A video came across my feed last week. A woman holding up a chart, walking through the water footprint of artificial intelligence. The chart put a thousand AI queries next to a glass of almond milk, an avocado, a beef burger. The burger used up more water by a mile, and she's right. Acknowledging the environmental cost of AI is honest. Singling it out as if it were uniquely guilty is something else.
The objections are coming from employees, college students and the twenty-somethings about to join the workforce.
The issue is not whether AI uses water. It does. It's whether using water consumption as an excuse to hate AI is grounded in truth.
I agreed with the video. Then I kept thinking about it for three days, because of who it was actually speaking to. In every engagement I've run, across every business owner and executive I've sat across from, not one has ever raised water or energy as a reason to avoid AI. Not once. The leaders are not the holdouts. The video was not made for them.
The objection to AI's energy use is real. The way it is being used is not about energy. It is the most defensible reason available to avoid learning something hard, and it is concentrated in exactly the people who will have to do the work.
Let me show you the math first, because the math is where this falls apart.
Nobody citing the water number agrees on the water number
The scary figure you've seen, the one that powers most of the outrage, traces back to a 2023 study from a University of California, Riverside research team. The headline version became "ChatGPT drinks a bottle of water per email." The actual paper estimated roughly 500 milliliters of water for a 100-word response, and a similar order of magnitude for a sustained conversation of ten to fifty exchanges. That number includes the water used to cool the servers and the water used upstream to generate the electricity. It was calculated from less efficient data centers using 2023 proxies.
Now hold that next to what the companies started measuring.
In August 2025, Google published the most detailed per-prompt accounting any major company has released. Their finding for a median text prompt: 0.24 watt-hours of energy and 0.26 milliliters of water. About five drops. They also reported that the energy cost of that median prompt had fallen 33 times in a single year as the models got more efficient. OpenAI's CEO put his company's average query around one-fifteenth of a teaspoon of water.
Sit with the spread. The frightening number and the measured number describe the same activity and differ by a factor of roughly two thousand.
Some of that gap is real apples and oranges. A 100-word email is a longer job than a median prompt, older models were thirstier, and the companies have every incentive to publish flattering figures. Critics correctly note that Google's number leaves out the water embedded in manufacturing the chips, which is not nothing. So the honest answer is a range, not a single clean digit. But the range runs from a few drops to a small fraction of a cup, and it is falling fast as the hardware improves.
That is not the profile of a planetary emergency at the level of a single query. When the most-cited statistic and the best-measured statistic disagree by three orders of magnitude, and almost everyone repeating it picked the scariest end without checking, you are not watching an environmental argument. You are watching a feeling that found a number to hold.
The real problem is aggregate, and it has a zip code
Here is where I refuse to wave the concern away, because the serious version of it is serious.
The per-query footprint is tiny. The aggregate is not. The Department of Energy's Lawrence Berkeley National Laboratory estimated that US data centers consumed around 17.4 billion gallons of water directly for cooling in 2023, plus roughly twelve times that indirectly through the electricity they drew. That figure could double or quadruple by 2028. The Riverside team projects global AI water withdrawal between 4.2 and 6.6 billion cubic meters by 2027, more than the annual withdrawal of several countries the size of Denmark.
And the water is not drawn evenly. It is drawn where the data centers sit, and many of them sit in places that are already dry. Microsoft has reported that a large share of its water consumption comes from regions classified as water-stressed. Google has reported a meaningful slice from high-scarcity areas. In 2024, after a fifteen-year drought, a Chilean environmental court forced Google to pause a planned data center near Santiago over its impact on a local aquifer.
That is a real problem. It is a siting problem, a grid problem, and a disclosure problem. It deserves regulation, transparency requirements, and pressure on the companies to build where the water is and to be honest about what they use.
Notice what kind of problem it is. It is a systems problem, owned by the people building the infrastructure. It is not a problem your staff solves by refusing to learn the tool, any more than you solve industrial agriculture's water footprint by refusing to understand how a tractor works.
And on the scale point, the comparison the video made holds up under scrutiny. A single pound of beef runs around 1,800 gallons of water by the Water Footprint Network's standard. A burger lands somewhere in the hundreds of gallons depending on how you count. One almond costs roughly a gallon on its own. Agriculture accounts for something like 70 percent of global freshwater use and as much as 80 to 90 percent of consumptive freshwater in the United States. A recent infrastructure analysis put one of the largest AI data centers in the world at the annual water footprint of about two and a half hamburger restaurants.
You can hold both of these at once. The aggregate is worth governing. The per-use panic is innumerate. Most people only hold the second one, and only for AI.
The cleanest reason to quit is usually not the real one
So why does AI get the scrutiny that beef, flights, and a closet full of cotton never receive? And why from this group in particular?
Because the environmental objection does something the others don't. It gives a principled-sounding reason to not do a hard thing, and it pays social dividends in the rooms where these objections actually circulate.
Learning to use AI well is uncomfortable. It exposes what you don't know. For a college student, it arrives wrapped in a campus culture that already codes AI as cheating, and stacking "and it's killing the planet" on top turns avoidance into a moral position the whole dorm applauds. For a younger employee, needing the tool can feel like admitting you can't do the job on your own. "I have ethical concerns about the environment" is a far more dignified place to put that discomfort than "I find this hard, and refusing it makes me look principled instead of scared."
Watch how selectively the principle gets applied.
Nobody who refuses to fly tells you they hate airplanes. Nobody skipping the burger says combustion engines are a fad. But "AI is killing the planet" lets a person walk away from the steepest learning curve of their working life and feel righteous on the way out. The objection is not load-bearing. The discomfort is. The objection is the costume the discomfort wears to look respectable in public.
This is not an energy position. It is an off-ramp with an ethics sign bolted to the front.
I want to be precise, because some people raising this are entirely sincere, and the concern itself is legitimate. The tell is not whether someone mentions the environment. The tell is whether they have changed any of their other consumption to the same standard, and whether the concern has made them curious or just made them stop. Real concern looks like questions. The exit-door version looks like a conclusion that arrived suspiciously fast and happens to require nothing further from the person holding it.
Look at who is walking through that door
This is where the cost stops being abstract.
A Harvard Business School meta-analysis pulled together 18 studies covering roughly 143,000 people across 25 countries, much of it students and working adults. It found women about 20 to 25 percent less likely than men to use generative AI. Women make up around 42 percent of ChatGPT users, 31 percent of Claude users, and roughly a quarter of mobile app downloads. The gap holds across nearly every region and occupation, and it persists even when access is equalized.
The driver is not capability. A significant part of it is ethical reservation. The research finds women are more likely to perceive AI use as cheating, more likely to weigh its social and environmental costs, more likely to hesitate on principle. The researchers are careful to say those concerns are not misplaced. The energy demands are real. The labor questions are real. The bias risks are real.
But look at the lever they found.
When researchers raised people's optimism about AI's societal impact, women's usage jumped from 13 percent to 33 percent. Building raw digital skills, by contrast, widened the gap, because it helped men more. Addressing the ethical frame moved adoption more than teaching the buttons did.
Read that against everything above. The environmental objection is an ethical reservation. The population most likely to act on ethical reservations skews younger and female, which is to say it skews toward the exact people most underrepresented in this technology and most central to its next decade. And the thing that actually moves them is not a tutorial. It is a clearer, more honest account of the stakes, the kind the panic actively prevents.
So the innumerate version of the water argument is not harmless. It is one more poorly built on-ramp, and it is turning away the people this field can least afford to lose. The license was never the hard part. The on-ramp is the product, and right now the environmental panic is parked in the middle of it.
What the outrage never gets around to mentioning
Here is the half of the ledger the chart left off entirely.
The same technology being blamed for the climate is being pointed directly at it.
In late 2023, a Google research team published a weather model in Science that outperformed the world's gold-standard physics-based forecasting system on about 90 percent of more than 1,300 tested measures. It predicts the tracks of cyclones and the onset of extreme heat earlier than the old methods, and it runs a full ten-day global forecast in under a minute. The European Centre for Medium-Range Weather Forecasts moved a machine-learning model into operational use in 2024.
The part I find most compelling is access. Running the old supercomputer forecasts costs hundreds of millions of dollars in infrastructure. Running the new model costs a few dollars an hour of cloud time. That means a national weather service in a poorer, more climate-exposed country can now get forecast quality that used to belong only to wealthy nations. In a warming world, earlier warning of a storm is measured in lives.
I'll hold myself to the same standard I'm asking of everyone else, so here is a fair example of the limits. A celebrated 2023 result claimed an AI system had found millions of new stable materials for batteries and solar cells, the equivalent of centuries of lab work. It was real progress, and it is also now contested, with other scientists arguing a meaningful share of the "discoveries" were duplicates. That is exactly how it should work. You make the claim, you show the math, and you let it get tested. The discipline I'm demanding of the people citing water against AI is the same discipline the field owes its own wins.
That is the whole point. Measure it. All of it. The cost and the value, on the same scale, with the same rigor. What nobody gets to do is total one column and call it a verdict.
What this means if you lead anything
You are not the problem here. You are reading this because you already see it. The problem is one layer down and one generation back.
The objection is spreading through the people who have to actually use what you deployed, and through the talent pipeline you hire from. You can buy every license in the building. If your staff won't touch the tool because a video told them it's draining the planet, you did not adopt AI. You paid for it. That gap, between the purchase and the use, is where most organizations quietly lose.
And the new hires are arriving pre-loaded. The graduate you bring on next year may walk in the door with a fully formed, morally framed reason not to use the system you just stood up. That is not a procurement problem. It is a human on-ramp problem, and it now has a generational edge.
So do three things.
Separate the real concern from the costume in your own people by asking what it has changed. Are they helping you build AI responsibly, or only refusing to learn it? The first is a contributor. The second is attrition in disguise.
Hold the standard evenly. If someone on your team raises AI's footprint, put it next to the travel budget and every other line nobody flinches at. Not to win the argument, but to model the honesty you want back.
Treat the ethics frame as the on-ramp it is. The research is blunt. People, especially younger people and women, move when the stakes are explained honestly, not when they are hyped and not when they are terrified. Give them the real numbers, the real limits, and the real upside, and far more of them cross over than any mandate would produce.
The argument was never really about water. It was about whether we are willing to do a hard thing, and whether we will tell the truth about why we won't. And the people most likely to dodge that truth are the ones you most need to bring along.
Hold all of it to the same standard. Including the standard of learning the thing before you decide to reject it.
