Emergency managers are quietly using AI chatbots to write public alerts, experts say

Left · 2Experts interviewed by reporter Jake Bittle say AI chatbots are likely already widely used by state and local emergency managers in the US to write public alerts and disaster messages, even though most agencies don't permit employees to use generative AI. 1 2 Left · 2Sarah Miller, a professor of emergency management at Pierce College in Florida, said official adoption is low but unofficial use is common, with staff using personal or free accounts. 1 2 Left · 2The article gives an example of an error: after Hurricane Isaias made landfall on the Gulf Coast as a Category 3 storm, a county emergency manager in Alabama published an AI-generated graphic of a "sting jet" threat that, according to Laura Myers of the University of Alabama's Center for Advanced Public Safety, was not meteorologically accurate. 1 2

Left · 2The case for AI rests on the claim that human-written alerts are often poor: they use jargon, omit information the public lacks, and include distracting detail, partly because agencies are short of staff. Examples cited include conflicting 2016 evacuation orders in adjacent California counties, Australia's confusing 2019 "Watch and Act" wording, and the 2017 Montecito mudslide, which killed 23 people after evacuation notices were read as optional. 1 2 Left · 2Rob Dale, who runs emergency management in Ingham County, Michigan, said AI would probably do better than an out-of-practice human in a crisis. He has used it to draft practice alerts but would want to control sending any alert himself. 1 2 Left · 2EM1, a startup selling AI software to emergency managers, says its model, trained using research by University at Albany disaster researcher Jeannette Sutton, is better than almost any human emergency manager. Its roughly 100 government and school-district clients are not using it to automate extreme-weather alerts and only use it for planning and internal preparation. 1 2

Left · 2In a demonstration, Sutton had the EM1 model revise a San Juan County, New Mexico flood alert that read "Breeched banks. The water is rising. EVACUATE NOW." The result named the sender, told people to leave for higher ground and not to drive through floodwater, and pointed to an update source. The article notes the original misspelled "breached" and gave no guidance on where to get more information. 1 2 Left · 2Even researchers who favour the technology warn it could backfire: people who distrust AI or the government may reject agencies that publish AI-generated content. Ginny Katz of the consultancy HazAdapt said such content does not feel authentic, and a union of public employees in Syracuse, New York sued the local government over using an AI service to answer non-emergency 911 calls. 1 2