but it happens once per version of the model.

Everyday use (the technical word is inference). Every time someone asks the model something and it answers. This is where most of AI’s energy goes.

LLM (large language model). The engine inside a chatbot that reads and writes text. It runs billions of calculations to produce each token (a small chunk of a word), one at a time. Full explainer: what an LLM is vs what AI means.

Watt-hour. How electricity is counted, the way gallons count water. A 10-watt LED bulb left on for an hour uses 10 watt-hours. A trillion is a million million.

How much water and power does AI use?

A United Nations University report from June 2026 compared data centers to countries. The numbers below come from the AP’s coverage of it, not the full report.

  • In 2025, data centers worldwide used 448 trillion watt-hours of electricity. That’s more than all but 10 countries.
  • Making that electricity released roughly as much carbon dioxide (the gas that traps heat in the air) as Argentina does, about 189 million metric tons.
  • Making that electricity also took about 1.2 trillion gallons of water. That doesn’t count the water used to cool the buildings, because the report didn’t measure it.
  • By 2030, data centers could use about 3% of the world’s electricity, nearly triple what Pakistan, Bangladesh and Nigeria use together. Only five countries would use more.
  • About 20% of data center energy goes to AI now. The report expects 40% by 2030.

Location matters too. Two-thirds of the data centers built or planned since 2022 are in water-stressed areas. Water-stressed means the area already doesn’t have enough water for its own needs. However these places are often rural and poor and do not have the resources to fight the companies that are building these data centers.

On the company side, Amazon reported that its data centers took in 2.5 billion gallons in 2025, its first full water number. That figure comes from a statistics roundup, not from Amazon’s own report. The same roundup says Microsoft declared itself “water positive” in June 2026, meaning it put back more water than it took out worldwide in fiscal 2025. A worldwide total can’t tell you what happened in any one town.

Why does it need so much?

Two reasons: the math and the cooling.

The math. Each token (a small chunk of a word) a chatbot’s LLM writes takes billions of calculations. They run on GPUs, and GPUs run hot. Sort of like a warehouse full of space heaters left on around the clock.

The cooling. To keep the chips from failing, data centers have to carry the heat away. The most common way is evaporative cooling, where water soaks up the heat and then evaporates. It costs significantly less than the alternatives, and it uses a lot of water. One modern AI data center can draw as much power as 100,000 homes.

Water cooling is a choice. Other methods exist:

  • Immersion cooling. Servers sit in a special liquid in a closed loop. It uses no water.
  • Direct liquid cooling. Coolant runs through pipes right to the chips. One industry guide rates its water use as moderate.

So why not use them? Well we all know the answer to that question. Cost. One industry guide says immersion costs 30 to 50% more to install than plain air cooling, and direct liquid costs 15 to 25% more. Evaporative cooling is the cheaper option, and the neighbors cover the difference in water. Sort of like a factory skipping the treatment system because the river is free. So while a lot of these companies are making huge profits, they also continue to take direct shortcuts that hurt not only our planet but the people in it.

Training vs. everyday use. Training takes a lot, but it happens once. Roughly 80 to 90% of AI’s total energy goes to everyday use, meaning people asking questions.

Does your use matter?

Yes, a little. Mostly no. Both are true.

The per-question numbers. You’ve probably seen that one chatbot question uses 10 times the electricity of a web search. That number is shaky, because it leans on a 2009 Google figure. Newer measurements put a quick text question at about 1.5 to 3 times a search.

In real-life terms: Google says a typical text question to its chatbot, Gemini, uses about as much electricity as watching TV for less than nine seconds, plus about five drops of water. That’s Google measuring itself, so treat it as a ballpark. Longer chats, images and video use more.

Scale is the real story. One popular chatbot handles about 2.5 billion questions a day. At that same ballpark, the electricity adds up to roughly what 25,000 U.S. homes use, all day, every day. That’s our own math, using 0.3 watt-hours per short question and the average U.S. home’s 10,791 kilowatt-hours a year. It assumes every question is a short text one.

A pattern worth knowing. In 1971, a famous ad from Keep America Beautiful showed a man in a canoe on polluted water, a tear rolling down his face as someone tossed trash from a car. The line: “People start pollution. People can stop it.” It made millions of Americans feel personally guilty about litter.

What the ad didn’t say: Keep America Beautiful was started in 1953 by the companies making the cans and bottles, including American Can and Owens-Illinois. Instead of changing their packaging, they backed a campaign that made litter your failing. They also lobbied against bottle bills, laws that would have brought back returnable containers. The man in the ad, Iron Eyes Cody, was an Italian American actor, not Native American.

The same framing shows up with AI. “Does my chatbot question hurt the planet?” puts the weight on the person typing. It moves attention off the companies that picked the cheapest cooling, built in water-stressed places, and often don’t say how much they use. Your one question is a drop. Their building decisions are the river.

Now that is not to say that you as a person cannot use this technology efficiently. Most searches do not need to go through AI. We’ve had Google for many years and I feel like running a search through AI is sort of like driving a tank down the street when a car would do just fine.

Who actually uses AI?

Pew Research Center surveyed 5,119 U.S. adults in February 2026:

  • 49% have used an AI chatbot, up from 33% in 2024 (the 2024 question was worded a little differently).
  • About 4 in 10 adults (42%) use chatbots to search for information. 38% of employed adults use them for tasks at work.

Use is up, but trust isn’t. A Gallup and Bentley University poll from May 2026 (3,270 U.S. adults) found 39% say AI does more harm than good, up from 31% a year earlier. Only 9% say it does more good. And 79% expect it to reduce the number of U.S. jobs.

The sharpest shift is among 18 to 29 year olds: 47% said AI does more harm than good, up 11 points in a year.

You can use a tool and still think it’s causing harm. Plenty of people do. Hell I’m writing a whole blog about it.

The people behind the curtain

Before a chatbot can learn to steer away from violent or explicit content, someone has to look at that content and label it by hand. This work is called data labeling. Reports say much of it went to workers in Kenya and other low-wage countries.

A 2023 TIME investigation found Kenyan workers labeling data for a major AI company were paid under $2 an hour. The OECD’s AI incident monitor (a tracker run by an international policy group) logs these conditions as an AI-related harm and describes psychological harm, pay disputes and little mental health support. We read the OECD’s summary of the TIME article, not the article itself.

The pay figure dates from 2023, so treat it as a past snapshot, not today’s rate.

Was it built on your work?

AI companies trained their models on text and images collected from the public internet: books, articles, blog posts, photos, artwork, code. Most of it was used without asking the people who made it or paying them.

Training data is the material a model reads to learn. For these models, it was a big chunk of the internet.

Authors, artists, news organizations and music publishers have sued companies including OpenAI, Meta, Anthropic and Google. The companies argue fair use, a legal idea that allows limited use of copyrighted work when the result is different enough from the original. Courts haven’t settled it. But I think we can all agree that’s a bit of bullshit. This is one of the reasons why I am particularly against generative AI. But again that’s another article.

Here is where the courts stood on October 31, 2026. In June 2025, a California judge ruled that training on books could count as fair use, but that keeping millions of pirated copies was a separate problem. Two days later, a second judge sided with Meta but said training “in many circumstances” would not qualify as fair use. Anthropic agreed to a $1.5 billion settlement with authors in December 2025. In September 2026, an appeals court ruled against a company that trained a legal research tool (not a chatbot) and left generative AI “for another day”. The results are mixed, and the chatbot question is still open. For you: if you’ve published captions, blog posts, photos or fanfiction, there’s a chance they were in the mix.

What happens next door

  • Water. A mid-sized data center can use up to 300,000 gallons a day. A large one can use 5 million, about what a town of 50,000 people uses.
  • Land. The average site in 2024 covered about 224 acres, up 144% since 2022. Communities in Ohio and Wisconsin have pushed back on farmland proposals, and one 300-acre rezoning in Ohio was withdrawn.
  • Jobs. Even the largest data centers employ fewer than 150 permanent workers. Some employ as few as 25.
  • Who lives nearby. Nearly half of U.S. data centers sit in areas with above-median environmental burdens and social vulnerability. In South Carolina, new data centers and the power plants built for them are often sited near Black and low-income communities.

Communities do have a say. The World Resources Institute (an independent research group) reports more than $64 billion in data center projects were delayed or canceled between May 2024 and March 2025 because of organized local opposition. That’s one reason the missing numbers matter: it’s hard to push back on a number nobody will share.

How you think about this is up to you

New to AI tools? Now you know what sits behind the chat window. That doesn’t mean stop. It means you’re choosing with your eyes open.

Use them every day? Shorter, clearer prompts use less energy. Check your privacy settings while you’re at it, and see how Claude and ChatGPT compare on privacy.

Does this make you uncomfortable? That’s a reasonable response. This post isn’t here to tell you what to do about it.

If one part of this surprised you, tell me which.

FAQ

Does AI waste water?

Data centers (buildings packed with computers that answer your questions and store the internet’s data) use billions of gallons of water, mostly to keep the chips from overheating. The most common method, evaporative cooling, lets water soak up the heat and then evaporate. Options that use little or no water cost more to build, so whether this counts as waste is your call.

Is AI bad for the environment?

Making the electricity for data centers produced about as much carbon dioxide (the gas that traps heat in the air) in 2025 as all of Argentina. Your one question is a tiny slice of that. The bigger weight is how many buildings are running and how they’re powered.

Is using a chatbot bad for the planet?

One short text question uses about as much electricity as watching TV for less than nine seconds, going by Google’s own numbers for its chatbot. Billions of questions a day add up, and longer chats, images and video use more. Shorter prompts use less, but the biggest choices belong to the companies running the data centers.

Was AI trained on copyrighted work?

Yes. Companies collected books, articles, photos and other work from the internet to teach their chatbots (that’s called training), mostly without asking or paying. Authors and artists have sued, and the companies say it’s fair use, a legal idea that allows limited use of someone’s work without permission. As of October 31, 2026, courts have ruled both ways.

Do AI companies exploit workers?

Reports say workers in Kenya were paid under $2 an hour to label violent and explicit content so chatbots could learn to filter it. That job is called data labeling, and workers have described psychological harm and little mental health support. The pay figure comes from a 2023 investigation, so it’s a past snapshot, not today’s rate.

Sources