Researchers have found that the training of one large language model — like Meta's Llama 3.1 — would generate as much air pollution as a car driving round trip from New York to Los Angeles 10,000 times. The total cost of AI's health impacts, they found, could reach $20 billion within six years.
The team of researchers from the University of California, Riverside, and the California Institute of Technology conducted what they say may be the first study of its kind assessing AI's impacts on air pollution. The paper, "The Unpaid Toll: Quantifying the Public Health Impact of AI," which is set to be released later Monday, found that the generation of electricity for data centers hosting artificial-intelligence applications could pollute the air so much that by 2030, an additional 1,300 people may die prematurely each year as a result.
That would be a 36% increase over the current annual asthma-related deaths in the country.
The researchers — led by Shaolei Ren of UC Riverside and Caltech's Adam Wierman — examined the release of nitrogen dioxide, sulfur dioxide, and particulate matter with a diameter of 2.5 micrometers or less, which can penetrate deep into the lungs, by power plants and diesel generators associated with AI facilities.
The boom in artificial intelligence has resulted in a spike in electricity demand. McKinsey & Co., a consulting firm, projects that data centers will use 11% to 12% of the total electricity consumed in the US in 2030, up from 3% to 4% last year. While the carbon emissions and water-usage implications of that growth have started to draw scrutiny, the direct health impacts of the air pollution these facilities generate have been mostly ignored.
"There is something like this, air pollution, which is affecting people right now," Ren said in an interview. "We aren't paying attention to it at all."
The researchers estimated that generating electricity for AI data centers could trigger roughly 600,000 asthma-symptom cases a year by 2030.
Last year, the researchers estimated, the generative-AI boom led to a public-health burden of $5.6 billion. By 2030, they calculated, AI's electricity-related public-health costs could top $20 billion. That's more than double the public-health costs of coal-based US steelmaking, they wrote, and would rival the emissions produced by California's 35 million cars.
The cost estimates are based on a risk-assessment tool developed by the Environmental Protection Agency that assigns a dollar figure to what it would take to avoid negative health outcomes, such as premature deaths, asthma symptoms, heart attacks, and missed days of school or work.
For Ren, whose field of study is responsible artificial intelligence, his interest in air quality dates to his childhood. He lived in a small mining town in China until the age of 6, where he saw a correlation between poor air quality and adverse health outcomes in his community, including lung cancer.
The paper is what's called a preprint, a standard practice in computer-science research in which researchers make a paper public before submitting it for peer review.
Diesel generators appear deadly in Virginia
The researchers also examined air pollution driven by emissions from diesel generators used by data centers for backup power and in the manufacturing of the silicon chips used in artificial intelligence.
To examine the impact of diesel generators, the researchers looked at those permitted in Virginia, home to one of the densest collections of data centers in the world. Generators, the paper said, produce 200 to 600 times the nitrogen dioxide per unit of power produced than a natural-gas power plant.
"Diesel generators represent a major source of on-site air pollutants for data centers and pose a significant health risk to the public," the researchers wrote.
Even assuming that emissions by Virginia-permitted generators were just 10% of what the commonwealth's regulations allow, they would cause an additional 13 to 19 deaths each year. If the diesel generators emitted 100% of what is allowed, they would lead to 130 to 190 additional deaths, the researchers found.
The public-health burden of Virginia's data-center generators amounts to $220 million to $300 million a year under the 10% assumption and as much as $3 billion a year under the 100% assumption, the study said.
The researchers found those health effects are not contained to the state, as air pollution travels. They found it's actually a county in Maryland — Montgomery County — that is most affected by Virginia's AI generators.
The effects, according to the EPA tool, would also be felt in West Virginia, New York, New Jersey, Pennsylvania, Delaware, Washington, DC, and as far away as Florida.
"We thought the air pollution was limited to a small area," Ren said. "That's not true. There is actually cross-state air pollution." He and the other researchers found that the harmful effects were felt disproportionately by "economically-disadvantaged communities."
Ren said that understanding the broad dispersion of negative health outcomes could encourage AI companies to alter their site locations or AI training schedules. Health impacts are higher during the day, he said, and some locations have higher or lower health effects.
The researchers also call for greater transparency from the Big Tech companies that lead large-language-model training.
Those companies, including Amazon, Google, Microsoft, and Meta, don't detail the air-pollution impacts of their AI operations in their annual sustainability reports, the researchers wrote.
"If you look at the sustainability reports from these companies, they mention carbon and water, but they don't mention anything about air pollution," Ren said. "They should start reporting this in the same way."
