Gemini improves fire alerts and radio transcription

Google has connected Gemini to Watch Duty to transcribe firefighters' radios faster and better monitor wildfires. At the same time, the model is coming to Mac to perform background desktop tasks.

Google has presented two very different uses of Gemini, but both revolve around the same idea: saving the user time without stealing their screen. On one hand, the AI helps Watch Duty transcribe firefighters' radio communications and broadcast fire alerts faster. On the other, it can perform tasks on Mac in the background, with human validation at every sensitive step. In both cases, the promise is simple: reduce noise, whether auditory or administrative.

Des radios de terrain enfin compréhensibles

The collaboration with Watch Duty, a wildfire tracking app, was announced by the Gemini team after six months of immersion with first responders. The system was tested in real conditions during the Railroad Fire in Washington State in July 2026, when Watch Duty triggered its first critical night alert from an almost inaudible 911 call. Gemini transcribed it in under 30 seconds. On the ground, this delay matters, as it can speed up alert dissemination to residents and help keep response maps up to date.

The project is not based on a simple generalization of a model trained on common texts. Google states that engineers used over 20,000 hours of Watch Duty radio recordings to teach the AI firefighting terms absent from ordinary dictionaries. As a result, the word error rate dropped by about 15% after integrating the model, while volunteer satisfaction with transcription tools rose from about 35% to 92%. Not bad for an AI listening to radios saturated with interference while everything burns elsewhere.

Une aide conçue pour ne pas remplacer les humains

Google emphasizes one point: Gemini is not there to decide on behalf of rescue teams in a life-threatening situation. Stephanie Chew, project lead at Google, summarizes the goal as follows: “This tool takes on the demanding work of listening, freeing Watch Duty’s volunteers to focus on what matters most: confirming fires and getting warnings to people while they can still act”. For his part, Dave Merritt, co-founder and CTO of Watch Duty, explains that the tool can also be used for other disasters: “This is something the tool also has the capability to do if reports of flooding are communicated via radio,” he adds, highlighting that first responders use Watch Duty to gain visibility on events spanning multiple counties.

The service remains free for app users, and the project was carried out at no cost thanks to a Google.org grant. In terms of load, Gemini currently processes about 200 hours of radio voice traffic per day from 400 streams, with a stated capacity to scale up to more than 10,000 streams if necessary. This scale changes the game for civil security, notably for 911 centers, local fire departments, or coast guard operations, provided strong human supervision is maintained.

Sur Mac, Gemini passe en mode exécutant

On computers, Google is pushing another feature called Computer Use, a function that allows Gemini to chain tasks on Mac in an isolated window. The user describes a goal, then the assistant clicks, types text, and navigates menus, forms, and websites. The whole process runs in the background, with a live preview, while the main screen remains available for work. Users choose accessible applications, can monitor everything in real time, pause at any time, and approve sensitive actions before execution.

Google cites several use cases: batch editing of photos and 3D scenes, filling shipping forms from local PDFs, or building presentations. The feature is deployed in English in the United States, starting now, in the Gemini app for Mac. Again, the benefit is clear: automate repetitive tasks without turning the computer into a black box, something product teams like a lot on slides and much less when they really have to click somewhere.

By the numbers

  • more than 20,000 hours — radio recordings used to train the tool, according to Google and Watch Duty.
  • about 15% — reduction in word error rate after model integration.
  • from 35% to 92% — volunteer satisfaction with transcription tools.
  • about 200 hours per day — voice traffic analyzed by Gemini, from 400 streams.
  • more than 10,000 streams — maximum announced capacity for the infrastructure.

To read