Tool · Warning-message structure
Message Element Analyzer
A complete warning message answers five questions: who is sending it, what the threat is, where it is, what to do, and when. This tool checks each verbatim first-burst campus alert for the presence of those five elements and shows how the patterns vary across incident types and over time.
These five are not our invention. They are the message-content elements that hazards and risk-communication research has identified as the components of a complete warning. The set (source, hazard, location, guidance, time) is the Warning Response Model that Dennis Mileti and John Sorensen synthesized for FEMA in 1990 after reviewing more than 200 studies of how people respond to warnings. The same structure is reflected in the CDC’s Crisis and Emergency Risk Communication (CERC) guidance and is the content structure FEMA recommends for Wireless Emergency Alerts. Sources are listed below so you can verify this yourself.
This page records only the binary presence or absence of each element. It does not score, weight, or rank messages, and it does not reproduce any specific published rubric or scoring instrument.
Sources for the five-element framework
- Mileti, D. S., & Sorensen, J. H. (1990). Communication of Emergency Public Warnings (ORNL-6609). Oak Ridge National Laboratory, for FEMA. The foundational synthesis; defines the message-content components. OSTI record.
- CDC. Crisis and Emergency Risk Communication (CERC) Manual (2014, upd. 2018). Its “Be Correct” and “Promote Action” principles map to naming the hazard and giving guidance. CDC CERC.
- FEMA IPAWS. Best Practices for Alerting Authorities Using Wireless Emergency Alerts. Recommends a WEA carry source, hazard, location, protective action, and time/expiration. FEMA guide (PDF).
- Lindell, M. K., & Perry, R. W. (2012). The Protective Action Decision Model. Risk Analysis, 32(4):616-632. The theory of why message content drives protective behavior. PubMed.
- National Academies of Sciences, Engineering, and Medicine (2018). Emergency Alert and Warning Systems. National Academies Press. Free full text.
How the coding was produced
In plain terms: rather than rely on a single opinion, we asked the same AI (Claude) to read each alert from scratch 25 separate times, with no memory of the other reads. Each time, it decided whether the alert answers each of the five questions and wrote a one-sentence reason. Think of it as 25 quick second opinions, with every opinion’s reasoning shown so you can check it. The more of the 25 that agree, the more settled the call.
More precisely: each verbatim first-burst alert was judged by 25 independent Claude Opus 4.8 passes. Every pass gave its own present or absent verdict for each element and a one-sentence justification, reading only the alert text. An element is marked present by majority vote; the agreement level you see (for example 23/25) is how many of the 25 passes agreed.
There are no keyword rules. Open any element to read all 25 passes’ verdicts and reasons, so you can audit a judgment and disagree with it. If you think a call is wrong, the reasoning is right there to check.
This is systematic, repeated AI effort at coding accuracy. It is not human-reviewed and is not validated research coding. Coding used only the alert text, never incident-type metadata. Treat it as a strong starting point and verify against the primary source.