Marcel HeadlamWhat One AI Question Costs the Planet, and Where Yours Fits
The average text prompt to Google's Gemini apps uses about five drops of water.
That's Google's own figure, and it sits awkwardly next to a thought you may have had halfway through asking AI for a recipe swap. Didn't I read that this uses a lot of water?
Both can be true.
The case for worrying
AI runs in data centres, which are large buildings full of computers that need electricity to run, and often water to stay cool. In April 2026, the International Energy Agency reported that "electricity demand from data centres soared by 17% in 2025".
It also predicts that "electricity consumption from data centres is set to double by 2030, and power use from those focused on AI is poised to triple" (IEA).
That's a prediction, and predictions change, but this one points to a real problem.
Then there's where the buildings go. Global Action Plan, a UK campaign group, looked at planned data centre sites in Britain. It found "84% of water-intensive data centre developments proposed for areas that are already water stressed or are projected to be water stressed by 2040" (Global Action Plan).
What Global Action Plan means. "Water stressed" means an area where people already use nearly all the water there is, or soon will. Most of the planned data centres that need a lot of water would be built in places like that.
Why it matters here. The count covers sites that are planned but not built yet, and it includes places expected to run short by 2040, not just places that are short today. Where the buildings go matters more than your one question.
Everyone can say "my question is tiny", but when everyone says it, you end up with billions of questions. Using AI less for that reason is like taking shorter showers, and that isn't silly.
The case for asking anyway
In August 2025, Google published its figures. It said the average text prompt to its Gemini apps "consumes 0.26 milliliters (or about five drops) of water". Each one also uses energy "equivalent to watching TV for less than nine seconds" (Google).
Now compare that with something in your home. Waterwise says "dual flush toilets typically use 4-6 litres of water" (Waterwise). Four litres divided by 0.26 millilitres is just over fifteen thousand. So one flush, at the low end, uses as much water as more than fifteen thousand short questions.
Don't treat this number as final. It's one company measuring its own system, and it leaves out the cost of training the AI in the first place. Other companies may be different, and not all of them share their figures.
Where the worry belongs
I don't think these two sides are fighting each other. A short text question is small next to a shower, a flush or an evening of TV. If guilt was stopping you from asking something useful, you can let it go.
The real damage comes from billions of questions, and from where the buildings that answer them are built. Companies, councils and governments make those choices, and that's where your worry can actually do some good.
Ask AI whatever helps you, and don't feel guilty about it. But do speak up about who checks how much water data centres use.
Putting it into practice
Know what kind of AI use you're doing. The five drops are for short text questions, and long chats, images and videos aren't included, so don't assume they cost the same.
Find out what's planned near you. You're allowed to ask, and this prompt can help you start.
Search for any data centres proposed or approved in [your area]. For each one, tell me who makes the planning decision, what the application says about water and electricity, and link the documents you found. Say plainly where you couldn't find something. Don't guess.
Then open the links and read them yourself.
Who decides whether a data centre gets built near you, and have you ever asked them anything?