# Groundsource: Google Uses Gemini to Turn News Into 2.6M Flood Records

Times of AI Desk · 2026-03-12 · Models

[https://timesof.ai/2026/03/google-deepmind-prepares-next-gemini-model](https://timesof.ai/2026/03/google-deepmind-prepares-next-gemini-model)

> Google Research introduced Groundsource, which uses Gemini to turn news reports into structured historical data. Its first open dataset holds about 2.6 million urban flash-flood events across 150+ countries; Google reports 82% of extracted events were accurate enough to be practically useful.

**Corrected October 14, 2026:** Rewrote the article around Google Research's Groundsource announcement. The original premise — that Google DeepMind was widely described as finalizing a new Gemini model — could not be sourced and was removed.

Google Research on **March 12** introduced **Groundsource**, a method that uses **Gemini** to turn unstructured news reports into structured historical data. Its first open dataset covers urban **flash floods**: about **2.6 million** historical flood events across **more than 150 countries**, from 2000 to the present.

## How it works

Google says the pipeline analyzes news reports where flooding is the main subject, isolates article text in **80 languages**, standardizes it into English with the Cloud Translation API, and then uses Gemini to classify real events versus warnings or policy talk, anchor relative dates against the article’s publication date, and map locations to standardized polygons with Google Maps Platform.

## Google’s validation numbers

- In manual review, **60%** of extracted events were accurate in both location and timing, and **82%** were accurate enough to be practically useful, per Google.
- Spatiotemporal matching found Groundsource captured **85% to 100%** of severe flood events recorded by GDACS between 2020 and 2026.
- Google says the data supports near-global urban flash-flood forecasts up to **24 hours** ahead, rolling out in Flood Hub.

## Limits

- Accuracy figures are Google’s own validation, not an independent audit.
- News coverage is uneven by region and language; Google says it is working to expand rural coverage.

## Sources

- [Google Research: Introducing Groundsource: Turning news reports into data with Gemini (March 12, 2026)](https://research.google/blog/introducing-groundsource-turning-news-reports-into-data-with-gemini/)