Google used Gemini to sort five million news articles and identify 2.6 million floods, turning written reports into a geo-tagged record for forecasting flash floods. The system gives weather models evidence in places where radar stations, local archives, and public infrastructure are thin. A newspaper, usually treated as a record of what happened, becomes part of a machine estimating what may happen next.

The method also changes the status of description. A reporter’s sentence about water crossing a road can enter a dataset, acquire coordinates, and help determine whether an alert appears over a city. The human account is compressed into a measurable event, but its usefulness depends on the uneven geography of journalism. Some places are documented constantly; others enter the archive only when damage becomes spectacular. The model can rebalance that map while inheriting the attention patterns that produced it. Flood prediction now reads the past as an instrument. The emergency message on a phone may begin with an old paragraph written after a road had already disappeared.