In short
Emergency managers are beginning to use AI chatbots to draft disaster alerts, but experts warn that errors, warning fatigue and lost trust could make public warnings less effective. The trend is growing quietly inside state and local agencies, even as most offices officially ban generative AI.
- Emergency managers are using AI to draft disaster alerts, but not usually to send them directly to the public.
- Experts say inaccurate or machine-like warnings could erode trust and make people less likely to respond.
- Most agencies ban generative AI, yet unofficial use appears widespread in disaster offices.
- Supporters say AI can improve speed and clarity if a human expert checks every output.
Emergency managers are increasingly leaning on AI chatbots to draft disaster alerts, even as experts warn that inaccurate wording, weakened trust and rushed automation could make warnings less effective when people most need them. The tension is now visible in real-world use, from a recent inaccurate hurricane graphic in Alabama to broader, unofficial adoption inside disaster offices across the United States.
What is changing in disaster communications?
AI tools are moving into one of government’s most sensitive jobs: telling people when to evacuate, shelter or prepare for fast-moving danger. According to WIRED, agencies and startups are marketing systems for flood assessment, resource coordination and public communication, while emergency managers under pressure are already using general-purpose chatbots to help draft alerts.
The appeal is straightforward. Disaster offices are often short-staffed, and the clock does not stop when a wildfire, flood or hurricane is closing in. Supporters say AI can quickly synthesize research, previous alerts and hazard information into draft messages that are clearer than the jargon-heavy warnings many agencies send now.
But the same speed that makes AI attractive also creates risk. A local emergency manager in Alabama recently used an AI-generated graphic about the threat from Hurricane Isaias; a meteorologist spotted that it was inaccurate, according to WIRED. The episode captures the central problem: a tool can produce something polished and still get the science wrong.
Why it matters
Disaster warnings are only as useful as the public’s willingness to believe and act on them. Emergency managers rely on credibility, and one badly written or incorrect alert can undo that trust quickly, especially in communities that already doubt government messages or distrust AI-generated content.
That matters because state and local emergency managers are the first line of defense in disasters, not FEMA. Their alerts are often the messages that decide whether people leave, stay put, or ignore the next warning entirely. If AI makes those messages faster but less authentic, officials could gain efficiency at the cost of compliance.
The stakes are not theoretical. Past communication failures have already shown how messy warning systems can be. WIRED points to conflicting wildfire evacuation orders in California in 2016, confusion around Australia’s 2019 “Watch and Act” wildfire messaging, and the Montecito mudslide after the Thomas Fire, where warning fatigue and unclear evacuation guidance helped produce devastating consequences. The Santa Barbara County mudslide killed 23 people and injured more than 160.
How do agencies say AI helps?
Supporters argue that AI can strengthen alerts by removing clutter and surfacing the key facts faster. Emergency managers often write under pressure, while also juggling incomplete information, exhausted staff and the need to issue warnings before a situation worsens. In that environment, a chatbot can act like a fast drafting assistant.
Jeannette Sutton, a disaster researcher at the University at Albany, said the pressure on public servants is intense. As she put it, “They’re just trying to get their job done without screwing something up.”
That is the practical case for AI in emergency communications: not replacing the human decision-maker, but helping them produce a clearer first draft, compare language across alerts and identify missing context. Some advocates say AI could also help agencies sort through large numbers of reports, research and prior messages more quickly than a human team can do alone.
Even so, the best-case version still depends on human review. The current expert consensus described by WIRED is that AI is useful only when a trained emergency professional supervises the output. In public safety, a faster draft is not valuable if it carries the wrong hazard details, the wrong geography or the wrong urgency.
What the reporting says about real-world use
Official policy has not caught up with unofficial behavior. WIRED reports that most agencies ban employees from using generative AI, yet experts believe it is already widespread inside state and local disaster offices. Some emergency managers are reportedly using ChatGPT or Gemini on personal or free accounts to produce high-quality draft alerts before revising them for official use.
One public example of cautious use comes from Ingham County, Michigan. Emergency manager Rob Dale has used AI to create draft practice alerts, but not to issue public warnings, according to WIRED. That distinction matters: training with AI is one thing, but sending AI-generated text to the public is another.
There is also a new commercial market forming around this workflow. EM1 says its AI system already performs better than nearly all human emergency managers on benchmark tests. The company says it has contracts with about 100 government agencies and school districts, though none are using it to automate extreme-weather alerts yet.
That leaves agencies in a middle ground. They are experimenting, sometimes quietly, but have not crossed the line into full automation for life-safety messages. For now, the technology appears to be functioning as a drafting aid rather than a replacement for judgment.
Why are emergency alerts so hard to get right?
Because public warnings have to do several things at once: be brief, accurate, actionable and calm. If they are too technical, people do not understand them. If they are too vague, they do not know what to do. If they are too repetitive, they can trigger warning fatigue and be ignored.
WIRED cites a review by Sarah Sutton of a New Mexico flood alert that found several clarity problems in the original wording. That kind of critique gets at one of the core frustrations in emergency management: many alerts are written in jargon, omit context or include details that do not help a resident decide whether to evacuate, prepare or stay indoors.
Those problems are often made worse by staffing constraints. Limited personnel and time pressure can push agencies to prioritize getting a warning out only after the situation is already deteriorating. By then, the message may be rushed, inconsistent or so generalized that it fails to move people.
AI, in theory, can reduce some of that friction. But if the model is trained on flawed examples, or if the user does not know enough to catch errors, the result may be a slicker version of the same problem.
What is the trust risk?
The trust problem may be the biggest reason experts are not ready to let AI speak for agencies. A warning that sounds machine-generated could feel impersonal, especially when people want to hear from a local official who understands the specific threat. Ginny Katz, founder of HazAdapt, told WIRED: “AI-generated content does not feel authentic.”
That authenticity gap matters because a warning is not just information; it is a signal that someone competent is paying attention. If residents think an alert was produced by a chatbot, they may question whether anyone actually checked it. In communities already skeptical of government, that doubt can spread quickly.
There is also the risk of overproduction. If AI makes it easier to issue warnings, agencies might send more of them. That could help in some situations, but it could also deepen warning fatigue, making people less responsive to later emergencies.
In other words, the efficiency gain could become a communication penalty if agencies use the technology to scale volume without improving judgment. A flood of alerts is not the same thing as better alerting.
Background: how disaster warning problems built up
The current debate did not begin with chatbots. Emergency communication has long been difficult because the job sits at the intersection of science, bureaucracy and human behavior. Officials have to make decisions with partial information, then turn those decisions into public instructions that ordinary people can understand in seconds.
WIRED points to several warning breakdowns that illustrate the challenge:
- In 2016, California counties issued conflicting wildfire evacuation orders.
- In 2017, the Thomas Fire aftermath and the Montecito mudslide brought fatal consequences, killing 23 people and injuring more than 160.
- In 2019, Australia’s “Watch and Act” wildfire messaging created confusion.
- In January 2025, erroneous evacuation alerts in the Los Angeles area sparked panic.
These events show that the problem is not new and not limited to one country or one storm type. Whether the threat is fire, flood or hurricane, the weakness is often the same: people do not receive a message that is timely, clear and consistent enough to act on immediately.
That is why some supporters see AI as an aid to better writing rather than a radical change in policy. If the biggest issue is clarity, they argue, a drafting tool might help strip out jargon and surface the basics faster. Critics respond that clarity is not the same as correctness, and in an emergency, correctness is the non-negotiable part.
By the numbers
| Figure | What it refers to | Why it matters |
|---|---|---|
| 100 | Government agencies and school districts with EM1 contracts | Shows how quickly AI vendors are entering public-safety workflows |
| 23 | People killed in the Montecito mudslide | Illustrates the cost of unclear or delayed warning systems |
| 160+ | People injured in the Montecito mudslide | Shows the scale of harm that can follow poor evacuation messaging |
| 2016 | Year California counties issued conflicting wildfire evacuation orders | A reminder that confusion in alerts predates generative AI |
| 2019 | Year Australia’s “Watch and Act” wildfire messaging caused confusion | Evidence that warning language itself can fail even when the hazard is clear |
| January 2025 | Los Angeles-area evacuation-alert panic | Shows how quickly false or confusing alerts can spread fear |
What happens next?
The key question now is not whether AI will appear in emergency offices — it already has — but how far agencies will let it go. WIRED’s reporting suggests the immediate future is likely to be cautious and hybrid: AI drafts, humans approve.
Several open questions will shape the next phase:
- How widely are agencies actually using AI for public-facing alerts versus internal drafting?
- Will any agency soon automate weather or disaster alerts with AI in production?
- How will public trust change if recipients know a warning was AI-written?
- Will regulators or agencies create formal rules for AI use in emergency communications?
Those questions matter because emergency messaging is a high-trust function. Even when AI helps write a cleaner alert, a human still has to own the decision, verify the facts and judge whether the message will help people act.
For now, the evidence points to a narrow but growing use case: AI as a behind-the-scenes drafting tool in an environment where speed matters, but mistakes can cost lives. The balance between those two pressures will determine whether the technology improves disaster communication or makes it less believable when the next warning arrives.
How officials are thinking about the tradeoff
Rob Dale’s practice-only use in Michigan suggests one cautious model: let AI help people train, rehearse and refine alerts without giving the system authority to warn the public directly. That approach preserves the human chain of responsibility while still capturing some of the drafting speed that makes the technology appealing.
At the same time, the fact that some officials are reportedly using consumer chatbots on personal accounts shows how quickly unofficial practices can outrun formal policy. Agencies may ban generative AI, but if staff are under pressure and the tools are easy to access, bans alone may not stop use.
That is why the broader debate is less about whether AI is available and more about whether public agencies can create guardrails before convenience becomes dependency. Emergency managers want faster drafting, but they also need to protect the credibility that makes public warnings work in the first place.
The central lesson from the cases cited by WIRED is that technology does not erase the hardest part of the job. People still have to decide what matters, how serious the threat is and whether the public will understand the message in time. AI can assist that process. It cannot be allowed to substitute for it.
Sources and reporting context
This article is based on reporting by WIRED, which examined how emergency managers are using AI chatbots for disaster alerts and how experts in public safety, disaster research and emergency management are responding to the shift. The reporting also references earlier coverage and examples involving Grist, Alabama, Michigan, California, Australia and New Mexico.
Across those cases, the pattern is consistent: AI may help officials write faster, but public safety agencies cannot afford to discover the limits of a chatbot in the middle of an evacuation.
Frequently asked questions
Are emergency managers already using AI to write disaster alerts?
Yes. WIRED reports that many emergency managers are already using AI chatbots to draft disaster alerts, even though most agencies officially ban generative AI. The use appears to be most common for internal drafting rather than direct public issuance.
Why are experts worried about AI disaster alerts?
Experts are worried because AI can produce inaccurate, overly polished or impersonal warnings that may reduce trust. In emergencies, credibility is crucial, and a mistaken alert can make people ignore future warnings or act on the wrong information.
Can AI make emergency alerts better?
Yes, but only in a limited way. Supporters say AI can help draft clearer warnings faster by summarizing information and removing jargon. The reporting says the current expert view is that a human emergency professional must supervise every output.
Have any agencies started automating alerts with AI?
Not yet, according to the reporting. EM1 says it has contracts with about 100 government agencies and school districts, but none are using it to automate extreme-weather alerts, which suggests agencies are still testing the technology cautiously.
What happens if public trust drops in emergency warnings?
If trust drops, fewer people may evacuate or take shelter when warned, which can increase injuries and deaths. Emergency managers depend on the public believing their alerts, so losing that credibility can make even accurate warnings less effective.
