Google released a study claiming its Gemini AI uses just five drops of water per text prompt, but environmental experts are calling the research misleading for omitting critical indirect water consumption from power generation. The controversy highlights growing scrutiny over AI's true environmental cost as data centers expand globally.
Google just sparked a fierce backlash from environmental researchers over its new study claiming Gemini AI uses barely five drops of water per text prompt. The tech giant's research, released amid mounting pressure over AI's environmental toll, puts water consumption at just 0.26 milliliters and energy use equivalent to watching TV for nine seconds. But leading experts are calling it a masterclass in corporate obfuscation.
"They're just hiding the critical information," Shaolei Ren, an associate professor at UC Riverside who studies AI's environmental impact, told The Verge. "This really spreads the wrong message to the world."
The controversy centers on what Google left out. The company's figures only account for direct water cooling in data centers, completely ignoring the massive indirect water consumption from electricity generation. According to Bloomberg analysis, the majority of a data center's water footprint actually comes from power generation - water that's used in gas and nuclear plants to create steam for turbines and cool reactors.
"You only see the tip of the iceberg, basically," says Alex de Vries-Gao of Digiconomist and PhD candidate at Vrije Universiteit Amsterdam, who has extensively studied data center energy consumption. When accounting for total water usage including power generation, previous research has found AI prompts can consume up to 50ml of water - nearly 200 times Google's estimate.
Google's methodology choices are equally problematic, experts argue. The study uses only "market-based" carbon emissions accounting, which factors in Google's renewable energy purchases, while ignoring "location-based" emissions that reflect the actual environmental impact where data centers operate. This approach follows corporate preferences rather than that call for reporting both metrics.
