Does every AI email drink a bottle of water?
Where the claim comes from
In September 2024 The Washington Post, working with researchers at UC Riverside, published “A bottle of water per email”. It said a 100-word email written by GPT-4 needs 519 millilitres of water, a little more than a standard bottle. The figure spread everywhere, and it still turns up in arguments about AI.
What’s actually true
- It’s a model, not a measurement. The method comes from the paper Making AI Less “Thirsty”. That paper estimated GPT-3 used about 500 ml of water for every 10 to 50 medium-length responses, depending on where and when the model runs. It counts water evaporated to cool the data centre and water used to generate the electricity it draws.
- Most of that water isn’t in the data centre. Critics who went through the method, including writer Andy Masley in correspondence with the paper’s lead researcher Shaolei Ren, found that most of the bottle is water used by off-site power plants. That includes evaporation from hydroelectric reservoirs. Ren later told the Deseret News that the figure shouldn’t be treated as an accurate measure of today’s AI. In his words: “We cannot just use a number from two years ago to describe today’s system.”
- Newer measurements are far lower. In August 2025 Google published its own figures: a median Gemini Apps text prompt used 0.26 ml of water and 0.24 Wh of energy, “about five drops”. In June 2025 OpenAI’s Sam Altman wrote that an average ChatGPT query uses about 0.000085 gallons, roughly a fifteenth of a teaspoon (about 0.3 ml). Both are company-reported figures, not independent audits.
- The total is still real. Billions of prompts a day add up. Training runs and new data centres also draw heavily on local water and power, especially in hot, dry places like the Gulf.
The truth
The truth: The bottle came from stacking worst-case assumptions about a 2020-era system, then running the result through a headline. Each prompt is closer to a few drops. The real story is the total: how many data centres get built, where, and what power and cooling they use. Ask about that, not about your email.
Sources
- The Washington Post: A bottle of water per email (Sept 2024)
- Li, Yang, Islam & Ren: Making AI Less “Thirsty” (arXiv 2304.03271)
- Deseret News via KSL: Fear of data centers outpaces knowledge about them (Ren and Masley quotes)
- Google Cloud: Measuring the environmental impact of AI inference (Aug 2025) · technical paper, arXiv 2508.15734
- Sam Altman: The Gentle Singularity (June 2025)
