In the United States, the use of language models for emergency notifications is being discussed. The reason is clear: during a fire or flood, local service employees often write messages in a hurry. The text may miss the addressee, the exact risk area, or a clear command. Researcher Jeanette Sutton showed on a real flood warning how the EM1 program corrects the structure: names the sender, indicates the evacuation area, suggests moving to higher ground and provides a source of updates. This is a useful editorial task, especially for small teams.
But the beautiful formulation does not check the original data. In Alabama, a local employee used a chatbot to create a hurricane threat chart with a meteorologically incorrect description of dangerous winds. An error could make a fictitious scenario look like an official forecast. According to WIRED and Grist, EM1 has about a hundred contracts with institutions, but none of them have yet transferred the system to automatically send emergency weather warnings. The program is used to prepare plans and internal work. A reasonable application scheme here is as follows: a person receives a verified forecast, determines the territory and action for residents, the model helps to shorten and clarify the message, and an authorized specialist verifies and sends it. Separately, you need to monitor the frequency of mailings: an excess of alarms leads to fatigue from warnings. Speed is important, but in an emergency message, accuracy is more important than smooth text.