AI Emergency Alerts Are Coming. Humans Might Not Be Ready

After Hurricane Isaias struck the Gulf Coast as a Category 3 storm, meteorologist Laura Myers — director of the University of Alabama's Center for Advanced Public Safety — took a call about a troubling artifact: a local emergency manager had published a graphic predicting a "sting jet" of high winds threatening the Alabama coast, and the graphic had been generated by an AI chatbot. Myers said the official had been trying to focus stakeholders on local risk. "It did what he wanted it to do, but it was not meteorologically accurate," she said.

Myers and most emergency management experts agree that generative AI use is probably already pervasive at state and disaster departments, even though most agencies don't officially permit it. Hundreds of vendors are selling AI tools to emergency managers for work like flood damage assessment and resource coordination, but public communication — the most visible task — is where both agencies and startups have the strongest interest. The appeal is understandable: emergency managers are often chronically understaffed, and many are simply bad at the writing itself.

The flaws are human, not technical. Emergency managers write in jargon, omit background details they assume residents already know, and pad alerts with irrelevant information. "A lot of agencies are really limited in resources and the number of people that they have," said Jeannette Sutton, a disaster researcher at the University at Albany who maintains a dashboard of disaster alert templates. "I think they recognize it as a priority, but they don't prioritize it until they realize they should have."

That matters because state and local emergency managers — not FEMA — are the first line of defense after extreme weather, and public communication is how they build or destroy what they call "our currency": trust. The history of human failure is long. In 2016, neighboring California counties issued "voluntary" and "mandatory" evacuation orders for the same wildfire. Australia's 2019 "Watch and Act" wildfire phrase later drew an official acknowledgment that it caused "significant confusion." And after the 2017 Thomas Fire, Santa Barbara County's alert to Montecito residents in the voluntary wildfire evacuation zone — "stay alert to changing conditions and be prepared to leave immediately" — reached people suffering warning fatigue who read it as optional. A subsequent mudslide killed 23 people and injured more than 160.

Sutton's warning frames the open question ahead: governments that issue bad warnings forfeit credibility. "They're losing the opportunity to be that voice to help people to be safe." What remains unresolved is whether AI-drafted alerts can be made meteorologically accurate, who verifies them, and whether agencies that ban generative AI in practice can govern its use where residents' lives depend on the wording.