When Hurricane Isaias struck the Gulf Coast as a Category 3 storm, a meteorologist in Alabama warned a state emergency manager that a graphic depicting a potential “sting jet” had been produced by an AI chatbot. The image proved inaccurate, illustrating how generative tools are already slipping into local disaster communications despite most agencies prohibiting their use.
Hundreds of firms now market AI applications for flood-damage assessment, resource coordination, and public messaging to state and municipal emergency offices. While formal policies often forbid staff from using such systems, many practitioners admit that the technology has become commonplace in day-to-day operations, especially for crafting alerts that must be issued quickly.
Human-written alerts have a history of missteps that can cost lives. In 2016, neighboring California counties issued contradictory evacuation orders for the same wildfire. Australia’s 2019 “Watch and Act” warning confused residents, and after the 2017 Thomas Fire, Santa Barbara County’s vague evacuation notice contributed to a deadly mudslide that killed 23 people and injured more than 160.
Proponents argue that AI could fix these shortcomings. Large language models can scan thousands of past alerts, synthesize best-practice research, and rewrite a draft in seconds. Rob Dale, a former meteorologist now heading emergency management in Ingham County, Michigan, said such a system would likely outperform a practitioner who has sent only one alert in fifteen years.
A startup called EM1 has trained a model on research from disaster-communication expert Sutton. When given a raw New Mexico flood warning that read, “Breeched banks. The water is rising. EVACUATE NOW…,” the AI produced a revised version that identified the sender, specified locations, clarified evacuation instructions, corrected spelling, and added a source for updates, meeting standard communication criteria.
EM1 reports contracts with roughly a hundred government agencies and school districts, though none currently rely on the system to automate live alerts. The tool is used for scenario planning and internal briefings, while many emergency managers reportedly draft alerts in seconds with ChatGPT or Gemini during training sessions, according to Sutton’s observations.
Nevertheless, experts caution that public acceptance may lag behind technical capability. In Syracuse, New York, a public-employee union sued the city for employing an AI service to answer non-emergency 911 calls, highlighting concerns that citizens could reject messages they perceive as inauthentic or generated by machines.