AI
One cow-flesh burger uses as much water as about 1.4 million AI prompts
By WhyStopHarmingAnimals.com
You would need about 1.4 million AI prompts to use as much water as one quarter-pound cow-flesh burger[1][2]. It takes two published figures to get there. OpenAI’s Sam Altman said an average ChatGPT query uses about 0.32 milliliters of water[1]. A University of Michigan life-cycle assessment by Heller and Keoleian found that a quarter-pound cow-flesh patty in the U.S. takes 437.7 liters of consumptive water[2]. Divide the patty’s water by the prompt’s and you get about 1.4 million[1][2].
Other estimates range from 8,800 to 5.5 million prompts
Depending on whose figures you use, one burger comes out to anywhere from about 8,800 to about 5.5 million prompts[1][3]. The low end comes from Li and colleagues, whose 2023 paper “Making AI Less Thirsty” estimated that GPT-3 used about a 500 milliliter bottle of water for roughly 10 to 50 medium-length replies, depending on when and where it ran[3]. So each reply used about 10 to 50 milliliters[3]. Measured against the Michigan patty, a burger works out to about 8,800 to 44,000 prompts[2][3].
The high end comes from counting more of the water that goes into cow flesh. The 437.7 liter figure only counts blue water, the water taken from rivers, lakes and wells[2]. It leaves out green water, the rain that falls on pasture and feed crops, so the full amount is higher. Mekonnen and Hoekstra’s global water footprint study, published in Ecosystems in 2012 and summarized by the Water Footprint Network[5], counts green, blue and grey water together and puts cow flesh at about 15,400 liters per kilogram[4]. That makes a quarter-pound patty about 1,750 liters, or around 460 gallons[4]. Against OpenAI’s 0.32 milliliters, that’s about 5.5 million prompts[1][4]. Many people assume burger figures are inflated by counting rain, but the 1.4 million figure leaves it out[2][4].
The bottle-per-email figure isn’t an average prompt
A 2024 Washington Post article said ChatGPT used about 519 milliliters of water, a little more than a bottle, to write a 100-word email[6]. The number spread widely. But it modeled GPT-4 writing a whole email, which is neither an average prompt nor the figure OpenAI later published[1]. The research behind the headlines, Li and colleagues’ 2023 paper, put a medium-length GPT-3 reply at about 10 to 50 milliliters[3], and the headlines turned that into a bottle per email. The 1.4 million figure uses OpenAI’s published average and the Michigan study[1][2], and the other studies give the range[1][3].
The plant burger used 99% less water
The same Michigan study compared a quarter-pound Beyond Burger with a quarter-pound U.S. cow-flesh burger[2]. The plant patty used 99% less water, 93% less land, 90% fewer greenhouse gases and 46% less energy[2][7]. In absolute terms, the Beyond Burger used 3.27 liters of consumptive water and the cow-flesh burger 437.7 liters[2]. Weighted for water scarcity, it was 1.1 liter-equivalents against 218.4[2]. Beyond Meat publicizes the study[7][9], and the comparison is one company’s burger against average U.S. cow flesh[2]. Still, the gap is in line with the other water figures for cow flesh.
Someone would have to use a chatbot for years to match the water in one burger[1][2]. Meanwhile, tens of billions of farm animals are fed and watered every day whether or not anyone opens a chat window, and a vegan burger that does the same job is already on the shelf.
Sources
[1]Altman, average query water
The Verge / OpenAI
[2]Heller & Keoleian, UMichigan Beyond Burger LCA
University of Michigan CSS, 2018
[3]Li et al., Making AI Less Thirsty, 2023
arXiv, 2023
[4]Mekonnen & Hoekstra, Ecosystems 2012
Ecosystems, 2012
[5]Water Footprint Network, farm animal products
Water Footprint Network
[6]Washington Post, bottle per email, 2024
Washington Post, 2024
[7]Beyond Meat, A Burger With Benefits
Beyond Meat, 2018
[8]Heller & Keoleian, UMichigan LCA PDF
University of Michigan CSS, 2018
[9]Beyond Meat release
Beyond Meat