Alaina Callahan’s roof was five years old. Her insurer told her to replace it or lose coverage. When she asked to see the aerial photos the company had used to make that determination, nobody would show them to her. When she asked how the AI that analyzed those photos reached its conclusion, nobody could explain it. “I had no recourse as a homeowner,” she told NPR. “None whatsoever.”
Callahan lives outside Houston, and her situation is not unusual; it is, increasingly, the standard experience of American homeownership in the age of algorithmic underwriting, driven by an infrastructure far larger than most homeowners realize.
An Airplane Photograph of Every Home in America
The Geospatial Insurance Consortium, an industry-funded organization initiated by the National Insurance Crime Bureau, operates an aerial imagery program that has photographed properties covering 99 percent of the U.S. population. It uses drones, manned airplanes, and high-altitude balloons to capture images of nearly every building in the country, most of the time without the owner’s knowledge or consent.
Computer vision models then scrutinize those images for underwriting hazards: damaged shingles, yard debris, overhanging tree branches, undeclared swimming pools, trampolines. Insurers review the flagged images and decide whether to keep covering the property. “We’ve seen a dramatic increase across the country in reports from consumers who’ve been dropped by their insurers on the basis of an aerial image,” Amy Bach, executive director of the consumer group United Policyholders, told the Wall Street Journal.
If you own a home in the United States, your roof has almost certainly been photographed, scored, and categorized by a model you didn’t know existed. Whether that score is accurate is a different question, and one that nobody is required to answer.
Fifteen States Have Approved AI Scoring for Rate-Setting
Cape Analytics, a geospatial property intelligence company, has secured Department of Insurance approvals in 15 states to use AI-derived roof attributes directly in premium calculations: Arizona, Colorado, Illinois, Indiana, Iowa, Kentucky, Michigan, Mississippi, North Carolina, Ohio, Pennsylvania, South Dakota, Tennessee, Virginia, and Wisconsin. In those states, an AI’s assessment of your roof condition can directly determine what you pay.
Cape’s own research, conducted using Nearmap imagery across millions of aggregate policy-years, found that roofs scored as severe or poor condition had 2.5 times the wind and hail claims frequency of excellent-condition roofs, and homes with significant tree overhang experienced 90 percent higher wind-related losses. Those correlations are real, and roof condition genuinely predicts claims at the portfolio level. But the accuracy of the AI’s condition assessment on any individual home is a separate matter entirely, one that Cape does not address in its published materials, because the company does not disclose property-level error rates.
In May 2026, LexisNexis Risk Solutions added fuel by launching Location Intelligence for Home, a neural-network model that scores individual properties across six perils: hail, wind, weather-related water damage, non-weather-related water damage, freeze, and collapse. Properties with the highest scores are 20 times more likely to generate claims than those with the lowest, according to the company. An optional roof condition grade bolts on top.
LexisNexis plans to file the model as a predictive tool for underwriting and rating in multiple states. When it does, insurers who adopt it will have a property-level risk score generated by a neural network trained on industry-wide claims data, embedded directly in their underwriting workflow. The homeowner will see only the premium.
What the Camera Can’t See
Here is the part that should concern anyone building, buying, or insuring a home: non-weather water damage accounted for 24 percent of all home insurance claims in 2025, according to LexisNexis’s own data, while weather-related water was only 4 percent. A burst pipe, a failed water heater, a slow leak behind drywall: these are the most common claim drivers, and aerial imagery cannot detect any of them. A drone photograph of your roof tells an insurer nothing about whether your 15-year-old supply lines are about to fail.
“Traditional property risk models were built around the signals that were easiest to get, not necessarily the ones most predictive of loss,” said Meredith Barnes-Cook, senior principal at Datos Insights. LexisNexis built Location Intelligence partly to address that gap. But many carriers still weight roof condition heavily in their underwriting, because it’s the attribute they can see from the air, not because it drives the most claims. A homeowner with a new roof and corroded supply lines looks like an excellent risk from above; a homeowner with a 12-year-old architectural reshingle but pristine plumbing looks like trouble.
When the Algorithm Is Wrong
Cindy Picos, a Northern California homeowner from Auburn, had her coverage canceled after aerial photos indicated her roof had “lived its life expectancy.” An independent inspector said the roof was good for another decade. Her insurer declined to reconsider and refused to show her the photos. Tracy Gartenmann in Austin received two grainy aerial images from Travelers Insurance and a two-month ultimatum to trim trees or lose coverage. “I thought it was a scam,” she said.
In Texas, NPR’s Audrey McGlinchy uncovered at least a dozen state investigations since 2023 into insurers using aerial imagery and AI to determine renewals, and in one case an insurer canceled coverage after examining the wrong image entirely: not the wrong roof, not a misread shingle, but a photograph of somebody else’s house.
The Regulatory Gap
The NAIC Model Bulletin on AI in insurance establishes two principles: don’t run AI models on bad data, and insurers cannot outsource accountability for vendor scoring. A computer vision model that systematically downgrades older roofs in certain neighborhoods can create what Troutman Pepper Locke’s legal analysis calls “systematic disparities.” Connecticut has advised insurers that cosmetic issues alone should not trigger underwriting action. Massachusetts affirmed homeowners’ right to see the photos and contest decisions.
But those are exceptions. Most states have no specific rules governing aerial imagery in underwriting. California saw State Farm drop 30,000 residential policies, partly driven by algorithmic risk reassessment in wildfire-prone areas, with no requirement to disclose the scoring methodology to affected homeowners. Three things would close the gap: mandatory disclosure of the images and scores used in underwriting decisions, a defined contestation process with ground-truth inspection rights, and published accuracy metrics for any AI model filed for rate-setting. Actuarial models in insurance are routinely validated. AI underwriting models should be held to the same standard.
The Missing Number
Cape Analytics proved that poor-condition roofs generate 2.5 times more claims. Nearmap’s Roof Age Gen2 claims 95 percent accuracy in detecting roof replacement year. Neither company publishes the rate at which their condition scoring disagrees with a competent human inspector on the ground, and no state insurance department tracks how many non-renewals were reversed after homeowners challenged an aerial assessment.
If the AI rates your roof as “poor” instead of “fair,” your premium might increase by $500 to $1,500 per year. If you’re non-renewed entirely, replacement coverage through a surplus-lines carrier or your state’s FAIR plan can cost double or triple the standard market rate, and a single wrong score that persists for three years could mean $1,500 to $4,500 in excess premiums before you discover and challenge it. Nobody has measured this at scale because nobody is required to. The error rate is the denominator that makes the entire system accountable, and it does not exist in any public record.
What This Means If You’re Building
If you’re building a new home, you are building for two audiences: yourself and the algorithm that will photograph it within months of completion. Architectural-grade shingles photograph differently from three-tab shingles that still meet code; light-colored roofing shows discoloration faster in aerial photos regardless of actual condition; metal roofs, which last 40 to 70 years, sometimes trip AI models trained predominantly on asphalt shingle data. Tree placement matters, too: a mature oak that shades your roof and cuts cooling costs by 25 percent also triggers the tree-overhang hazard flag correlated with 90 percent higher wind losses, and the AI doesn’t know whether your arborist visited last month.
Document everything. If you are buying an existing home, get a roof inspection before closing, not for structural peace of mind alone, but so you have timestamped professional documentation of your roof’s actual condition that predates whatever the insurer’s AI decides when your policy comes up for renewal.
Limitations of This Analysis
No insurer publishes its AI roof scoring accuracy or appeal outcomes. Homeowner accounts cited here are self-reported through media interviews, and we cannot independently verify the condition of their roofs. Cape Analytics’ 2.5x claims frequency correlation for poor roofs is drawn from carrier-contributed data; the methodology for defining “poor” is proprietary and not independently audited. LexisNexis Location Intelligence launched in mid-2026 and has no independent validation. State regulatory responses remain fragmented and, in most jurisdictions, nonexistent.
The cost estimates for misclassification are illustrative, derived from publicly available premium ranges by roof condition tier. Without published error rates, the aggregate financial impact is incalculable. That incalculability is, itself, the problem.