WASHINGTON — The debate over AI water use is getting a reset after journalist Karen Hao acknowledged a unit mix-up that overstated the water needs of a proposed Google data center in Chile by about 1,000 times, refocusing attention on where AI infrastructure is built and what it discloses. The correction has reignited calls for tech companies and local officials to publish clear, comparable water data before projects are approved, Dec. 22, 2025.
AI water use: what changed after the correction
In a Wired report, Hao said the mistake in her book “Empire of AI” stemmed from misunderstanding units while describing a proposed Google facility near Santiago. She said she is working with her publisher to correct the error, which had circulated widely in online arguments about AI water use.
The fix does not make AI water use a nonissue. It does, however, make the key question harder and more local: How much water will a specific project withdraw, consume and return — and during which months? (Water planners often separate withdrawals from consumption, which is water effectively lost to evaporation.)
Why AI water use is a local issue, not a single number
New work summarized in a Vrije Universiteit Amsterdam release estimates AI systems alone could be associated with roughly 312.5 billion to 764.6 billion liters of water use in 2025 when researchers include both on-site cooling and the water used to generate electricity. De Vries-Gao called the lack of disclosure “highly problematic” and urged “full transparency” at the data-center level.
U.S. policymakers are also trying to pin down baselines. A Congressional Research Service FAQ notes U.S. data centers used about 176 terawatt-hours of electricity in 2023 — about 4.4% of total U.S. consumption — and highlights how water demand can swing sharply depending on cooling design and local conditions.
That variability can turn into a flashpoint. Guardian reporting from the Great Lakes describes residents who say new data center proposals could strain municipal systems even as water levels fluctuate and infrastructure ages.
Transparency push: from NDAs to standardized reporting
A Brookings analysis argues water planning can’t remain an ad hoc, site-specific fight as AI-driven data centers expand. It urges more regional coordination among utilities, economic developers and regulators so communities can compare projects on consistent terms.
The debate has been building for years. A 2023 University of California, Riverside summary helped popularize early estimates of AI water use; a 2024 Washington Post investigation put the issue into household terms with its “bottle of water per email” estimate; and a 2022 Reporters Committee for Freedom of the Press account showed how hard it can be to get water-use records when cities and companies treat them as trade secrets.
What comes next: If AI water use is going to be debated honestly, communities need the basics up front:
Expected withdrawals vs. consumption, plus summer peaks and contingency plans.
The water source (municipal, groundwater or reclaimed water) and where heated wastewater goes.
The cooling design, and what “water-free” marketing means in practice.
A separate tally of the water tied to the facility’s electricity, not just on-site use.
Without that, one wrong number can drown out the harder, location-specific truth — and the places facing the sharpest AI water use pressures will keep finding out last.
