From Warning to Survival: What Does AI Add to Disaster Management?

Science today has no way of predicting the moment of an earthquake, and the U.S. Geological Survey's fact sheet on early warning says so plainly. What it has is something else and narrower: it can detect the first waves, the fastest and weakest, and send a warning before the stronger ones arrive. The difference between the two is the first thing to grasp when we speak of AI and disasters.

Seconds that make a difference

Electronic signals travel faster than seismic waves through the Earth, which opens a window of between a few seconds and a few minutes depending on your distance from the epicentre, and possibly none near it. In that window a train can stop, gas valves can close, and people can move away from windows.

Google has extended this capability by turning the accelerometers in Android phones into a sensing network. In its research blog (July 2025) the company reports that the system was running in 98 countries by the end of 2023, had detected more than 18,000 earthquakes between magnitudes 1.9 and 7.8, issued alerts for more than 2,000 of them, and sent 790 million alerts; and that the number of people with access to early warning worldwide rose from about 250 million in 2019 to about 2.5 billion. In the magnitude 6.7 earthquake in the Philippines (November 2023), the first alert came after 18.3 seconds, giving about 15 seconds near the epicentre and up to a minute farther away. These are the company's own figures, and we have not found independent verification. The blog acknowledges a trade-off between speed and accuracy: an early underestimate of magnitude may miss people in danger, and an overestimate generates false alarms.

Floods: five days, not seconds

A flood is slower than an earthquake, which allows forecasting further ahead. In a paper published in Nature in 2024, researchers reported that their machine-learning model extended the reliability of current real-time forecasts to a lead time of five days, and that forecast skill in Africa became similar to what is available in Europe. The paper also reported that the model performs in basins with no measuring stations at a level close to what existing methods achieve in measured basins.

How much does warning save?

The United Nations initiative Early Warnings for All, launched by the Secretary-General in 2022 with the aim of protecting everyone by the end of 2027, cites the Global Commission on Adaptation for the point that 24 hours' notice before a hazardous event can reduce damage by about 30%. Citing the 2022 Global Status Report, it also says that disaster mortality is eight times lower in countries with extensive early-warning coverage than in those with limited coverage, and that 119 countries are now covered by a multi-hazard early warning system.

These figures concern early warning in general, not AI alone, but they show the value of what AI speeds up and widens.

When an accurate warning fails

The story of the shepherd who cried wolf until no one believed him when the wolf really came fits here. An accurate model is no use unless it becomes a message that arrives in time, that people trust, and that reaches a body that acts. Repeated false alarms erode trust, and response requires training and clear lines of responsibility.

The human role cannot be handed to a model: a person weighs a probability against local context and human cost. The model gives a probability; the decision is the responsibility of whoever holds it. Knowledge of the unseen belongs to God alone, and everything we build is an estimate we act on by taking the means.

What does a successful system look like?

1. Integrated data: sensors, satellites and linked records of known quality.
2. A short decision loop: from model to warning to action, without a jam of orders.
3. Transparency and training: a public that knows why it is asked to evacuate, and regular drills.
4. A human in the loop: an expert who monitors and corrects.

In Saudi Arabia, Arab News (2025) reports that the Baseer system, from the Ministry of Interior with SDAIA, uses computer vision and machine learning to analyse the movement of more than one million worshippers a day in the Grand Mosque, in order to anticipate crowding. It is an example of linking monitoring to decision in a setting where millions gather.

Survival is built brick by brick: a model that detects, a system that delivers, a person who decides, and a community that trusts and responds.