Does a single AI query use a bottle of water?

Episode Summary

Executive Summary: The episode examines claims about how much water AI chatbots consume per prompt, tracing them to research on data-center cooling and electricity generation. It finds the popular “small bottle per query” claim is an oversimplification: estimates vary widely by model, query type, and what water sources are counted, with most use often occurring off-site in power generation rather than directly in data centers.

Main Topics: Origins of the AI water-use claim (Priority: 5/5): Paul Connolly explains why listeners are asking about AI’s environmental cost and introduces the “small bottle of water” claim as the central number under scrutiny. How the original research estimated water use (Priority: 5/5): Dr. Shaolei Ren describes the 2023 study 'Making AI Less Thirsty,' which inferred water consumption from limited public data on GPT-3 and Microsoft/OpenAI infrastructure. Direct vs indirect water consumption (Priority: 5/5): The episode distinguishes water used on-site for data-center cooling from off-site water evaporated while generating electricity, noting that most of the total can come from power stations. Why GPT-4 estimates exploded (Priority: 4/5): A Washington Post extrapolation from Ren’s method to GPT-4 produced about the same total water use as the GPT-3 estimate, but for one query instead of many, highlighting the uncertainty of closed models. Limits of single-number claims (Priority: 5/5): Experts argue that a universal 'average query' figure is misleading because AI prompts vary widely in length, complexity, and model size; a range would be more meaningful than one number. Emerging transparency from AI companies (Priority: 3/5): The episode ends with Mistral releasing a per-response water-use figure for its chatbot, suggesting the industry may gradually disclose more environmental data.

Key Arguments: The viral claim that one AI prompt uses a small bottle of water is not straightforwardly true; it depends on model, location, and what water is counted. Ren’s 2023 study used extrapolation because firms like OpenAI do not disclose enough internal data, so the result is necessarily approximate. Most of the reported water footprint is indirect—evaporated in electricity production—rather than only the cooling water inside data centers. GPT-4 is more powerful and less transparent than GPT-3, making its water-use estimates more uncertain and harder to verify. A single average figure for all prompts is misleading because AI usage ranges from tiny text queries to long, complex, multi-step tasks. Early corporate disclosures such as Mistral’s may help, but they still need independent fact-checking and clearer methodology.

Data Points: Water use claim: about a small bottle of water per prompt - The widely circulated claim the episode investigates Study title: Making AI Less Thirsty - 2023 paper by Dr. Shaolei Ren and colleagues Water consumption estimate for medium-sized language model queries: 500 milliliters - Ren’s estimate for roughly 10 to 50 queries to equal this amount Average number of questions per 500 ml: around 30 queries - Estimated for ChatGPT-like systems using US data centers Share of water use that is off-site: 87% - For an AI query going to an average US data center, most water use is in electricity generation Washington Post GPT-4 estimate: 519 milliliters - Applied Ren’s method to GPT-4 for one 100-word email query Altman’s average-query water figure: 0.000085 gallons - Sam Altman’s blog-post claim about average query water use Altman’s figure in teaspoons: roughly one-fifteenth of a teaspoon - The podcast translates Altman’s gallons figure into a more intuitive measure Mistral typical response water use: 45 milliliters per response - For a 200–300 word response from Le Chat Mistral responses for a small bottle: 10 responses - Quick calculation showing about a small bottle’s worth of water

Pivotal Quotes: "We want to bring a more complete picture of the environmental impacts of AI to the public." — Dr. Shaolei Ren: Explaining the motivation behind his 2023 water-impact study "I think it doesn't make sense to give a single number for all queries." — Sasha Luciani: Critiquing simplistic claims about an “average” AI prompt "I think the true number is somewhere between Altman's number and the thirsty AI number." — Sasha Luciani: Summing up the uncertainty around AI water-use estimates

Implications: Listeners should treat AI water-use headlines cautiously: the footprint varies by model and query, and much of it may occur outside data centers. For industry, better disclosure and standardized metrics are needed to make environmental claims comparable.