Wildfire AI Is Moving From Prediction Toward Coordinated Action

A wildfire model can now turn live camera feeds, weather, terrain and fuel conditions into a spread map within minutes. During the Palisades Fire on January 7, 2025, the University of California San Diego’s WIFIRE team says its first predictive model reached Los Angeles responders within minutes of the first sign of smoke. Yet Los Angeles County’s official review of public alerts, warnings and evacuations—not the full suppression response—found outdated policies, inconsistent practices and communications vulnerabilities. Better prediction did not automatically produce better coordination.

That gap is the subject of a new peer-reviewed perspective in *Communications Earth and Environment*. The authors argue that agentic AI could help agencies allocate scarce crews, aircraft and equipment across simultaneous fires, rather than optimizing each incident in isolation. Their proposed system would connect local incident commands, regional operations and state or federal coordinators through a shared objective and a human approval layer.

The idea is important—and easy to overstate. This is a governance proposal, not a deployed wildfire command system. It reports no controlled operational trial, no measured reduction in acres burned and no evidence that an AI allocation recommendation has improved a live incident outcome. Its real contribution is to name the bottleneck: seeing the fire is not the same as deciding, across jurisdictions, who should send what, where and when.

Research references: [2] [3] [1]

Prediction has moved closer to the fire line

Wildfire intelligence used to arrive slowly. The WIFIRE program describes a system that now combines weather forecasts, ground sensors, satellite imagery, historical fire records and camera data to model how a fire may spread. Its Firemap outputs feed official channels through California’s Fire Integrated Real-Time Intelligence System, or FIRIS. Aircraft can add infrared perimeter measurements when conditions permit.

For the Palisades Fire, a 911 call at 10:30 a.m. triggered an automatic alert to the WIFIRE team. The team says it used the ALERTCalifornia camera and sensor network to triangulate the ignition and issue an initial predictive model within minutes. Over the three-day wind event, WIFIRE says it addressed about 230 fire starts and published models for 12 verified fires considered capable of becoming large. Those are implementer-reported operational facts, not an independent evaluation of lives or structures saved.

The physical sensing network is substantial. UC San Diego reported more than 1,140 ALERTCalifornia cameras in early 2025, including 115 cameras at 52 sites in Los Angeles County. The cameras can pan, tilt and zoom, operate at night with near-infrared imaging and provide a shared visual layer to public-safety agencies.

These capabilities improve the “state-estimation” problem: where is the fire, how fast is it moving and what could it reach? They do not solve the allocation problem that emerges when several incidents compete for the same engines, crews, aircraft or evacuation support.

Research references: [2] [4]

Wildfire response is a network of decisions

The United States already has a three-tier mobilization system. More than 250 local dispatch centers coordinate initial response. Ten geographic-area coordination centers can mobilize resources across broader regions. When those regions are strained, the National Interagency Coordination Center moves resources nationally, while the National Multi-Agency Coordinating Group sets national priorities.

This hierarchy exists because local optimization can weaken the whole system. A commander has strong reasons to request enough resources for the fire in front of them. But under simultaneous incidents, the last crew or aircraft assigned locally might prevent more damage elsewhere—or might need to remain available for a new ignition. The value of that decision depends on information and priorities that no single incident command possesses.

The Los Angeles County after-action review illustrates the operational stakes without reducing the disaster to one failure. Six fires ignited between January 7 and January 9, 2025. The Eaton and Palisades fires together burned about 37,000 acres, destroyed more than 16,000 structures and killed 31 people. Nearly 250,000 residents were placed under evacuation warnings or orders. The review examined more than 150,000 alert and evacuation records, unit-location data, responder interviews and community listening sessions. It found no single point of failure; it identified a series of weaknesses across policy, practice and communications.

The County’s recommendations included clearer roles, stronger emergency-management staffing, standardized training, interoperable communications, field-accessible dashboards and more reliable satellite and internet connectivity. These are prerequisites for any AI coordination layer. A model cannot reconcile unclear authority, repair a radio network or create mutual-aid agreements by itself.

Research references: [5] [3]

What an agentic coordination layer would do

The new perspective proposes a multi-agent architecture mapped to the existing command structure. Local incident-command agents would represent the conditions and requests at individual fires. Regional agents would combine incidents across a geographic area. A central state or federal agent would evaluate competing requests against a shared reward function intended to minimize total damage and deployment cost. Human decision-makers would approve, override or reject the plan.

In principle, such a system could continuously compare alternative allocations as forecasts change. It might notice that an aircraft is producing diminishing benefit at one incident while a new ignition has crossed a threshold elsewhere. It could identify replacement resources needed to keep a community covered when units deploy under mutual aid. It could also expose the system-wide cost of keeping equipment in reserve rather than treating reserves as unused capacity.

This is different from today’s predictive AI. A spread model estimates what the fire may do. A coordination agent would recommend what organizations should do together. The latter must represent not just weather and fuels, but travel time, crew qualifications, aircraft restrictions, water access, fatigue, communications, local coverage, legal authority and the uncertainty of suppression effectiveness.

That last input is a major constraint. The perspective notes that reliable empirical evidence about how different firefighting resources perform across diverse conditions remains limited. An optimizer is only as credible as the consequences it can estimate. If the system cannot quantify the benefit of sending a particular resource to a particular incident, a sophisticated allocation may still rest on fragile assumptions.

Research references: [1]

The evidence boundary is unusually important

No evaluated outcome in the perspective establishes that agentic AI improves wildfire coordination. The architecture, shared reward, governance hierarchy and staged deployment path are design targets. The operational wildfire platforms discussed by the authors provide evidence that rapid detection and forecasting are possible, but claims about their coverage or performance are reported by their developers or operators unless independently evaluated.

The proposed reward function also encodes values. “Minimize total damage” sounds neutral until damage must be measured across lives, homes, ecological resources, cultural sites, critical infrastructure, firefighter exposure and communities with unequal capacity to recover. A single numerical objective could make those trade-offs less visible rather than more legitimate.

Central coordination can reduce resource hoarding, but it can also concentrate power. Local commanders may know road, terrain or community conditions that a central system misses. Tribal nations, counties, utilities and federal land agencies operate under different authorities. A system that optimizes across them needs prior agreements on data, objectives, appeals and accountability—not merely API access.

Research references: [1]

Human oversight must be operational

The paper argues for a human-in-the-loop design. That phrase should mean more than placing an approval button after an automated recommendation. The reviewer needs enough time, information and authority to understand the recommendation, see its uncertainty, compare alternatives and stop the system safely.

The European Union’s AI Act requires human overseers of covered high-risk systems to understand capabilities and limits, monitor anomalies, recognize automation bias, interpret outputs and disregard or override a recommendation when appropriate. That law does not necessarily govern a California wildfire deployment; its Article 14 is used here as a governance benchmark. The United States Department of Homeland Security’s nonbinding framework for AI in critical infrastructure similarly emphasizes responsibility across system design, data governance, safe deployment, testing and continuing performance monitoring.

For wildfire coordination, meaningful oversight would require at least five things: a traceable explanation of why resources were ranked; visible uncertainty and stale-data warnings; a way for local commands to add missing facts; explicit override and safe-stop procedures; and a record suitable for after-action review. The system must also degrade safely when cameras, aircraft, cellular service or power fail—the conditions most likely during an extreme event.

Research references: [6] [7]

A safer path begins as a shadow system

The most credible near-term deployment is not autonomous command. It is a shadow mode that watches the same incidents as human coordinators without controlling resources. Its recommendations can be compared with real decisions and outcomes across fire seasons. Researchers can then measure whether it identifies conflicts earlier, reduces decision time, preserves local coverage and avoids unsafe or inequitable recommendations.

A second stage could present ranked options to coordinators with uncertainty, constraints and counterfactuals: what changes if one aircraft is unavailable, travel time doubles or a new ignition appears? Only after prospective evaluation should an agent be permitted to execute bounded administrative steps—and even then, operational command should remain with accountable officials.

The breakthrough in wildfire AI may therefore be a shared decision layer that helps humans see the system-wide consequences of their choices while keeping responsibility visible. The technology to model the fire is moving quickly. The institutions needed to coordinate around those models will determine whether that intelligence becomes resilience. The Perspective reports support from a Cambridge Humanities Research Grant, the Centre for Human-Inspired AI and the CRASSH climaTRACES Lab, and its authors declare no competing interests.

Research references: [1]