Emergency alerts have to do more than describe a threat. People need to understand what is happening, whether they should act, and where to find reliable updates. AI tools may help officials write those messages quickly, but a flawed AI-generated graphic during Hurricane Isaias shows why speed alone is not enough.
Clear warnings are hard to write under pressure
Emergency managers handle disaster response and recovery at the state and local level. Their agencies are often the first line of defense after extreme weather, and the public messages they send can shape whether residents trust official guidance.
Yet writing a useful alert is difficult. Messages can rely on jargon, assume readers know details they do not, or include information that distracts from what people need to do. Agencies may also have limited staff and resources, leaving little time to improve public communication while responding to an unfolding emergency.
Past communication failures show the consequences. In 2016, adjacent counties in California issued different evacuation instructions for the same wildfire: one voluntary and one mandatory. In Australia in 2019, the phrase “Watch and Act” caused “significant confusion,” according to the government.
After the 2017 Thomas Fire, residents in Montecito faced a deadly mudslide during a heavy rainstorm. Some people in a voluntary evacuation zone were told to stay alert and be prepared to leave, even though most should have left immediately. Some read the notice as optional; others did not want to evacuate again. The mudslide killed 23 people and injured more than 160 others.
AI may help turn a draft into a usable warning
Generative AI could assist with the part of emergency work that often gets squeezed for time: making instructions clear. A large language model can review many past alerts and use research about effective warnings to suggest edits quickly. FEMA provides training on alert writing, but officials may not recall every practice when phones are ringing and a hazard is approaching.
Rob Dale, who runs the emergency management department in Ingham County, Michigan, has used AI to draft practice alerts. He said he would consider having a chatbot draft incident communication during a major event, while keeping control of whether the alert is sent. That distinction matters: the tool can help prepare words, but an official remains responsible for the warning.
One example came from a New Mexico county alert about flooding on the San Juan River. The original message said: “Breeched banks. The water is rising. EVACUATE NOW. McGee Park open for people and pets, RV.” It was urgent, but it did not clearly identify who needed to leave, where to go, or how to get updates.
After being asked to revise it, EM1’s model produced a more structured version: “San Juan County: Floodwater is rising near [location]. People in [evacuation area]: LEAVE NOW for higher ground. Do not walk or drive through floodwater. McGee Park is open for people and pets. Check [official update source] for updates.” The revised message names the jurisdiction, specifies an action, warns against crossing floodwater, and points readers toward updates.
Accuracy and trust still depend on people
AI-generated communication can also be wrong. During Hurricane Isaias, a local emergency manager used a chatbot to create a graphic about where a “sting jet” of high winds might threaten the Alabama coast. Laura Myers, director of the University of Alabama’s Center for Advanced Public Safety, said the graphic was not meteorologically accurate.
The incident illustrates a central risk: a polished message or visual can still misrepresent the hazard. If residents receive inaccurate or confusing warnings, agencies may lose trust precisely when people need them to be a dependable source. Sutton has described that trust as something governments risk losing when their warnings fail to help people stay safe.
AI tools are already being marketed for emergency management tasks including flood damage assessment, resource coordination, and public communication. EM1 says its model, trained using Jeannette Sutton’s research, has matched or exceeded human capabilities on most of its benchmark tests, including scenarios involving wildfires and tsunamis. Those company-reported test results do not remove the need for careful review in real emergencies.
Use AI as support, with human oversight
The practical case for AI is strongest when it helps officials make a draft more direct and complete under pressure. A review process can check that the alert identifies the affected area, tells residents what action to take, gives essential safety instructions, and points to an official source for updates.
But the Hurricane Isaias graphic is a reminder that AI output needs to be checked against reliable emergency information. An alert is not just a writing exercise: it connects people to decisions about evacuation and safety. Keeping a human official responsible for the final message gives agencies a chance to catch errors before they reach the public.
As disasters become more frequent and widespread, emergency agencies will face continued pressure to communicate clearly with limited resources. AI may help them do that work faster. Whether it improves public safety will depend on the quality of the information it receives, the clarity of its drafts, and the judgment of the people who approve them.