The moral high ground has a golf course on it.

First, a clarification: "AI" isn't one technology — it's a marketing umbrella. ChatGPT and Claude are AI, but so are Netflix and Spotify recommendations, autocorrect, Grammarly, Shazam, and game matchmaking. The person who says "I don't use AI" almost certainly does every day — through tools that just don't advertise it. The debate is already running on a false premise.

Of all those charges, the one with the most heat — and the most confusion — is water: "every prompt drinks a bottle," "data centers are draining rivers." It sounds damning, until you make the one distinction almost nobody does. Water comes in three kinds. Green water is rain that fell on land and got used by plants — never pulled from a river, lake, or aquifer, and it falls whether or not anything's growing there. Blue water is water actually withdrawn — irrigation, drinking water, industrial processing. Grey water is the freshwater needed to dilute the pollution something creates.

Start with that burger. The global average water footprint of beef is about 15,400 liters per kilogram — but roughly 94% of it is green water (rain on pasture and feed crops). That distinction is the whole game: count only blue water — the fair fight — and beef drops to about 550 liters per kg. Worldwide, that still adds up to roughly 30× all the world's AI (on the order of 7–9 trillion gallons a year against AI's ~264 billion). That viral "beef uses 80× more water than AI" figure? It pits U.S. beef — rain included — against all of global AI: apples to oranges twice over. Counted honestly, like for like, beef still wins — just by about 30×, not 80×.

Now, about that latte. Almonds don't get a green-water pass — California's orchards sit in a near-desert and run almost entirely on irrigation. A gallon of almond milk takes roughly 84 gallons of real, withdrawn water, so one cup is about 5 gallons. Compare that to AI's least flattering credible number: UC Riverside researchers put a 100-word AI answer at about 500 mL — a full water bottle, electricity included. That one cup of almond milk still outdrinks about 35 of them — and that's giving AI the worst-case figure. Same kind of water, no asterisk.

AI runs on withdrawn water too — none of it rain — so it's comparable to almond milk in a way it never was to beef. The real question isn't "AI vs. beef." It's: of everything made from withdrawn water, where does AI actually rank?

Here's the same comparison at full scale — three things that are all genuinely blue water, all genuinely huge, and only one of them is "technology":

This is the comparison that survives scrutiny. Almonds and golf are both genuinely withdrawn water — no green-water asterisk — and neither is "technology," so nobody can twist this into "ban computers." It's just: water goes mostly to food and recreation, at a scale that makes AI's current footprint look almost rounding-error small. (And yes — that's where this issue's headline comes from. The moral high ground really does have a golf course on it.)

But "almost" is doing real work in that sentence — because the curve is bending up fast:

This is already happening. Take Joliet, Illinois: the state Water Survey flagged its aquifer back in the 1970s, and a 2018 projection said the wells would be dry by 2030. A $1.5 billion pipeline to Lake Michigan finally broke ground this year — finishing in 2030, the same year the aquifer runs out. Then, in March 2026, the city approved the largest data center in Illinois, sited right on top of it.

The pattern repeats. In Virginia — the data center capital of the world — a bill letting communities weigh data-center water impacts was vetoed; in 80% of its data-center towns, NDAs hide the terms residents live under. Tucson's council voted unanimously against a $3.6 billion Amazon data center — and it kept moving anyway.

The anger is legitimate — it's just aimed at the wrong target. The ask shouldn't be "no AI"; it's "no freshwater withdrawals in drought zones, no NDAs, full disclosure of water contracts." That's winnable. "Ban AI" isn't, and it protects no one's water.

And the freshwater part is solvable. Microsoft's newest data centers use closed-loop, "zero-water" cooling — filled once, then recirculated — cutting a facility's yearly draw to about what a single restaurant uses. China just switched on a wind-powered data center on the seabed off Shanghai that cools itself with seawater. The technology to take AI off the freshwater grid already exists; what's missing is the requirement to use it.

Which is more or less what the moral authority everyone keeps quoting actually said. Pope Leo XIV's Magnifica humanitas (May 2026) gets cited as a condemnation of AI. What he actually wrote:

He presented it alongside Anthropic co-founder Christopher Olah. Not a ban — a call for accountability from the people building this, not abstinence from the people using it.

The friend with the latte isn't wrong to care about AI's water use — they're just not thinking about their own cup. Concern without context makes people feel guilty about the wrong things and lets the real decision-makers off the hook. And it only works if it's rigorous: rainfall and withdrawal aren't interchangeable, and every comparison here holds up because both sides are the same kind of water — real, withdrawn, no asterisk.

That's the move this whole issue turns on: not the loudest number, but the one that survives a fact-check.