Google has become one of the first big tech companies to officially share the environmental impact of its AI model, Gemini. In their study, Google claims Gemini consumes .26 milliliters of water and produces 0.03 grams of CO2 emissions per prompt. This is significantly less than what most environmental experts predicted—but do these figures really tell the whole story? 

In an interview with The Verge, two co-authors of one of the papers Google cited in its study argue Google is vastly underestimating Gemini’s water usage. The data centers' new AI models require use a significant amount of electricity. Google’s study doesn’t take into account the water consumed in the power plants these data centers rely on. When French tech company Mistral considered these factors in an environmental study of their own LLM, their estimated water usage per prompt was 50 mL—nearly 200 times Google’s estimate for Gemini.

Without transparency from tech companies on exactly what resources these models are using, it’s difficult, if not impossible, to reduce AI’s environmental impact. Regardless of how accurate Google’s numbers are, hopefully, this study will encourage others in the AI space to be upfront about their own carbon footprint so we can move AI towards a greener future.