IT processing requests for artificial intelligence, or AI, result in increasing levels of fatal atmospheric pollution from power plants and diesel generators which permanently provide electricity to the increasing number of computer processing centers.
This air pollution should lead to 1,300 premature deaths per year by 2030 in the United States, and its public health costs linked to cancer, asthma and other diseases, as well as At missed work and school days, $ 20 billion a year are approaching. .
These are the conclusions of a study By UC riverside and Caltech scientists published online this week. However, these human and financial costs seem neglected by the technological industry.
“If you look at these reports on sustainable development written by technological companies, they focus only on carbon emissions, and some of them also include water, but there is absolutely no mention of air pollutants harmful to health and these pollutants already create a burden on public health, “said Shaolei RenAssociate Professor UCR of electrical and computer engineering and corresponding author of the study.
The authors, including a Caltech teacher and a computer scientist Adam WiermanRecommend the adoption of standards and methods forcing technological companies to report air pollution caused by their energy consumption and their emergency generators.
They also recommend that the hardest communities affected by air pollution due to the production of electricity of data processing centers are correctly compensated by technological companies for the health burden.
The authors also found that air pollution resulting from AI disproportionately affects certain low -income communities, partly because of their proximity to power plants or rescue generators in data processing centers. In addition, pollution crosses the boundaries of counties and states, creating impacts on the health of communities around the world, said Ren.
“The data centers pay local land taxes to the county where they operate,” said Ren. “But this impact on health is not limited to a small community. In fact, it crosses the whole country, so these other places are not at all paid.
For example, pollution from rescue generators of Northern Virginia data centers is propagated to Maryland, Western Virginia, Pennsylvania, New York, New Jersey, Delaware and the Columbia District, causing costs of Regional public health of around 190 to 260 million dollars per year. . If these rescue generators emit at their maximum authorized level, the annual cost will be multiplied by 10 and will reach 1.9 to 2.6 billion dollars.
In some regions, the public health cost associated with AI processing centers exceeds what technological companies pay for electricity, according to the study.
While technological companies rush to provide AI services that reshape our way of working and playing, air pollution that results in the form of fine particles penetrating into the lungs – those of size less than 2 , 5 micrometers – and other pollutants regulated by the federal government, such as nitrogen oxides, is expected. to increase sharply. According to the study, the burden of public health by 2030 should be double that of the American steel industry and compete with that of all California cars, bus and trucks.
“AI growth leads to a huge increase in demand for data and energy centers, making it the fastest growth energy consumption sector in all industries,” said Ren.
For example, Ren and his colleagues calculated the programs linked to the formation of a large language
Model, or LLM, on the LLAMA-3.1 scale of META, an advanced LLM with open weight published by the owner of Facebook in July to compete with the main proprietary models such as the OPENAI GPT-4. The study revealed that the production of electricity to cause this model produced air pollution equivalent to more than 10,000 round trips by car between Los Angeles and New York.
The authors estimate health costs, including premature deaths, using statistical methods developed by the American environmental protection agency, which take into account the known epidemiological risks associated with air pollution from power plants and rescue diesel generators. The 1,300 annual deaths expected by 2030 are in the middle of a range between 940 and 1,590.
“If family members suffer from asthma or other health problems, air pollution from these data centers may affect them at the moment. This is a public health problem to which we have to attack ourselves urgently, “said Ren.
The newspaper title is “The unpaid assessment: to quantify the impact of the AI on public health. »» In addition to Ren and Wierman, the authors of the article are also Yuelin Han, Zhifeng Wu and Pengfei Li, all three of the Bourn’s College of Engineering of the UCR.
The article follows the previous revolutionary research of the REN team which revealed the IA water consumption imprint.