'The Unpaid Toll: Quantifying the Public Health Impact of AI' byYuelin Han, Zhifeng Wu, Pengfei Li, Adam Wierman and Shaolei Ren comments
The rise of artificial intelligence (AI) has numerous potentials to play a transformative role in addressing grand societal challenges, including air quality and public health [1, 2]. For example, by integrating multimodal data from various sources, AI can provide effective tools and actionable insights for pandemic preparedness, disease prevention, healthcare optimization, and air quality management [1, 3]. However, the surging demand for AI — particularly generative AI, as exemplified by the recent popularity of large language models (LLMs) — has driven a rapid increase in computational needs, fueling the unprecedented expansion of energy-intensive AI data centers. According to McKinsey projections, under a medium-growth scenario [4], the U.S. data centers are anticipated to account for 11.7% of national electricity consumption in 2030, a substantial increase from their current share of less than 4% in 2023.
The growing electricity demand of AI data centers has not only created significant stress on power grid stability [5,6], but also increasingly impacts the environment through escalating carbon emissions [7,8] and water consumption [9]. These environmental impacts are driven primarily by the “expansion of AI products and services,” as recently acknowledged by Google in its latest sustainability report [10]. To mitigate the challenges posed to both power grids and the environment, a range of strategies have been explored, including grid-integrated data centers [6, 11], energy-efficient hardware and software [12–14], and the adoption of carbon-aware and water-efficient computing practices [9,15–17], among others.
The hidden toll of AI. While the environmental footprint of AI has garnered attention, the public health burden, a hidden toll of AI, has been largely overlooked. Across its entire lifecycle — from chip manufacturing to data center operation — AI contributes substantially to air quality degradation and public health costs through the emission of various criteria air pollutants. These include fine particulate matter (PM2.5, particles measuring 2.5 micrometers or smaller in diameter that can penetrate deep into lungs and cause serious health effects), sulfur dioxide (SO2), and nitrogen dioxide (NO2). Concretely, the AI hardware manufacturing process [18], electricity generation from fossil fuels to power AI data centers, and the maintenance and usage of diesel backup generators to ensure continuous AI data center operation all produce significant amounts of criteria air pollutants. Moreover, the distinct spatial-temporal heterogeneities of emission sources suggest that focusing solely on reducing AI’s carbon footprints may not minimize its emissions of criteria air pollutants or the resulting public health impacts (Section 5).
Exposure to criteria air pollutants is directly and causally linked to various adverse health outcomes, including premature mortality, lung cancer, asthma, heart attacks, cardiovascular diseases, strokes, and even cognitive decline, especially for the elderly and vulnerable individuals with pre-existing conditions [20–23]. Moreover, even short-term (hours to days) PM2.5 exposure is harmful and deadly, accounting for approximately 1 million premature deaths per year from 2000 to 2019 and representing 2% of total global deaths [24].
Globally, 4.2 million deaths were attributed to ambient (i.e., outdoor) air pollution in 2019 [25]. Air pollution has become the second highest risk factor for noncommunicable diseases [26]. Notably, according to the latest Global Burden of Disease report [27], along with high blood pressure and high blood sugar, ambient particulate matter is placed among the leading risk factors for disease burden globally in every socio-demographic group.
While the U.S. has generally better air quality than many other countries, 4 in 10 people in the U.S. still live with unhealthy levels of air pollution, according to the “State of the Air 2024” report published by the American Lung Association [28]. In 2019 (the latest year of data provided by the World Health Organization, or WHO, as of November 2024), an estimate of 93,886 deaths in the U.S. were attributed to ambient air pollution [29]. In fact, even compliance with the U.S. Environmental Protection Agency (EPA) air quality standards does not necessarily guarantee healthy air that meets the WHO guidelines. Concretely, the EPA’s recently tightened primary standard for PM2.5 sets an annual average limit of 9 µg/m3, considerably higher than the WHO’s recommended level of 5 µg/m3 [30,31]. In addition, the EPA projects that 53 U.S. counties, including 23 in the most populous state of California, would fail to meet the revised national annual PM2.5 standard in 2032 [32].
Further, criteria air pollutants are not confined to the immediate vicinity of their emission sources; they can travel hundreds of miles through a dispersion process (i.e., cross-state air pollution) [33,34], impacting public health across vast regions — pollutants from the 2024 Canadian wildfires significantly degraded air quality across much of the U.S. and reached as far as Mexico and Europe [35].
Importantly, along with transportation and industrial activities, electricity generation is a major contributor to ambient air pollution with substantial public health impacts [26, 36, 37]. For example, a recent study [38] shows that, between 1999 and 2020, a total of 460,000 excess deaths were attributed to PM2.5 generated by coal-fired power plants alone in the U.S. As highlighted by the U.S. EPA [36], despite years of progress, “fossil fuel-based power plants remain a leading source of air, water, and land pollution that affects communities nationwide.” Moreover, according to the U.S. Energy Information Administration (EIA) projection [39], the coal consumption by the electricity sector in 2050 will still be about 30% of the 2024 level in the baseline reference case, and the number will exceed 50% in the high zero-carbon technology cost case. Indeed, the growing energy demands of AI are already delaying the decommissioning of coal-fired power plants and increasing fossil-fuel plants in the U.S. as well as around the world [6,40,41].
The public health outcomes of AI due to its emission of criteria air pollutants lead to various losses, such as hospitalizations, medication usage, emergency room visits, school loss days, and lost workdays. Moreover, these losses can be further quantified in economic costs based on epidemiology and economics research for the corresponding health endpoints [22,42]. In contrast, the environmental impacts of AI, e.g., carbon emission from fossil fuels and water consumption for data center cooling, often do not cause the same immediate health impacts. For instance, while anthropogenic carbon emissions could also pose risks to public health, such impacts are often second- or third-order effects through long-term climate change which can then threaten the human well-being by affecting the food people eat and facilitating the spreading of pests, among others [43]. Nonetheless, despite their immediate and tangible impacts on public health, the criteria air pollutants of AI have remained under the radar, entirely omitted from today’s AI risk assessments and sustainability reports [10,44,45].
Quantifying the public health costs of AI. In this paper, we uncover and quantify the hidden public health impacts of AI. We introduce a general methodology to model the emission of criteria air pollutants associated with AI tasks across three distinct scopes: emissions from the maintenance and operation of backup generators (Scope 1), emissions from fossil fuel combustion for electricity generation (Scope 2), and emissions resulting from the manufacturing of server hardware (Scope 3). Then, we analyze the dispersion of criteria air pollutants and the resulting public health impacts across different regions.
