Disaster AI fails before the model runs

The bottleneck in AI for emergency management sits before any model runs. A new RAND landscape assessment, commissioned by the Markle Foundation, finds that the field's failure mode is not capability. It is evaluation. Wildfire coordinators, hurricane planners, and flood responders are being sold tools faster than their agencies can tell which ones work, and RAND's actual argument is about the layer above the algorithm.

Most products in the AI-for-emergency-management market are general-purpose AI tools repurposed for disasters. Roughly half are purpose-built or have documented emergency-management use cases. The rest are adjacent. Either way, a county emergency manager shopping for the next incident tool has no shared performance benchmarks, no standard procurement documentation, and no clearinghouse to compare them. RAND's own framing makes the point: prior literature catalogues what AI could do. Almost none catalogues what is actually available, or what adoption requires.

So the reusable mechanism is procurement-shaped. When a buyer cannot evaluate, the cheapest products win on demos, not on outcomes. Disasters do not pause for that.