Among the fierce debate on fees for the environment of artificial intelligence, Google published a novel one test This says that his assistant AI Gemini only consumes the minimum amount of water and energy for each text hint. But experts say that the claims of the technological giant are misleading.
Google estimates that the median of Gemini’s text uses about five drops of water or 0.26 milliliters and the same number of electricity as watching TV for less than nine seconds, about 0.24 Watt hour (WH), which produces about 0.03 grams of carbon double emissions.
Google estimates are lower than previous studies on intensive data centers in water and energy that are subject to generative AI models. This is partly due to the performance improvements that the company has made over the past year. But Google also omitted key data points in his study, which led to an incomplete understanding of Gemini’s environmental impact, experts say The Verge.
“They just hide critical information.”
“They just hide critical information,” says Shaolei Ren, professor of electrical and computer engineering at the University of California in Riverside. “This really spreads the wrong message.” Ren studied water consumption AND Air pollution Related to AI and is one of the authors of the paper, which Google mentions in the Gemini study.
The main problem, which indicated that Google omits the indirect water consumption in its estimates. His study included water that data centers employ in cooling systems to stop the servers from overheating. These cooling systems caused worries Over the years, how data centers can tighten water deficiencies in regions susceptible to drought. Now the attention changes how much more May need electricity data centers To accommodate novel AI models. The growing demand for electricity caused a number of novel construction plans gas and nuclear power plants, which also consume water in your own cooling systems AND To turn the turbines with steam. In fact, Most of the water center water consumes due to the use of electricity – which Google overlook in this study.
As a result, along with the respect of Google water: “You only see the tip of the iceberg,” says Alex de Vries-Gao, founder of the website Digiconomist and a doctoral student at VRIJE Universiteit Amsterdam Institute for Environmental Studies, who examined the demand for the energy of data centers used for cryptomying and artificial intelligence.
Google skipped another vital record when it comes to energy consumption and pollution. The article only divides “on the market“A measure of carbon dioxide emissions, which takes into account the obligations that the company undertakes an raise in renewable energy in energy networks.
It would also be a more holistic approach to embracing “Location“The measure of carbon dioxide emissions, which is considering the impact of the data center, wherever it works, taking into account the current mix of spotless and soiled energy of the local energy network. The emissions based on location are usually higher than market emissions, and offers a greater view of the local environment.” This is ground grounding “, Ren says that both Ren and de Vries Google should include the location, on the international side, observing the international, based on international impact, and then based on the location of the Metric company, observe International. Standards determined by the greenhouse gases protocol.
Google’s article cites previous research conducted by Rhenium AND De Vries-Gao And he claims that he can ensure a more correct representation of environmental impact than other modeling tests that do not have the first page data. But Ren and de Vries-Gao say Google compares apples to Oran. Previous work was based on medium, not on the median, from which Google uses, as well as the defects of Ren Google for not sharing numbers (the number of words or tokens under text prompts) for how the median happened. The company writes that it bases its estimates on the median of prompt to prevent the protruding value, which consume very more energy from the results of the distortion.
“You only see only the tip of the iceberg.”
When it comes to calculating water consumption, Google claims that its determination of 0.26 ml of water per text is “rows of size smaller than previous estimates”, which have achieved up to 50 ml in REN. This is a misleading comparison, he claims that again, because the paper co -author takes into account the total direct and indirect water consumption in the data center.
Google has not yet sent a novel article to review, although spokeswoman Mara Harris told We -Mail that it was open in the future. The company refused to answer the list of other questions The Verge. But the study and accompanying blogs say that Google wants to be more lucid in the field of water consumption, energy consumption and carbon dioxide emissions in AI Chatbot and offer more normalized parameters, how to measure environmental impact. The company claims that it goes further than previous research, taking into account the energy used by sterile machines and supporting infrastructure in the data center, such as cooling systems.
“Although we are proud of innovations related to our previous performance benefits, we commit ourselves to further improvement in the coming years,” says Amin Vahat, vice president of AI and infrastructure for Google Cloud and Jeff Dean, the main scientist of Google Deepmind and Google Research blog.
Google claims that it has significantly improved the energy efficiency of the Gemini text monitor between May 2024 and May 2025, reaching 33 times reduction of electricity consumption to a hint. The company claims that the Swift Median carbon trace has fallen 44x in the same period. These profits also explain why Google estimates are now much lower than the research from previous years.
However, enlarge and the real image is more gloomy. Profits from performance can continue to lead to greater pollution and generally employ resources – unfortunate Phenomenon known as the Jevons paradox. The so -called “carbon emissions” Google increased by 11 percent and 51 percent since 2019 last year, because the company still aggressively implements AI Sustainable development report. (The report also notes that Google began to exclude some categories of greenhouse gas emissions from climate goals this year, which in his opinion is “peripheral” or outside the company’s direct control).
“If you look at the total numbers that Google publishes, it’s really bad,” says De Vries-Gao. As for the estimates that were spent on Gemini today, “it does not tell the full story.”
48 Comments