Google recently published a paper on the environmental impact of AI, where it argued that the existing measurement methods are too "narrow." The company proposed a "full stack" methodology that includes factors like idle machines and data center overhead, applying it to its production AI model, Gemini.
If you"re in a hurry, here"s what its "comprehensive" measurements found for a median Gemini text prompt:
- It consumes 0.26 mL of water, which is about five drops.
- It uses 0.24 Wh of energy, less than watching 9 seconds of television.
- It generates 0.03 grams of CO2e.
- Google claims it achieved a 44x reduction in total emissions per prompt between May 2024 and May 2025.
But according to The Verge, several experts are skeptical of the numbers Google presented, with Shaolei Ren, an associate professor at UC Riverside who studies AI"s environmental impact, accusing the company of "hiding the critical information." Ren is the co-author of Making AI Less "Thirsty": Uncovering and Addressing the Secret Water Footprint of AI Models, a paper Google cited in its study.
Ren"s paper claimed that training GPT-3 in Microsoft"s state-of-the-art U.S. data centers consumed an estimated 700,000 liters (about 185,000 gallons) of clean freshwater. His work also estimated that when you talk with ChatGPT, a conversation consisting of 20-50 questions and responses could "drink" a 500 ml (17 oz) bottle of water.
Experts like Ren and Alex de Vries-Gao of Digiconomist argue that Google omits indirect water use from its calculations. The company"s figure only includes the water used on-site for cooling its data centers. It completely ignores the off-site water consumed by the power plants that generate the massive amounts of electricity these data centers require. This makes the five-drop figure a "tip of the iceberg" estimate at best, in Vries-Gao"s words.
They also point out that Google used a "market-based" measure of carbon emissions instead of a "location-based" one. Market-based accounting allows a company to subtract the renewable energy credits it purchased, making its footprint appear smaller. The experts say a location-based measurement is more transparent because it reflects the actual carbon intensity of the local power grid supplying the data center, offering a real-world look at the company"s local impact.
In the study"s accompanying blog, Google says it wants to be more transparent about energy use and water consumption. The company"s paper is not yet peer-reviewed, and it declined to answer specific questions from The Verge on the record.