Google Publishes White Paper on AI Disaster Prediction

The white paper's goal fits in one sentence. The two authors wrote that "no one should be surprised by a natural disaster," and Google Research put the same sentence forward as its vision when it announced the paper on X. The numbers behind that sentence are heavy. According to the paper, in 2025, one of the three warmest years on record, more than 350 major weather-related disasters affected more than 110 million people worldwide, and record heat in Asia reached 770 million people. The true cost of natural disasters, including indirect losses, exceeds $2 trillion a year, and over the past decade the number of people affected by weather-related disasters rose 75%, leaving more than 250 million people displaced from their homes.

To a sociologist's eye, the structure of the paper is itself an argument. It sets technology, access and partnerships side by side and insists that prediction is only half the battle, because warnings have to reach people. According to the paper, Google works with official alert agencies in more than 100 countries, and in 2025 it connected people with crisis information more than 10 million times a day on average. The view running through the whole document is that who receives a warning and what they do with it shapes the outcome of a disaster more than the accuracy of the prediction model does.

The technology is organized by hazard. Riverine floods are forecast up to seven days ahead, urban flash floods up to 24 hours ahead and cyclones up to 15 days ahead, and wildfire boundaries are tracked in near real time. Riverine flood forecasts cover more than two billion people in around 150 countries, and urban flash flood forecasts were added to Flood Hub in 2026. The method behind them, Groundsource, uses Gemini to turn millions of historical flood records spanning more than 150 countries, along with news reports and public records, into structured data. METAL previously reported that Google AI forecasts riverine floods seven days out and urban flooding 24 hours out.

The most striking part of the paper is where warnings turn into money. In Nigeria, GiveDirectly and the International Rescue Committee used Google's forecasts to deliver anticipatory assistance to households before flooding peaked, and in Adamawa State the UN humanitarian agency OCHA ran a $7 million initiative along the Benue River that helped about 350,000 people prepare. In Kenya, the fintech company Atram ran a pilot with around 10,000 borrowers in which local flood alerts automatically paused loan repayments. The moment a forecast becomes cash and a loan deferral, disaster response moves from a meteorology problem to a problem of social institutions.

The monsoon case in India is on a different scale. According to the paper, in 2025 the Indian government worked with an international research team to select a monsoon forecast model, and Google's NeuralGCM and the European Centre for Medium-Range Weather Forecasts model did the best job of predicting when the monsoon would start. A system that blended the two models with rain-gauge statistics from the India Meteorological Department predicted the onset of the monsoon up to a month in advance and even spotted an unusual mid-season dry spell ahead of time. The Indian government delivered the forecasts to 38 million farmers, and the paper says that kind of proactive decision-making has the potential to nearly double annual farmer income.

The cyclone evidence is the Melissa story from the opening. According to the paper, the experimental model built on WeatherNext generates up to 1,000 scenarios in minutes to predict a cyclone's formation, track, intensity, size and shape 15 days ahead, and its three-day forecasts are as accurate as prior models' two-day forecasts, giving forecasters an extra day. Evan Thompson, principal director of the Met Service Jamaica, said in the paper, "Because of that early warning, we were able to give advance notice to the public to say move from certain areas," adding that it saved lives and livelihoods. METAL previously reported on how DeepMind's model caught a hurricane's rapid intensification a day early.

Earthquakes are a matter of detection in seconds, not prediction. The accelerometers in Android smartphones are pooled into something like a global network of mini-seismometers that catch the P wave and send alerts before the more destructive S wave arrives. The paper says that during the magnitude 7.2 and 7.5 doublet earthquake in Venezuela in June 2026, the first alert went out three seconds after the earthquake began, and more than 11 million alerts were delivered anywhere from a few seconds to two minutes before the S wave, depending on distance from the epicenter.

Wildfires have moved to satellites. Google maps wildfire boundaries in 34 countries, added seven of them in 2026 alone, and refreshes the maps every 15 to 20 minutes depending on the region. In 2025 it generated more than 520 wildfire crisis alerts on Search that reached more than 75 million users. FireSat, led by the nonprofit Earth Fire Alliance, followed its first proto-satellite in early 2025 with three more satellites in July 2026, and once the constellation of more than 50 satellites is complete, the goal is to detect fires the size of a five-by-five-meter shipping container anywhere on Earth at intervals of 20 minutes or less. John Mills, CEO of Watch Duty, said in the paper, "Speed matters in emergencies," and said the Google fellowship would use Gemini to get life-saving information to people faster.

The flow of money is in the document too. According to the paper, Google.org contributed more than $250 million to crisis support from 2004 to 2025, and as of 2026 it has put more than $40 million into the AI Collaborative: Wildfires coalition. Within that, Watch Duty and the Woodwell Climate Research Center each received $2 million, the University of California San Diego received $1.8 million, and $5 million went to the World Resources Institute's Cool Cities Lab. After a disaster, a damage-assessment model called SKAI compares satellite imagery to score building damage, and after Melissa hit Jamaica the tool assigned preliminary damage scores to more than 385,000 buildings.

The institutional partners' words stress relationships over technology. Kamal Kishore, head of the UN Office for Disaster Risk Reduction, said in the paper that Google has "the ability to extend early warning services to communities and countries that have long been underserved," and Celeste Saulo, secretary-general of the World Meteorological Organization, said what makes the collaboration exciting is that it anchors Google's AI innovation "in trusted and accountable National Meteorological and Hydrological Services." Both statements carry the same division of roles: final authority over warnings stays with government agencies, and the company supplies the models.

The last chapter of the 26-page paper METAL reviewed is about communication rather than prediction. The document diagnoses command centers during a disaster as buried under too much data rather than too little, and says generative AI can distill road closures, rising water levels and blocked evacuation routes into concise briefings, always with human oversight and verification checks. It also raises the balance problem: missing a warning puts lives at risk, while too many false alarms make people tune out. The paper closes with a line stating that Google's crisis models are predictive, experimental and provided for informational purposes, and are not a replacement for official local government emergency instructions.

What the paper shows is that disaster response has changed from work between meteorology and government administration into work in which a data company holds one of the pillars. Each time a forecast turns into cash assistance, a loan deferral or a planting date, the company's model moves one step further into social institutions, and the question of who issues warnings and who is accountable for them grows in importance. The paper leaves that answer with government agencies, and whether that line holds is the place to watch as this work continues.