The hidden costs of AI

[MUSIC] Welcome to Science Sessions, the podcasts of the proceedings of the National Academy of Sciences, where we connect you with Academy members, researchers, and policymakers. Join us as we explore the stories behind the science, high Matthew Hardcastle. In recent years, the development of artificial intelligence, including large language models and generative AI has captured the attention of the public and industry alike. However, the purported benefits of AI come with a cost that is often invisible to consumers. In this special episode, we will hear from computer scientists and engineers studying the impact of AI on sustainability. A considerable amount of data and computing power are required to train large language models. Providing answers to users in real time requires a construction of data centers, which can rival the energy demands of entire towns. By some estimates, querying the large language model can use ten times more energy than a traditional internet search engine. Generating images and videos is even more energy intensive. In areas without reliable access to renewable energy sources, much of this energy demand is provided by fossil fuels, producing significant carbon emissions. Bronis Dessepinski is the chief technology officer at Livermore Computing, the Supercomputing Center of Lawrence Livermore National Laboratory in California. Dessepinski describes how the energy demands that AI data centers have increased in recent years. He also explains why improvements in data center energy efficiency will not necessarily reduce total energy expenditures. The AI data centers are all talking about being one gigawatt are larger centers, which is 100 times more power than we were using for the largest supercomputers 20 years ago. Even Dwarfs would have really widely considered the fastest system on the planet today. A gigawatt is enough to power a city of 3 million people. A big factor is how are you going to generate that power? What we hear is a lot of the plants are using generators that they're deploying right there, either natural gas or even worse diesel. Energy efficiency as a metric has improved by roughly a factor of 25. Let's say in the last 20 years, as we improve energy efficiency, you might not easily think that that's going to drive down the amount of energy that we use. But in fact, it tends to actually drive things in the opposite direction as you get your work done more efficiently. You get it done in less time. So now you've finished one piece of work and you have a resource that you're not going to leave sitting there idle. As you deploy more cycles and more compute capability, the amount of work that gets done increases to fill how many cycles are available. So your energy goes up. The total energy usage and associated carbon emissions from data centers have risen dramatically, but how will trends continue in the next decade? These AI processes and technologies are developing so rapidly, simply extrapolating from past trends is unlikely to provide reliable estimates. Another issue is a lack of transparency around the exact specifications of many data centers, which are often considered proprietary information by corporations. Eric Massinet, a mechanical engineer at the University of California, Santa Barbara, and Lawrence Berkeley National Laboratory in California studies the sustainability impact of data centers. Massinet explains the difficulty of tracking and projecting the energy usage of AI data centers. In the US, where we have arguably the largest data center fleet in the world, electricity use of data centers was something like 1% of national electricity use for quite a long time. It went up from around 1% to 2% a decade ago to more than 4% in the present day, and that trend is also playing out globally, largely due to investments in new data centers running these very power hungry AI servers. Researchers have been trying to predict the energy trajectory of data centers for a long time, and invariably we get it wrong. This is a sector that's very tough to analyze, even when we have good data. The issue for this particular sector is we don't have great data. A lot of data center operators don't report their energy use. Even when they do, we really don't understand what's going on under the hood, so to speak. We need models that enable us to make rough projections of where the sector is going. In the US, we've only performed three national studies in the last 20 years of data center energy use, and that pace is not nearly quick enough to keep pace with the developments that are happening on the ground. Every day we're hearing about new data center investments in the AI Gold Rush, even given all of this uncertainty. We have enough rough evidence to get moving on some solutions. A lot of those announced data centers haven't been built yet. The time to act is really now. However, the missions are not the only sustainability impact associated with AI. Data centers also use water during the production of energy and to cool excess heat produced by computer systems. Data center energy sources generate air pollutants, which can have significant regional health effects. Shallay, Red, an electrical and computer engineer at the University of California Riverside studies the intersection of AI, energy, and public health. And an article published in communications of the association for computing machinery in the Journal of a US-based non-profit professional membership group, ran in his colleagues estimated that while water usage can vary significantly, some large language models consume the total of 500 milliliters to answer 10 to 50 queries, equivalent to a standard size bottle of water. Ran describes the water usage and air pollution associated with AI data centers. If you look at the overall water number for data center industry, some of the reports from the Lawrence Berkeley National Lab says that the overall water consumption was roughly 17 beating gallons. It's a very small number at the national level, but water is a highly seasonal and regional resource. Some of the reports by the Bloomberg News shows that over two-thirds of the new data centers since 2022 in the US were built in regions with relatively high water stress levels. Some data centers do not even have a water meter. And that's probably not very fair because residents are meter-based on their actual water consumption. When we talk about the emissions from data centers, the first thing that often comes to people's mind is the carbon emission. But actually, there's also some emissions of criteria air pollutants like PM2.5, sulfur dioxide, nitrogen oxides. Scope 1 is for the direct emission for data centers that is primarily from the on-site generators and typically diesel generators. And scope 2 is the electricity usage emission. For the direct emission we use the Virginia as example, we show that assuming that the actual emissions are about 10% of the permitted level, then the overall health impact is roughly 200 to 300 million dollars per year. For scope 2 emission, the overall health impact is even more substantial. High exposure to this air pollutants can create asthma, heart attack, lung cancer, and even premature death. Given the transformative potential of machine learning, one possibility is that AI might be used to find solutions to some of the problems created by AI. Tefic Kosar, a computer scientist at the University at Buffalo in New York, studies ways to improve the performance and sustainability of data centers. Kosar explains how AI might be leveraged as a tool to help address climate change and sustainability challenges. Artificial intelligence has already been quite instrumental in the transformation of climate action and in the development of effective solutions for climate change. But all of this comes with a cost. Training a single large language model emits as much carbon dioxide as a small town with several hundred homes would emit in a year. Using AI to improve the efficiency and sustainability of AI may sound like a paradox, but it is actually not. AI models can be used to significantly improve the efficiency and sustainability of next-generation AI technologies. For example, AI models can be used to track the energy consumption and carbon emission of AI. They can be used for predicting and allocating renewable energy sources for AI workloads. There are also some new efforts in using AI techniques to generate more efficient AI architectures. For example, governments can establish standardized procedures for measuring the environmental impact of AI, light bulbs, TVs and all appliances have energy ratings. So why shouldn't AI models have similar ratings? The rapid growth of the power hungry AI data centers has come at the cost of significant carbon emissions, in addition to regional water stress and air pollution. Even if, as some have suggested, the current acceleration of demand for AI proves to be above all, the long-term sustainability impact of AI could be significant without proactive regulations and development policies. Thanks for tuning in to Science Sessions. You can subscribe to Science Sessions on iTunes, Spotify, Google Play or wherever you get your podcasts. If you like this episode, please consider leaving a review and helping us spread the word. [Music]