In Altadena, California, a 1,750-square-foot ranch house built in 1958 burned to its foundation on January 8, 2025. The Eaton Fire consumed it in under an hour. Fourteen thousand acres. Nearly 7,000 housing units destroyed. On that same parcel, First Street Foundation’s Fire Factor had assigned a risk score of 9 out of 10—a property with a greater than 26% cumulative probability of wildfire exposure over 30 years. The score had been publicly available since 2022.
Nobody checked it before the house was sold. Nobody checked it before the previous renovation. Nobody checked it during any of the four real estate transactions the property had undergone since First Street published its data. The soil report was on file. The termite inspection was on file. The fire risk score existed in a database that anyone with an internet connection could access for free, and it sat there, untouched, while a family bought a home on a hillside that an algorithm had flagged as near-certain to burn.
The Score That Insurance Companies Already Use
First Street Foundation assigns a Fire Factor score of 1 to 10 for every property in the United States—over 140 million parcels. The model incorporates satellite-derived vegetation density, historical fire perimeters, topographic slope analysis, local climate projections, and building material assessments. It runs 30-year forward projections accounting for climate change. The methodology is peer-reviewed and published in their documentation.
The numbers are sobering. According to First Street’s 5th National Risk Assessment: 20.2 million US properties carry moderate wildfire risk, 6 million carry major risk, 2.7 million carry severe risk, and 1.5 million carry extreme risk—a cumulative burn probability exceeding 26% over three decades.
Cape Analytics, now owned by Moody’s after a January 2025 acquisition, takes a different approach: computer vision applied to aerial and satellite imagery. The system identifies building materials, roof condition, vegetation encroachment, and defensible space compliance at the individual address level. Insurers use Cape’s data to make underwriting decisions—which is a polite way of saying they use it to decide whether to offer you a policy at all.
That’s the asymmetry. Insurance carriers have been using AI wildfire risk models for years. Homebuyers and builders, largely, have not. The data is public or purchasable. The analytical tools exist. The gap is adoption—and the consequences of that gap landed on 12,585 families in Los Angeles in January 2025.
What $13,000 Buys You
A December 2025 study by Headwaters Economics and the Insurance Institute for Business & Home Safety examined the actual cost of building a wildfire-resistant home from scratch. Their reference case: a one-story, 1,750-square-foot house in Southern California with a $500,000 construction budget. The findings:
| Standard | Additional Cost | % Premium |
|---|---|---|
| California Chapter 7A (CWUIC) | $13,000 | 2.6% |
| IBHS Wildfire Prepared Home | $9,000 | 1.8% |
| IBHS WFPH Plus (enhanced) | $15,000 | 3.0% |
Fifteen thousand dollars. On a half-million-dollar build, that’s the cost of upgrading the kitchen countertops from laminate to quartz. It buys you ember-resistant vents, a Class A fire-rated roof assembly, noncombustible gutters, enclosed eaves, and tempered glass in exterior windows. It buys you a house that, when paired with defensible space landscaping, has a meaningfully higher probability of surviving a wildfire than the one next to it that was built to minimum code.
The CoreLogic 2024 Wildfire Risk Report found that combining individual property mitigation with community-wide defensible space can reduce expected wildfire loss by 75%. Seventy-five percent loss reduction for a 3% construction premium. Run the math on a $500,000 home in a Fire Factor 7 zone with a 14% cumulative burn probability: the expected loss over 30 years without mitigation, using CoreLogic’s median damage estimates, is roughly $70,000. With mitigation, that drops to approximately $17,500. The $15,000 investment pays for itself once over 30 years even before accounting for insurance premium savings.
Why the Code Doesn’t Require What the Data Demands
California Chapter 7A applies only in designated WUI zones—areas where wildland vegetation meets urban development, as mapped by CAL FIRE. After the January 2025 fires, a broad coalition of fire scientists, insurers, and the California Building Industry Association urged Los Angeles to apply Chapter 7A to all reconstruction in Pacific Palisades and Altadena. The code was already required in Palisades. It was not required in Altadena.
The irony is structural. WUI maps are redrawn infrequently and based on vegetation surveys, not predictive fire behavior models. A parcel can sit outside the WUI boundary—exempt from Chapter 7A—while carrying a Fire Factor score of 8 or 9. AI risk models update continuously. Building code jurisdiction maps do not. The Altadena homes that burned were not in a designated WUI zone. They were, by every computational fire risk model available, in a zone of severe to extreme wildfire exposure.
This is the regulatory gap that AI has exposed but that building codes have not yet absorbed. CAL FIRE’s WUI maps are binary: you’re in or you’re out. First Street’s Fire Factor, Cape Analytics’ property assessments, and newer platforms like FireScore.ai operate on continuous risk gradients with parcel-level precision. The code gives you a boundary line. The AI gives you a probability distribution.
The Insurance Market Already Made the Decision
While building codes lag behind AI risk data, insurers do not. State Farm paused new homeowner policy applications in California in May 2023. Allstate had already stopped. By late 2024, the California FAIR Plan—the state’s insurer of last resort—had seen enrollment surge as private carriers withdrew from fire-prone areas. California’s Proposition 103 reforms now allow insurers to use forward-looking catastrophe models, including AI and machine learning-based wildfire projections, in their rate-setting.
For new construction, this creates a stark reality: you might be able to get a building permit on a lot where you cannot get homeowner’s insurance at any reasonable price. The permit office checks the WUI map. The insurer checks an AI risk model. The two systems are looking at different data, drawn at different resolutions, updated at different frequencies, and reaching different conclusions about the same parcel of dirt.
If you’re building new and haven’t checked your lot’s Fire Factor score before you pour the foundation, you’re doing less due diligence than your insurance company did before deciding whether to cover you.
What This Means If You’re Building
Check the score first. First Street’s Risk Factor is free for any address. Cape Analytics data flows through insurance carriers. FireScore.ai and other platforms offer parcel-level assessments. Look at the 30-year projection, not just the current snapshot—climate models show wildfire risk expanding eastward and into areas historically considered low-risk.
If your lot scores above a 5, build to at least IBHS Wildfire Prepared Home standard regardless of whether Chapter 7A applies. The $9,000 premium on a $500,000 home is 1.8%—less than most contingency budgets. If you’re in a high-risk zone, the enhanced WFPH Plus standard at $15,000 adds enclosed eaves and noncombustible gutters that address the two most common ember entry points in wildfire structural ignition.
And if you’re buying a lot: run the fire risk score before you run the soil test. The soil report tells you what kind of foundation to pour. The fire risk score tells you whether the house will be standing in 30 years—and whether anyone will insure it.
What This Analysis Didn’t Prove
The ROI calculation above uses simplified assumptions: CoreLogic’s 75% mitigation figure comes from a combined individual-plus-community intervention, not individual hardening alone. A single hardened home surrounded by non-hardened neighbors captures only a fraction of that benefit—IBHS research suggests individual mitigation alone reduces risk by roughly 40-50%, not 75%. The $70,000 expected loss figure is a rough estimate based on median damage ratios applied to the construction cost; actual losses vary enormously depending on fire intensity, wind speed, suppression response, and whether the fire reaches the structure as ember exposure or direct flame contact.
First Street’s Fire Factor uses climate projections that carry their own uncertainty bands. The 30-year burn probability for a parcel scoring 7 could reasonably range from 8% to 20% depending on the emissions scenario. The scores are useful as relative risk indicators, not as actuarial certainties.
This analysis also does not address the equity dimension: the highest-fire-risk parcels in California tend to cluster in two categories—expensive hillside properties and lower-income rural communities. A 3% construction premium is trivial on a $500,000 build. On a $150,000 manufactured home in a rural WUI area, $9,000 is a different conversation entirely.
The Strongest Case Against
AI wildfire risk scores could accelerate redlining by another name. If insurers use these scores to withdraw coverage and lenders follow by refusing mortgages, entire communities could become financially unbuildable—not because the risk is unmanageable, but because the market decided the risk isn’t worth managing. Paradise, California, rebuilt after the 2018 Camp Fire. If AI risk scores had existed in their current form in 2019, it’s an open question whether insurance availability would have permitted that rebuilding at scale.
There’s also the accuracy question. AI models are trained on historical fire data and vegetation imagery. They struggle with recently cleared lots, new defensible space landscaping, and community-wide fuel reduction projects. A homeowner who spends $20,000 on ember-resistant construction and defensible space may still carry the same Fire Factor score as their neighbor who did nothing—because the models update slowly and individual mitigation efforts are not yet systematically captured.
The tools are better than what we had. They are not yet as good as what the decisions they inform require.