Alert! Crafting Evidence-Based Warning Messages With Generative AI

Many emergency managers have seen or sent alerts like the below that missed the mark: acronyms that mean nothing to most people, unclear protective actions, and jargon that wastes precious characters while confusing recipients.

RCEMA ALERT 08:15 09/17/25: LVL 3 HAZMAT INCIDENT following train derailment at MP 28.3 BNSF corridor. Active release of Anhydrous Ammonia (NH3) confirmed. ERZ established for sectors 7A & 7B, generally N of Highway 54 & W of I-49. All persons in ERZ must initiate SIP protocols immediately. DO NOT ATTEMPT TO EVACUATE. Monitor EAS and local media for further instruction.

Alert sent by simulation county, Riverside County, Missouri, 2025.

In a dire situation, this kind of alert could cost lives. However, alerts like these are not outliers – they are symptoms of a systemic challenge.

Emergency managers know the scramble: a disaster hits and suddenly alerts need to go out across a dozen different platforms – each with its own quirks, limits, and rules. In the middle of the chaos, staff are juggling the requirements of the Integrated Public Alert and Warning System, (IPAWS): 90-character wireless emergency alerts (WEA), 360-character extended WEA alerts, broadcast Emergency Alert System (EAS) messages, commercial mass messaging platforms, and social media posts. One small formatting slip and a message that will reach thousands can cause more harm than good.

Effective tools, such as the Federal Emergency Management Agency’s (FEMA) Message Design Dashboard (MDD), provide substantial evidence-based guidance on emergency alerts and warnings for all hazards to those who write alerts. However, the scale of the challenge is huge. FEMA’s public alert system now links more than 1,800 agencies nationwide, all handling their own emergencies with different resources and expertise. The system itself works since most broadcasters passed the Federal Communications Commission’s 2024 test, but writing and sending those alerts remain too complex and slow in practice. The knowledge base and messaging infrastructure exist, but until recently, the tools available were limited to alert writing training, generic templates, and best intentions. Most of these tools are also focused solely upon cell phone–based alerts, while most emergency managers send alerts and warnings on a wide variety of platforms.

Many emergency managers are still learning how to effectively use artificial intelligence (AI), especially large language models, for complex tasks like alert generation. One challenge is that these models were trained on publicly available data containing examples of alerts that experts would consider substandard. Because large language models generate text by reproducing patterns found in their training data, they can easily replicate those weaknesses. Achieving high-quality, evidence-based alerts, therefore, requires careful prompt and context engineering guided by domain expertise.

Although emergency management professionals are often initially skeptical of AI, their willingness to adopt it grows significantly with hands-on experience. Adoption offers clear benefits, such as greater efficiency, clearer messaging, and real-time data integration. However, concerns about accuracy, context, and cultural sensitivity persist.

Developing AI tools that can generate strong messages and flag weaker ones could ease the cognitive burden on message creators, enabling them to focus on higher priorities while delivering faster, more effective communications. The priority should be developing practical tools that emergency management professionals can actually use – but how to get there matters just as much as what is created. By aligning technology, expertise, and will, companies can better meet those needs and build solutions that last.

For the rest, click here.