The 'Gallons Per Query' AI Water Myth, Examined

You will have seen the claim somewhere on social media in the last year: every time you ask ChatGPT a question, it drinks a bottle of water. Some versions of the figure put it at 10 gallons per image. The figure is wrong. Like many wrong things about AI, it is wrong in an interesting way, and the truth still leaves room for legitimate concern.

Where the number actually lands

According to research compiled by Understanding Your AI and a recent analysis by Andy Masley, a typical chatbot query uses somewhere between 15 and 60 millilitres of water at the data center cooling stage. That is a shot glass, give or take. The ten-gallons-per-image figure that has gone viral is not supported by credible measurements. It appears to have spread from a misreading of a research paper that was itself describing full training-run and lifecycle estimates, not per-query usage.

OpenAI's CEO Sam Altman called the social-media versions of the claim “completely untrue” and “totally insane” at a February 2026 summit appearance. He is not exactly a neutral party, but the underlying figures match what independent measurement studies have reported, including the original arXiv paper that started a lot of this conversation in the first place.

What the honest picture actually is

The whole global data center sector, including streaming, banking, gaming, and AI inference together, accounts for less than 1% of global water withdrawals by most credible estimates. AI is a fast-growing slice of that 1%, but it is not the dominant slice yet, and the per-query numbers are small.

That does not mean the environmental story is fine. Two things can be true at once. The viral per-query water figure is fake, and AI infrastructure is still drawing on water and energy in specific regions where local impact matters. Communities in Phoenix and The Dalles, Oregon, among others, have pushed back against data center expansion, and the concern there is local water stress rather than the per-query maths.

Why the myth keeps spreading

It spreads because it is sticky. A vivid figure (“a bottle of water for every email!”) travels further than a calibrated one (“around a shot glass at the cooling stage”). It also spreads because the people most worried about AI's footprint are right to be worried about something. The instinct to look for the environmental cost is correct. The number itself is just wrong.

The pattern is familiar. We covered the related chatbot accuracy myth recently, where a confident-sounding answer from a model tricks users into trusting it more than the evidence warrants. The water claim works the same way, in reverse. A confident-sounding figure trips trust the other direction, into excessive alarm. Both errors come from the same source. People do not have time to check.

What to do? When you see the next viral AI environmental statistic, look for two things. Where did the original measurement come from, and what is it actually measuring? If the answer is a per-query figure derived from a training-run paper, treat it the way you would treat a financial statistic that took the lifetime profit of a company and divided it by one transaction. Mathematically valid. Practically useless.