The World Faces a Sharp Rise in Extreme Weather. Can AI Help?

Alice Hill is the David M. Rubenstein senior fellow for energy and the environment at the Council on Foreign Relations. Colin McCormick is the chief innovation officer at Carbon Direct, a commissioner with the Washington DC Commission on Climate Change and Resiliency, and an adjunct professor at Georgetown University.

The Los Angeles wildfires this January reached a size and ferocity that puts them as some of the most destructive fires California has seen. Entire communities have been wiped out, and more than 160,000 people were under evacuation warning. At least twenty-eight people have died. Emergency warning systems failed to fully meet the moment; a false alarm sent to millions of residents sparked panic while some warnings came twelve hours late, and a lack of time stamps made it impossible to tell if alerts were current or outdated. 

Organizations tracking climate change say such extreme weather events are likely to become more common, say . But the ability to adapt may have a crucial new tool—emerging artificial intelligence (AI) programs propelling early warning systems.

Forecasts in a Hotter World

In the past twenty years, human-caused climate change has resulted in more than half a million deaths and massive economic damage. Rising temperatures have led to bigger storms, heavy precipitation known as “rain bombs,” larger wildfires, longer periods of extreme heat, and extended drought. According to the UN Office for Disaster Risk Reduction, from 1980 to 1999, the world suffered about 4,200 disasters that caused 1.19 million deaths and $1.63 trillion in economic losses. From 2000 to 2019, the number of major disasters soared to 7,348, with 1.23 million people killed and losses nearly doubling to $3 trillion. While some of the rising losses may be attributable to better reporting, inflation, and more people living in riskier areas, climate change also shares some of the blame.

Indeed, today’s disaster intensification loop reflects what climate scientists have long predicted. A study published in Nature estimated that climate change will cause $38 trillion in annual global economic damage by 2050, based on climate impacts in 1,600 subnational regions worldwide over the past forty years.

In the face of accelerating climate change, effective early warning systems are needed more than ever to reduce mortality and economic harm. Such systems provide time to evacuate, lessen damage to homes, and take precautions like moving vehicles and livestock to higher ground to reduce business disruption. Providing just twenty-four hours of advance notice could reduce damage by 30 percent, says the UN Environment Program, which estimates that investing $800 million in early warning for developing countries could prevent losses of $3–16 billion.

Yet only about half of the world’s countries have the resources to provide early warning about imminent weather conditions. To close the gap, the United Nations launched an ambitious “Early Warning for All” initiative in 2022, aiming to ensure that every person on the planet is protected by these systems by the end of 2027. 

But there are significant barriers. Effective systems require accurate weather forecasting, timely issuance, and appropriate infrastructure for disseminating warnings. For many countries, progress is stymied by a lack of technical expertise to develop, operate, and maintain tracking systems; unclear decision-making protocols to reach local communities; and a knowledge gap in metrological monitoring, historical data, and forecasting ability. Africa has the least developed land-based weather observation network, with many existing systems outdated and poorly maintained. According to the World Meteorological Organization, Africa has less than forty radar stations, unevenly distributed, with only about half able to provide accurate short-term forecasts. Up to 60 percent of Africans are not protected by early warning systems. 

The Potential of AI

In both developing and developed nations, AI can play a meaningful role in improving forecasting. This technology can rapidly analyze and synthesize vast amounts of historical weather data, learning to recognize subtle patterns that play out repeatedly to make highly accurate predictions about future weather. However, the introduction of AI also faces challenges, including limited observational weather data in many parts of the world, security issues, excessive power demands, and potential bias in outcomes. 

Conventional, state-of-the-art weather forecasting uses detailed scientific calculations of the physics and chemistry of the atmosphere, land, and oceans. Major national weather services such as the U.S. National Oceanographic and Atmospheric Administration (NOAA), and intergovernmental bodies such as the European Center for Medium-Range Weather Forecasts, operate expensive and energy-intensive supercomputers to run these calculations, which can take hours to complete. Typically, these elite institutions produce and publish ten-day-ahead forecasts several times a day, using up-to-the-minute weather data from satellites and other systems for calibration.

Very few countries have access to supercomputers and highly trained weather scientists, so those countries often rely on less accurate methods that can run on normal computers. This problem has inspired many research groups to try using AI to make accurate, timely forecasts. While training AI models is a time-consuming, energy-intensive process, using them to make forecasts is often hundreds of times faster and less energy-consuming than conventional calculations. This research has paid off dramatically over the past year; several emerging AI weather-forecasting systems are highly accurate and can run using a simple laptop or a tiny amount of online cloud computing.

One important example is GraphCast, an AI weather model released in late 2023 by Google DeepMind. During the 2024 Atlantic hurricane season, GraphCast proved remarkably accurate at predicting the path of hurricanes, identifying the landfall location of Hurricane Beryl in July and Hurricane Milton in October well before conventional weather models. (Google has recently released an even more capable AI model called GenCast.) Other large technology companies such as Huawei and Nvidia, and a growing number of smaller tech startups, have also developed advanced AI weather models with impressive forecasting accuracy. 

Government weather services are working hard to figure out how to incorporate these new models into their forecasting services. The European forecasting center has adopted GraphCast on an experimental basis, offering AI-based weather forecasts on its website. NOAA has held workshops with leading academics and technology companies on AI weather forecasting, and NASA is collaborating with tech company IBM to develop foundational AI weather models. 

Meanwhile, innovative new partnerships are emerging to deliver AI-based weather forecasting to regions outside the Organization for Economic Cooperation and Development (OECD), including:

• The UN World Food Program and scientists at Oxford University collaborating with African trade bloc the Intergovernmental Authority on Development’s Climate Prediction and Applications Center to bring AI extreme weather forecasting and early warning systems to countries in east Africa, including Kenya and Ethiopia;
• The Philippine Atmospheric, Geophysical and Astronomical Services Administration’s plans to adopt AI weather forecasting; and
• A coalition of international development banks and national governments announcing the $1 billion Agriculture Innovation Mechanism for Scale initiative at the COP29 climate conference in November 2024 to help national weather services across the globe adopt AI-supported weather forecasting with a focus on providing advanced forecasts to farmers.

The Limits of AI

While these AI models offer enormous promise for bringing accurate, specific weather forecasts to more countries and regions, they are not a panacea. AI-based forecasting continues to face funding, energy, and implementation challenges.

Funding for data collection. Measurements from local weather stations are needed to calibrate AI model forecasts, and these are not available in many parts of the Global South. Also, while AI models have shown very good ability to predict storm paths, they are not as accurate in predicting some other weather features like maximum storm wind speeds. AI-based forecasting depends on increased international funding to support the development of local weather data collection infrastructure. Just as governments need to invest in local meteorological expertise, they will need assistance to attain access to AI capacity to generate the forecasts. 

Dirty energy in AI. Global electricity use is surging with the rise of AI, much of which is powered by fossil fuels. The energy drain in the near term from AI will also require vigilance by governments to ensure that power demand does not conflict with climate change emissions reduction goals. 

Private-public cooperation. It is also important to remember that even though private-sector technology companies like Google and Nvidia have led the development of AI weather models, those models depend heavily on historical and current observational weather data collected by government weather services and research programs. This dependence highlights the need to resolve important policy questions about the best way to structure ongoing public-private cooperation. Widespread adoption of AI-assisted forecasting will require greater collaboration between forecasters and local organizations to tailor global forecasts to regional needs.

Accessibility of forecasting. Another challenge is making sure that forecasts improved by AI are readily accessible and available to everyone—and that they are acted upon. In December 2024, residents of Mayotte, a densely populated archipelago between the African continent and Madagascar, reportedly did not credit warnings issued by MeteoFrance, a regional meteorological service, fifty hours before Cyclone Chido made landfall, potentially increasing the loss of life. 

Accurate weather forecasting is, and should remain, a freely available public good. If people, communities, or companies want to pay for added services, they can, but core life-saving services should be free of charge. AI has the potential to greatly improve weather forecasting. With better forecasting, early warning systems have the potential to save more people’s lives, livelihoods, and homes. 

This work represents the views and opinions solely of the authors. The Council on Foreign Relations is an independent, nonpartisan membership organization, think tank, and publisher, and takes no institutional positions on matters of policy