The Specimen
An essay about AI and water that spends more words taking its headline back than making it — and why that is the most honest thing about it
First in a series on how debates about scarce resources get told as stories about personal guilt.
You have probably met the number by now. Writing a single hundred-word email with a chatbot uses about a bottle of water. It is a good number. It travels well, survives paraphrase, fits in a headline, and lands with a small satisfying click of guilt — the sensation of having been shown the hidden cost of a thing you did this morning without thinking. It has been repeated enough that most people file it, as one does with a fact heard three times, under true.
The strange thing about the article that gave the number its most recent life is that the article itself does not believe it. Or rather: it prints its assertion in the largest type on the page and then spends most of its length explaining, carefully and correctly, why you should not take that figure literally. The number is not a fixed property of a chatbot reply, it tells you. It changes with the model, the data center, the local weather, the cooling system, the electricity source, and the method of accounting. The same prompt has a different footprint depending on where and when it is answered. All of this is right. It is also a quiet retraction of the sentence at the top of the page.
That gap — between what a headline promises and what its own body is willing to defend — is the subject of this essay, and the reason it is the first in a series. The article is not a villain. It is a specimen. Held up to a light, it shows a mechanism by which a debate about a shared physical resource gets converted into a story about individual consumption, and it shows that mechanism unusually clearly, because this particular article is more honest than most and the seams are therefore visible.
Let's start with the headline, because everything is in the headline. It makes three claims at once: that a hundred-word email costs roughly a bottle of water; that global AI could consume something like half the United Kingdom's annual water withdrawal by 2027; and that much of that water is being drawn from places already in drought.
Next, watch what the body of the article does with each.
The bottle it takes back, in the terms above. The half-a-United-Kingdom figure it also walks toward the exit: that number comes from a 2023 study modeling a range of possible 2027 withdrawals — 4.2 to 6.6 billion cubic meters — under a set of scenarios its own authors describe as scenarios, not a forecast and certainly not a measurement of anything happening now. The article, to its credit, says so. Which leaves the third clause, the one about drought — and this is the claim the body actually stands behind, develops, and is right about. A liter of water taken from a wet basin is not the same as a liter pulled from a stressed aquifer in a dry July. Water, unlike carbon, is local. Where and when you take it is most of what matters.
So the article's strongest, best-supported argument is the one printed smallest — demoted to the tail of the headline and softened with a hedge, much of — while its weakest, most-equivocal number rides in sixty-point type at the top. The rhetoric runs exactly opposite to the evidence. The thing the piece can defend is buried; the thing it cannot defend is the hook. That inversion is not a flaw peculiar to one blog post. It is the house style of an entire genre, and once you have seen it you will see it everywhere in this series.
It is worth slowing down on why the numbers disagree, because the disagreement is not a scandal and pretending it is one would be its own kind of dishonesty.
Set the bottle beside its opposite. A 2025 study authored by Google reported that the median text prompt to its Gemini system used about 0.26 milliliters of water — five drops — under the company's own production accounting. Five drops and one bottle are not close. But they are not a lie and a truth; they are two answers to two different questions. They involve different models, different years, different workloads, and, decisively, different boundaries. Some estimates count only the water evaporated on site to cool the servers. Others add the water consumed off site to generate the electricity those servers draw — the indirect footprint hidden inside the power line. Some count every liter withdrawn from a source; others count only the fraction consumed, removed from immediate reuse. A figure can be technically sound and still be incomparable to the one beside it, because the two were built to measure different things.
This is the discipline the rest of this series will try to keep, so it is only fair to keep it here: the bottle is a framing of a 2024 estimate for one model writing one email at an average American data center, and its full calculation — which model, which location, which cooling, direct or indirect water, withdrawal or consumption, one reply or a whole editing session — was never reconstructed, so it should be read as a memorable illustration and not a conversion rate. The 2027 UK water withdrawal figure is a modeled scenario. The five drops is a vendor's number under a vendor's boundary. None of these is the answer. Each is an answer to a question someone chose. And the choosing of the question is where the politics lives, long before any number appears.
With all that we've said, we must emphasize that the article is good. Against the reflexes of its genre, it names the difference between water withdrawn and water consumed. It notes that even a major newspaper had to correct itself after paraphrasing a "500 milliliters per ten-to-fifty responses" estimate as though it applied to every single prompt — a small slide, quietly made, exactly the kind this essay is about, and the article flags it rather than hiding it. It refuses the tidy moral. It does not tell you that the answer is to write fewer emails; it says, more or less, that individual restraint cannot substitute for knowing where the water is coming from. On the merits, it is a responsible piece of work.
And that is precisely why it makes such a clean specimen. The useful question about a piece like this is not whether it means well — plainly it does — but what it does regardless of what it means. What the article intends is to inform you. What it produces is a small transferable guilt, aimed at your own hands, and a headline that will outlive every correction printed beneath it. The distance between the intention and the product is not the author's private failing. It is the thing this whole series is about, and it will turn out to matter more than anyone's sincerity, here and in every essay after.
Because the honest version of the story has no headline. It depends on where, and when, and how it's cooled, and it's probably falling per query, and locally it can matter a great deal is the true sentence, and it will not fit at the top of a page and will not be repeated three times until it files under true. So the bottle gets reached for anyway — even here, even by an author who knows better and says so four paragraphs down. The memorable simplification outlives its own caveats. That is not a moral failing of one writer. It is a tax the whole genre pays for needing something a reader can hold.
Which raises the question this series exists to ask. Why does the number that survives always seem to be the one that lands on you — your email, your bottle, your life choices — rather than the one that lands on the system that decided where the servers went?
Because a per-you number does a particular kind of work. It takes a question about infrastructure — who gets to draw how much from which source, on what terms, with whose permission — and quietly reissues it as a question about your personal budget. It moves the debate from the aquifer to the conscience. And a conscience is a much easier thing to address than a water district, because a conscience has no lawyers and asks for nothing back.
That is what the bottle is: not a lie, but a decoy. And if it is a decoy, the useful move is to ask what it draws the eye away from. The article's own best paragraph tells you, and then leaves it lying there. The thing that actually matters is not how much water a use consumes in the abstract. It is where the water is, when it is taken, whether it was going to be scarce there anyway, and who else was drinking from the same source. Those are not questions about consumption. They are questions about allocation — about who was granted a claim on a shared resource, and who was not. And allocation, unlike an almond or an email or a bottle, has no checkout aisle. There is nothing for a guilty individual to refuse.
The quietest party in the whole argument is the one with the largest stake in it: the stressed basin itself, the household downstream, the person who will need that aquifer in twenty years. None of them has a marketing budget, so none of them sets the volume of the debate — but the deeper trouble is that none of them has a channel into it at all. There is no seat where they sit, no price that carries their need, no vote that registers it. The argument would proceed in exactly the same way whether they were in the room or not, and that — more than any number — is the surest sign of whom it was built to be about. The bottle has a marketing budget; it is doing free advertising for the idea that this is a problem about you. The basin has none.
So this is where the series begins, and where it is going. The bottle made the resource visible, and that was a real service; the interface really does hide the pipes, and it is worth being reminded that text on screens (this one included) runs on chips and cooling towers and somebody's water. But visibility is not the same as understanding, and a number that makes you feel the cost is not the same as a number that tells you who is paying it. From here the trail runs downward — first into water and dirt, into a crop that took the guilt while a thirstier one drank in peace, and eventually into the grid, the substation, and the line item on a stranger's power bill. The method will stay the same each time, and it is the method this essay has tried to model on its own numbers: name the year, name the place, name what was actually measured, notice who chose the question — and, when a thing plainly means one thing and does another, learn to trust what it does.
Follow the water. But keep an eye on the well.