A particular kind of conversation happens whenever a new technology becomes culturally visible. Someone reads something troubling, shares it, and within days a general consensus forms among one group that the thing is bad and among another that it is being unfairly targeted. The positions form faster than the arguments, which means most people arrive at their opinions through a kind of social triangulation: they notice which side the people they trust have taken and update accordingly. The behavior is normal and possibly adaptive, but it creates a specific vulnerability. A position can be correct while its reasons are wrong, and that matters far more than the people holding the position usually acknowledge.
The water use debate around AI works as a useful stress test for this. Water is not the most important environmental concern raised by large language models, but it is one where the numbers have been made unusually legible, and where the sources of confusion are followable once explained.
In late 2024 and into 2025, a claim circulated in both directions. Sam Altman stated that a ChatGPT query uses roughly 0.000085 gallons of water, about a fifteenth of a teaspoon. Separately, a Morgan Stanley projection estimated that AI data centers could consume close to a trillion liters of water annually by 2028. Both figures were accurate in their own framing. The two simply describe different things using the same word.
The teaspoon figure counts only the water consumed during inference, meaning the moment of exchange between a user and the model. That figure is small and accurate. The trillion-liter figure attempts to describe something closer to the full resource footprint of the industry: training runs that continue while current models are in use, the buildout of cooling infrastructure across hundreds of data centers, and in some analyses, the water passing through thermoelectric power plants supplying electricity. Both figures count something real, but they are not counting the same thing.
Hank Green made this point in his dissection of the debate, and it deserves more credit than it typically receives: choosing which water to count is not a neutral technical decision. The selection is rhetorical. To make the number look small, count only the query. To make it look large, include industrial water flowing through power plants, much of which returns to its source and is never consumed the way municipal water evaporates off a cooling tower. Both approaches cite methodology, and neither is complete.
What kind of water matters is also part of the question. Municipal water (which runs through treatment plants, gets delivered through city infrastructure, and competes directly with households) is not the same as water drawn from a river to run a turbine cycle and returned at a higher temperature. Both are forms of use, and neither is trivial. Collapsing them into one figure because they share a unit of measurement obscures the actual question: whether a specific use in a specific place is creating strain that would not otherwise exist.
Where the argument tends to go wrong
AI discourse tends to go sideways here. What most people take away from a headline about water use is not a set of conditional figures dependent on scope and methodology. They take away a simple directional claim: AI uses a lot of water, or AI uses very little. The directional claim then gets folded into a pre-existing position about whether AI is good or bad, extractive or efficient, something to be concerned about or something being unfairly targeted.
The problem is not just inaccuracy. A position built on a simplified version cannot be revised when the underlying facts become clearer. If you are skeptical of AI's environmental footprint because you believe every query consumes meaningful amounts of drinking water, your skepticism will survive any correction to that specific figure, because the actual concern is not the teaspoon. The concern is something closer to a general sense that AI companies are not being honest about their costs, which may well be legitimate but is not what was argued. When someone corrects you with the accurate figure, they have not actually persuaded you of anything; they have only given you a reason to think they are operating in bad faith.
The same pattern runs the other direction. Someone who accepts the teaspoon figure at face value and uses it to dismiss environmental concerns about AI has not engaged with training footprints, or with the geographic specificity of where data centers are being sited, or with power demand, which is probably a larger concern than water given its share of existing infrastructure and its implications for both carbon budgets and energy costs. The person has only learned to cite a small number quickly, which is not the same as having a position.
What the honest version looks like
Green's video resists the temptation to resolve the question cleanly, somewhat unusually for popular science content. It does not conclude that AI water use is bad or acceptable. The conclusion is that water is not one thing, that use is not one thing, and that anyone giving you a single clean number about either is making choices they are not disclosing. The conclusion is more uncomfortable than a verdict, but it is probably the honest one.
The most clarifying reframe in the discussion is geographic. A data center drawing from a watershed that is already fully allocated creates a different kind of problem than one built in a region with genuine surplus. A facility using non-potable reclaimed water for cooling is not competing with the same systems as one drawing municipal supply. These distinctions do not make the concern disappear; they make it more precise, which is what good environmental reasoning requires. The question is whether a specific use, in a specific context, creates strain that would not otherwise exist, and whether that strain is being managed honestly.
What this means for how you think about AI
For anyone trying to figure out what to believe about artificial intelligence as a factual question about its costs and effects (rather than as a policy question or a corporate-behavior question), the water debate is a useful model. The easy version of the concern is not the accurate version, and the dismissal of the easy version is also not. Critics and advocates both have access to figures that confirm what they already believe. The question worth asking is what scope the cited figures are drawing, and whether that scope is honest given what the person citing them is trying to argue.
This matters particularly now because the terms of the AI debate are still being established. The categories people use to evaluate it (is it wasteful, is it useful, does it eliminate jobs, does it create them) are forming faster than the evidence can support, and positions that form fast tend to calcify. Build your opinion slowly enough that you know which part of it would have to change if the evidence shifted.
Whether AI turns out to be worth its investment, resource cost, and social disruption is genuinely uncertain. The honest answer is that the people building it do not know, the people opposing it do not know, and the projections on both sides are serving interests that are not yours. What you can control is whether your own reasoning about it is clear enough that you would notice if the picture changed.
That is not a small thing to ask, but it is more useful than a clean verdict.