In the ongoing conversation about the environmental impact of artificial intelligence, Google has published a new report asserting that its Gemini AI assistant consumes only a negligible amount of water and energy for each text input. However, industry experts have raised concerns over the accuracy of these claims.
According to Google, a typical Gemini text prompt consumes approximately five drops of water (0.26 milliliters) and the same amount of electricity that would power a television for less than nine seconds (about 0.24 watt-hours), generating an estimated 0.03 grams of carbon dioxide emissions.
These figures are significantly lower than those in previous studies regarding the resource-intensive data centers that support generative AI models. Although Google has made strides in efficiency over the last year, experts suggest that crucial information was omitted from the study, leaving a gap in understanding Gemini’s total environmental footprint, as noted by sources within Technology News.
“They’re just hiding the critical information.”
Shaolei Ren, an associate professor at the University of California, Riverside, emphasizes that important indirect water usage is missing from Google’s claims. While the report includes water utilized for cooling systems in data centers, it does not address how these facilities could worsen water shortages in regions facing drought. Furthermore, the increasing electricity demand from new AI models raises concerns about the necessity for additional energy sources, including gas and nuclear power plants, which also draw on water resources for cooling.
Experts highlight that Google’s estimate offers only a partial view of the issue. “You only see the tip of the iceberg, basically,” says Alex de Vries-Gao, founder of Digiconomist and a PhD candidate at Vrije Universiteit Amsterdam. His research focused on energy consumption in data centers connected to cryptomining and AI.
Google’s report also presents a limited view of its carbon footprint. It references a “market-based” carbon emission metric that accounts for the commitment to renewable energy. However, a more comprehensive assessment would include a “location-based” measure that reflects the actual energy mix in the specific locations where data centers operate, providing a clearer picture of their environmental impact.
This is crucial for understanding their local effects, according to both Ren and de Vries-Gao, who argue that Google should adhere to the standards established by the Greenhouse Gas Protocol and include this information in its assessment.
Google’s report references prior studies by Ren and de Vries-Gao, claiming it reflects a more precise environmental impact than others relying on average estimates. However, Ren argues that comparing averages to Google’s median method presents an “apples-to-oranges” situation, expressing concern that necessary details, such as data on the length of text prompts, were not shared in the report.
“This is not telling the complete story.”
Google’s assertion that a text prompt requires just 0.26ml of water is starkly lower than estimates from Ren’s research, which could reach as high as 50ml. Ren attributes this discrepancy to Google’s focus on direct water consumption without accounting for the broader context of water use at data centers.
The technology company has yet to submit its findings for peer review, though a spokesperson mentioned willingness for future scrutiny. While refusing to answer several inquiries from Technology News, the company stated its aim to enhance transparency regarding its AI chatbot’s environmental impacts and to establish well-defined metrics for evaluating such effects. Google maintains that its calculations surpass past studies by factoring in energy utilized by inactive machines alongside cooling systems.
“We are proud of our efficiency improvements and are committed to making significant advances in the future,” remarked Amin Vahdat, VP/GM of AI & Infrastructure for Google Cloud, alongside Jeff Dean, chief scientist of Google DeepMind and Google Research, in a blog post.
Between May 2024 and May 2025, Google claims it achieved a 33-fold reduction in electricity consumption per Gemini prompt, with a reported 44-fold decrease in carbon footprint for a median prompt. These advancements partly explain why Google’s current estimates are considerably lower than earlier studies.
However, a broader perspective reveals a more troubling scenario. Enhanced efficiencies can still lead to greater overall resource consumption and pollution, a concept encapsulated by what is known as Jevons paradox. Google’s reported “ambitions-based carbon emissions” have risen by 11 percent in the last year and 51 percent since 2019 as the company intensifies its focus on AI, according to their latest sustainability report. Notably, this report indicates that Google has excluded specific categories of greenhouse gas emissions from its climate objectives, categorizing them as “peripheral” or beyond the company’s direct influence.
De Vries-Gao expresses concern, stating, “If you look at the total numbers that Google is posting, it’s actually really bad. This situation concerning Gemini estimates does not portray the full story.”
