This story was produced by Grist and co-published with WIRED. It is part of Distorted, a Grist series exploring the spread of misinformation in the wake of extreme weather, and is published with support from the CO2 Foundation.
Earlier this month, after a major hurricane made landfall on the Gulf Coast as a powerful Category 3 storm, emergency management expert Laura Myers received an unexpected phone call from a meteorologist stationed in Alabama. Myers, who serves as the director of the University of Alabama’s Center for Advanced Public Safety, maintains a steady dialogue with state and local emergency managers throughout her region to navigate the complexities of disaster response. During the call, the meteorologist raised a red flag: a local emergency manager had just published a graphic illustrating where the hurricane might threaten the Alabama coastline with a destructive "sting jet" of high-velocity winds.
The underlying issue was not just that the weather pattern was complex, but how the graphic came to be. The local emergency manager had generated the image using an artificial intelligence chatbot, and the visual data was fundamentally inaccurate.
"That county emergency manager was trying to get his stakeholders to focus on the risk in that county, and so he had asked AI to help him," Myers said. "It did what he wanted it to do, but it was not meteorologically accurate."
While the vast majority of official emergency management agencies maintain policies strictly prohibiting employees from using generative artificial intelligence for public-facing communications, Myers and other leading disaster management experts agree that the technology is likely already pervasive across state and local disaster departments. Today, hundreds of technology companies actively market specialized artificial intelligence tools to emergency managers, pitching capabilities ranging from automated flood damage assessment to complex resource coordination. Yet, public communication remains one of the primary areas where both government employees and tech startups show the greatest enthusiasm.
The primary driver behind this trend is straightforward: most emergency managers are not naturally skilled at drafting high-pressure public alerts or communicating rapidly during unfolding crises. This limitation has nothing to do with their professional competencies and everything to do with human nature. Under immense stress, people tend to write in dense, bureaucratic jargon that non-experts struggle to decode. They frequently leave out vital information under the faulty assumption that the general public already possesses background knowledge, while simultaneously including irrelevant details that can easily distract citizens from the critical actions they need to take.
"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 has developed a comprehensive dashboard providing standardized templates for disaster alerts. "I think they recognize it as a priority, but they don’t prioritize it until they realize they should have. They’re just trying to get their job done without screwing something up."
Emergency managers are the government officials on the ground charged with leading disaster response and recovery efforts. At the state and local levels, these officials and their agencies form the absolute first line of defense following an extreme weather event, long before federal entities like the Federal Emergency Management Agency can mobilize. Because of this frontline role, emergency managers often emphasize that "trust is our currency," noting that it is precisely through their communications with the public that this essential trust is either built or permanently squandered.
As climate-driven disasters become increasingly frequent and widespread, history offers numerous examples of flawed human communication severely compromising emergency responses. In 2016, for instance, adjacent counties in California issued contradictory "voluntary evacuation" and "mandatory evacuation" orders for the exact same active wildfire. Similarly, during the devastating 2019 bushfires in Australia, officials issued warnings utilizing the phrase "Watch and Act." The Australian government later acknowledged that the wording caused "significant confusion" because residents had no clear understanding of what the directive actually required them to do.
In the worst-case scenarios, inadequate communication can spell the difference between life and death. This tragic reality played out in Montecito, California, in the aftermath of the 2017 Thomas Fire. Just weeks after the destructive blaze was contained, a heavy rainstorm triggered a catastrophic and deadly mudslide. The government of Santa Barbara County struggled intensely to communicate the rising risk to residents who were still deeply traumatized by the fires and suffering from what researcher Jeannette Sutton defines as "warning fatigue."
During the event, the county sent alerts to residents living within the designated "voluntary evacuation zone" established for the previous wildfires, advising them to "stay alert to changing conditions and be prepared to leave immediately." While most of those residents should have fled immediately, many perceived the evacuation notice as merely optional. Others simply felt emotionally and physically exhausted by the prospect of evacuating for a second time in a single month. Ultimately, the subsequent mudslide claimed 23 lives and left more than 160 others injured.
Governments that release confusing or poorly formulated warnings run the severe risk of permanently eroding public trust, Sutton noted. "They’re losing the opportunity to be that voice to help people to be safe."
In theory, drafting emergency alerts represents a textbook use case for artificial intelligence, particularly the sophisticated generative models that have evolved at a breakneck pace over recent years. While FEMA regularly conducts training sessions to instruct emergency managers on best practices for synthesizing and distributing public alerts, many officials find it difficult to recall those detailed protocols when an actual emergency breaks out. A large language model, by contrast, operates differently. It can ingest hundreds of thousands of historical alerts in a matter of minutes, digest the entirety of academic research on effective crisis communication, and apply those insights to edit a human-drafted alert in mere seconds.
"If you’ve only sent one [alert] in your life, and that was 15 years ago, and you need to remember all your training, and you’re thinking ‘Hey, there’s a chemical plume coming my way, the phone’s ringing, the radio’s blaring, I need help,’ the AI is probably gonna do a better job than you are," said Rob Dale, a former meteorologist who currently runs the emergency management department in Ingham County, Michigan.
Dale has utilized AI tools to draft several practice alerts behind the scenes, though he has never been forced to deploy one during an actual crisis. Ingham County, located in central Michigan, is largely shielded from catastrophic weather phenomena like hurricanes and major wildfires. If such an event were to occur, Dale stated he would have no fundamental objection to utilizing a chatbot to assist in drafting incident communications, provided he retained ultimate control over approving and sending the final alert.
Some industry experts believe the technology has already advanced to a point where it is ready for broader deployment. EM1, a software startup that designs artificial intelligence tools for emergency managers, has incorporated Sutton’s academic research to train a proprietary model. Executives at the company assert that the resulting software already outperforms almost any human emergency manager in specific testing scenarios. Having tested the model in-house against various simulated crises—including fast-moving wildfires and tsunamis—the company reports that the technology consistently matches or exceeds human capabilities across standard benchmark tests.
To demonstrate the software’s practical application, Sutton shared an example using an actual alert recently issued by a New Mexico county attempting to warn local residents about severe flooding along the San Juan River.
The original alert read starkly: "Breeched banks. The water is rising. EVACUATE NOW. McGee Park open for people and pets, RV."
When Sutton asked the EM1 software model to evaluate and revise the message, the system generated an overhauled alert within seconds. The revised version addressed the primary shortcomings of disaster communication: "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 AI model successfully identified and corrected multiple critical deficiencies in the original text. The initial message failed to state which agency was issuing the warning, omitted the specific geographic location experiencing flooding, neglected to clarify exactly who needed to evacuate, and provided no practical safety guidance on how to evacuate safely. Furthermore, the original text misspelled the word "breached" and omitted any reference to where residents could find subsequent updates.
While EM1 maintains contracts with approximately one hundred government agencies and school districts, none have yet fully automated their extreme weather alert systems using the technology. For now, these agencies rely on the software strictly for workshop planning and internal preparation. Nevertheless, every expert interviewed for this reporting expressed absolute certainty that numerous emergency managers are already utilizing commercial chatbots to draft public messages, even if they remain reluctant to admit it publicly. Sutton noted that emergency management professionals attending her alert training seminars frequently present her with polished, high-quality messages that they acknowledge generating in seconds using tools like ChatGPT or Gemini.
"Official adoption is pretty low, but people are unofficially using it, absolutely," said Sarah Miller, an emergency management professor at Pierce College in Florida and a longtime instructor at the National Disaster Preparedness Training Center. "They’re using their own personal accounts, or they’re using free ones."
At first glance, the trade-off appears straightforward: if artificial intelligence can draft better alerts faster than humans, and the margin between an effective alert and a poor one can determine life or death, utilizing AI for every heat wave and wildfire seems imperative. Yet even researchers who remain optimistic about chatbot technology warn that an uncalculated shift could easily backfire. Members of the public who harbor inherent skepticism toward artificial intelligence—or who distrust government institutions to begin with—might outright reject agencies that rely on AI-generated graphics or social media posts. In Syracuse, New York, for instance, a public employee union recently filed a lawsuit against the local government for deploying an AI service to field non-emergency 911 calls.
"We trust things that we perceive [are] authentic and care about us," said Ginny Katz, founder of HazAdapt, a consultancy that advises disaster officials on technology adoption. "AI-generated content does not feel authentic." Additionally, experts point out a distinct risk that chatbots could enable emergency agencies to issue more alerts at a much faster rate, potentially exacerbating public warning fatigue.
As is true across many sectors currently adopting artificial intelligence, disaster experts emphasize that the software remains only as capable as the human operator wielding it. A seasoned emergency manager might successfully leverage a chatbot to streamline public communication by drafting a series of baseline alerts. Conversely, an inexperienced official rushing to broadcast an urgent warning may lack the technical literacy required to accurately evaluate the output generated by models like ChatGPT or Claude.
Even though the technology still requires a vigilant human expert firmly in the loop, that limitation has not stopped understaffed emergency management departments from turning to chatbots for help. Regarding how the widespread proliferation of AI-generated content is ultimately shaping public trust in government institutions, disaster experts maintain it is simply too early to tell.
"If you had a human sending you a text message versus a computer sending you a text message, which one would you believe is being more accurate?" asked Justin Snair, founder of Preppr, another firm providing AI software solutions to emergency management agencies.
The answer to that question, he noted, depends entirely on the individual receiving the message.