America’s wildfire risk data quietly puts millions of homes in danger

A lot of us might assume that most homes that are destroyed by wildfires were in obvious, high fire-risk areas, like on the edge of forests that frequently burn. But wildfires are a faster-growing and much closer threat than we may realize — burning in places that rarely used to see them.

For instance, many homes that remain in the neighborhoods that burned in the historic Los Angeles wildfires last year are still considered as having “low risk” in assessments from the Federal Emergency Management Agency (FEMA) despite the charred remains of their neighbors showing how vulnerable they might be to embers blowing from miles away. 

It raises an urgent question: Do we actually know which homes face the most danger of burning? 

Government risk maps are too coarse for the way wildfire works now. But new tools powered by AI are giving us a clearer picture. They could reshape how we understand the dangers that lie ahead and force a reckoning over where we live and how we build and protect our homes — if we choose to listen.

How AI helped risk modelers zoom in

For decades, modelers calculated wildfire risk by looking at historical patterns of wildfires, but it’s increasingly evident that this vastly understates the scope of the problem. In fact, until 2023, California prohibited insurers from using forward-looking catastrophe models that included factors like future climate change to set their rates. “Wildfires have very complex dynamics, and a backward-looking approach is not sufficient,” said Firas Saleh, director of North America wildfire models at Moody’s, a financial analytics firm. 

Now, one company, ZestyAI, says they have a new model that fills in “blind spots” in the government’s fire risk calculations, providing a sharper picture of the threats wildfires can pose to individual homes.   

“Early on in our journey, we realized that insurance companies were writing property insurance without having a deep understanding of the properties themselves,” said Kumar Dhuvur, chief product officer at ZestyAI. “A lot of times, their way to get that understanding was to ask agents or the homeowners, ‘Hey, do you have a tree next to your house? Do you have a swimming pool?’”

To produce their estimates, ZestyAI used satellite images of neighborhoods to examine structures, vegetation, and terrain. They combined this information along with historical fire records and climate variables to train their AI model. This allowed them to calculate risks for specific houses.

“You’ve got to be very granular in your assessment of risk,” Dhuvur said. “There could be whole neighborhoods where the resolution is too low and becomes a no-go zone for an insurance company.”

When using its model to analyze the regions burned in Los Angeles wildfires last year, for example, ZestyAI found that more than 3,000 properties that were labeled as low or no risk in assessments from FEMA showed up as having an elevated fire risk in ZestyAI’s model. These properties have an estimated value of $2.4 billion. Across California, there are 1.2 million properties worth around $940 billion that were labeled as low risk in FEMA’s National Risk Index. ZestyAI found all of them to face greater danger.

It’s an alarming result when you put it in the greater context. Sprawling, destructive wildfires are extracting a massive and growing toll from the global economy. In 2025, fires around the world burned through 390 million hectares — more than 90 percent of the land area of all the countries in the European Union. The price tag of wildfires has been surging in recent decades, and the Los Angeles wildfires last year may be the most expensive disaster in US history.  

This is a big jump forward from conventional fire risk models. 

FEMA’s National Risk Index, for example, calculates threats over census tracts or counties (The National Risk Index has now been migrated into the new Resilience Analysis and Planning Tool). The dataset groups high-risk and low-risk homes together in ways that miss a lot of important differences between them. Some houses may have fire-resistant shingles and a wide defensible space that give them more protection. Others may have shared wooden fences with neighbors that create pathways for fire to travel, leaving those homes vulnerable to fires that start far away. 

That was clear in the aftermath of the Los Angeles fires, when some homes were left standing despite just about the rest of the neighborhood turning to ash. 

For its part, the Federal Emergency Management Agency told Vox that the National Risk Index is intended to be a baseline, not an absolute measure of risk. 

“FEMA welcomes efforts by others to develop additional datasets that support communities in preparing for all hazards, including wildfires,” a FEMA spokesperson wrote in an email. “Increased research and data collection on risks enables communities to enhance their preparedness and resilience before disasters happen.”

But even though ZestyAI’s analysis provides a clearer picture of which homes are in danger, it can still leave some important complexities of wildfires, said Hussam Mahmoud, who leads the Vanderbilt Center for Sustainability, Energy and Climate and studies risks to communities. 

Fire risk is not just a function of individual homes but of how whole neighborhoods and environments interact. A group of homeowners might clear a wide defensible space around their own homes, upgrade their sidings, and protect attic vents from cinders, but if one of their neighbors falls short, it could endanger the whole community when flames arrive. Even fire-resistant homes that meet upgraded construction codes can burn if they are pummeled for hours with waves of embers on hurricane-force winds, as the 2025 Los Angeles fires showed. 

“I think AI is a very promising technology,” Mahmoud said. “It has limitations to how it can be used with a physics-based model.”

There are also tradeoffs between how precise risk estimates can be and how much they cost. Inspecting individual homes in person can yield the sharpest picture, but it’s intrusive, time-consuming, and expensive to send people to examine millions of homes. And in-person inspections still don’t tell the whole story. 

“When you’re on the ground assessing buildings and looking if the building has good roof material versus good siding versus something else, you’re assuming that this building is a recipient of fire,” Mahmoud said. “You’re not looking at how the fire is propagating across the community.”

Wildfire risks aren’t just increasing. They’re evolving.

The story we tell about wildfires is shifting. They are not just a problem in wilderness areas that happen to spill over into cities and neighborhoods. Houses are now as much a part of the landscape as pine trees and chaparral. They are both fuels and sources of ignition, even far from forests and shrublands. The vast majority of wildfires are ignited by human activity, and when entire neighborhoods ignite, fires behave in hard-to-predict ways not seen in nature. That was evident as the Los Angeles fires last year engulfed coastal mansions in Pacific Palisades and entire neighborhoods in downtown Altadena. 

But better models like ZestyAI’s can make a difference — if we’re willing to make hard decisions and act on them.  

Some communities have already used specialized fire forecasts to target certain properties with mitigation measures to reduce their odds of igniting and then leveraging that to lower their insurance rates. And with regulatory reforms like allowing insurers to use fire models that look ahead, California is starting to lure some insurance companies back to the state.

The biggest challenge may yet be getting people to acknowledge their risks at all. 

The home listing site Zillow last year decided to remove climate risk scores from property listings under pressure from California real estate groups that complained that the scores were hurting the resale value of some homes. It makes sense: A better map of fire dangers might not be in your interest if you’re trying to sell your home and its value suddenly drops because it shows up as having a higher risk of igniting. An insurance company might also use that information to raise your premiums or drop your coverage entirely. 

There are more reasons why people might not want to think too hard about future fires. Faced with an urgent housing shortage, Los Angeles is under immense pressure to build as much as possible, as fast as possible. Yet despite all the efforts to speed up construction, especially in the wake of the devastating wildfires last year, building in Southern California is still an agonizingly slow process.

Anyone with money and time who has lost their home in a fire can afford to wait to assess their risks and rebuild their homes to be more resilient, or move. However, many lower-income fire victims don’t have a choice other than to try to go back to the same conditions that put them in danger in the first place. That’s part of why there have been more permit applications to date for rebuilding in low and middle-income communities — like Altadena, for example — that burned last year, and fewer in wealthier enclaves like Pacific Palisades. 

“Families that are displaced from Palisades do have the wealth and means to look for alternatives as opposed to Altadena residents, for whom that’s their only option,” said Minjee Kim, an assistant professor of urban planning at the University of California Los Angeles. 

AI and more advanced models can help us predict these risks and understand them better, but no algorithm can extinguish financial denial or do the political heavy lifting required to stop us from building tomorrow’s homes in burn zones that are only getting bigger.