Google taught AI to predict floods from old newspaper reports

Google has developed an unconventional approach to flash flood forecasting: instead of building costly sensor infrastructure, the company uses archived news reports as a source of historical data. A language model translates qualitative text descriptions of natural disasters into quantitative indicators suitable for training hydrological models.

Climatologists know the problem this method addresses well. Accurate flood prediction models require multi-year records of water levels, precipitation and terrain. But in most developing countries, where floods are deadliest, such infrastructure simply does not exist. Sensors cost money and need maintenance and electricity, all of which are in short supply in vulnerable regions.

Historical accounts of floods do exist, though, in newspaper archives, local outlet reports and government bulletins. The catch is that this data is qualitative: "a severe flood submerged three villages" is not a number a traditional model can work with.

Google proposed using an LLM as a translator: the model reads a historical text and extracts structured numerical estimates of the event's scale, duration and geography. The resulting synthetic numerical series are then used to train a flood forecasting model. In effect, archival journalism stands in for decades of instrumental measurements.

This is a fundamentally new way to handle data scarcity in climate problems, and one that could potentially apply well beyond hydrology.

Google is already testing the system in regions of Africa and South Asia, where flood warnings are practically nonexistent. The company has been developing Flood Hub since 2023 and now covers more than 80 countries. The new method is meant to extend coverage to areas that were previously left out of forecasting because they lacked historical numerical data.

It is a telling example of how LLMs are changing the very structure of scientific data. Previously, the line between "data exists" and "no data" was set by the presence of measuring equipment. Now language models can push that line back, because they can extract hidden quantitative information from texts written with no scientific purpose at all.