Imagine being the person responsible for keeping a city safe when the sky turns orange and the wind starts carrying smoke. You have minutes, not hours, to decide whether to tell people to evacuate, shelter in place, or simply wait and watch. Your phone is buzzing, the radio station is waiting, and social media is already full of rumors. The old way of doing things means sitting down with a keyboard and carefully crafting a warning, choosing words that will make people take action without panicking. But there is another way now: open a chatbot, type in the facts, and let it write the message for you. That scenario is not hypothetical. EM1, a company that sells AI software to public safety agencies, says it has contracts with around a hundred government agencies and school districts. Yet none of those clients are using the technology to automate extreme weather alerts just yet. For the moment, they are only using it to workshop plans, prepare for internal drills, and test how they might respond to hypothetical disasters. It is a cautious, low-risk start—a way to get comfortable with the technology before trusting it with the public’s safety. Officials can ask the software to draft a hypothetical evacuation order or craft a warning about a mock chemical spill, and then study how the AI phrases things. These are useful experiments, but they are a far cry from turning over the microphone to an algorithm during a real emergency. Even so, every expert interviewed for this story said they are certain that many emergency managers are already using chatbots to write messages, even if they won’t admit it. The gap between what agencies officially approve and what individual employees actually do is wider than most people realize. A manager might use a personal ChatGPT account at home to draft a heat advisory, then tweak it and send it out. There is no policy against it because there is no policy at all. And in an understaffed office, no one is asking too many questions.
Sutton, a researcher who has spent years teaching emergency managers how to write effective alerts, sees the evidence of this unofficial use all the time. In her training sessions, people often hand her polished, professional-looking messages that they have drafted in seconds using ChatGPT or Gemini. The writing is clean, the tone is appropriate, and the structure follows all the best practices that used to take years to learn. It is hard to tell at a glance that a machine wrote it. Sarah Miller, a professor of emergency management at Pierce College in Florida and a longtime instructor at the National Disaster Preparedness Training Center, says the situation is clear: “Official adoption is pretty low, but people are unofficially using it, absolutely. They’re using their own personal accounts, or they’re using free ones.” The temptation is easy to understand. Emergency managers are often stretched thin, covering huge territories with small teams. A chatbot can instantly generate multiple versions of an alert—one for the general public, one for vulnerable communities, one for Spanish speakers, one with a lower reading level. It can make the language more urgent or more reassuring with a single prompt. In a field where every minute counts, that kind of speed is hard to resist. And the people doing it are not lazy or careless. They are often the ones who care the most, the ones who lie awake at night worrying about whether their community is prepared. They see AI as a way to do more with less, to close the gap between what they should be doing and what they actually have time to do. Even the managers who use chatbots rarely tell their supervisors, because they worry about being seen as cutting corners or relying on something that hasn’t been approved. So the practice stays hidden, unofficial, and largely unexamined.
On the surface, this looks like an easy trade-off. If AI is better and faster than humans, and if the difference between a good alert and a bad alert can be life and death, then shouldn’t we be using AI for every heat wave, wildfire, and flood? Many researchers who are optimistic about chatbot technology say yes—but with a major warning. The switch could backfire. Members of the public who are skeptical of AI, or who already distrust the government, may refuse to take advice from an agency that publishes AI-generated graphics or social media posts. Trust is fragile, and people notice when a message feels robotic. There is a recent example of this tension: in Syracuse, New York, a union of public employees sued the local government for using an AI service to answer non-emergency 911 calls. The lawsuit wasn’t just about whether the AI could do the job; it was about the idea that a machine was handling something so human. People want to know that someone is actually in charge, not a black box. Ginny Katz, the founder of HazAdapt, a consultancy that helps disaster officials decide how to use technology, explained why that matters: “We trust things that we perceive [are] authentic and care about us. AI-generated content does not feel authentic.” No matter how perfect the grammar is, if people sense that a robot wrote the alert, they may not listen. And in an emergency, ignoring a warning can be deadly. That is the paradox: the technology might be able to write a better message, but if the public doesn’t believe it came from a person who cares, it loses its power. Authenticity, not accuracy, may be the real difference between an alert that saves lives and one that gets deleted with a swipe.
Another risk is that AI could make the problem of warning fatigue even worse. The entire purpose of an emergency alert is to stand out from the constant stream of notifications and messages that fill our lives. But if chatbots make it easier and cheaper to send out alerts, emergency managers might send more of them, and more often. When people receive a dramatic evacuation order for a small storm that turns out to be nothing, they start to tune out. The next time the alert is real, they may ignore it. A machine that can produce a hundred messages a minute doesn’t understand that sometimes the best warning is the one you don’t send. There is also the issue of human judgment. AI software is only as capable as the person using it. A seasoned emergency manager can ask a chatbot to draft a series of alerts, review the results with a practiced eye, and catch subtle errors or harmful phrasing. But an inexperienced manager, under pressure and in a hurry, might not know how to evaluate what the model produces. ChatGPT and Claude can generate responses that sound authoritative but are subtly wrong. They can invent details, misstate evacuation routes, or use language that confuses people. They don’t know the history of a community, or how residents reacted to past warnings, or which neighborhoods have learned not to trust the government. Without a human expert in the loop, a mistake can spread to thousands of people in seconds. The software can be a powerful tool, but it is not a substitute for judgment. It is more like a very fast intern with no context and no common sense, and it needs someone older and wiser to keep it in line.
Even though the technology is a long way from being able to operate on its own, that hasn’t stopped many understaffed emergency management departments from using it. The pressures are real: budgets are tight, staff turnover is high, and the number of extreme weather events is increasing. A department that can barely afford to keep its radios working is not going to invest in a custom AI system. But a free chatbot on a personal phone costs nothing and can save hours. That is why the unofficial use is likely to continue growing, regardless of what official policies say. As for how all of this is affecting public trust in government, experts say it is simply too early to tell. The effects of AI-generated content are complicated. Some people may feel that a machine-written alert is more objective and data-driven. Others may see it as cold and impersonal. Justin Snair, the founder of Preppr, a company that offers AI software to emergency managers, poses a simple but revealing question: “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?” The answer, he says, depends on the person receiving it. There is no single response, and that is exactly what makes the future so uncertain. Some people are reassured by technology; they assume that a computer must be right because it has no emotions. Others are deeply suspicious of it, and anything that feels automated makes them doubt the message. Emergency managers have to navigate that split, knowing that the same AI tool that helps one person feel informed might make another person feel manipulated.
Where does that leave us? The honest answer is that we are in an awkward middle period. AI is already part of emergency communication, even if it is not always visible. It is drafting messages, helping with plans, and quietly shaping how the public receives urgent information. The challenge is not to decide whether to use it—that decision has already been made, unofficially—but how to use it responsibly. That means keeping a human in the loop who understands the community, can look at a chatbot’s output and say, “This doesn’t sound like us,” and knows how to adapt a message for the people who need to hear it. It means being transparent about when AI has been used, not as a way to hide, but as a way to show that the technology is a tool, not the voice of authority. It means resisting the urge to send more alerts just because we can, and remembering that the goal of an emergency alert is not simply to be fast or accurate; it is to move people to act. That requires trust, and trust is built through authentic, human connection. A chatbot might help write the words, but the responsibility, the care, and the accountability still have to come from a person. We may look back at this moment as a time when we stumbled toward a new kind of communication, learning the hard way that machines can do a lot but cannot replace the feeling of another human looking out for you. Until we figure out how to preserve that trust, the safest path is probably the least glamorous one: use AI as a helper, not a replacement, and let the humans take the blame when things go wrong—and the credit when they get it right.
