State Farm Patents a GAN That Invents the Roof Geometry a Drone Scan Missed

A 3D point cloud built from a drone pass over a roof is almost never complete: obstructed sightlines, reflective shingles and overhanging trees leave gaps a human reviewer would flag as missing data. US Patent 12,705,712, granted to State Farm on August 11, 2026, trains a generative adversarial network to fill those gaps by inventing the missing geometry and labeling roof, window and wall elements in the same pass, then routes the reconstructed surface into both underwriting and claims decisions.

What Claim 1 Actually Describes

The patent's independent method claim is specific about the sequence: obtain a 3D point cloud of a structure, augment it with data describing a structural element's characteristics, then generate "a gap-filled semantically segmented 3D point cloud from the 3D point cloud as augmented with the data using the trained generative adversarial network, wherein the trained generative adversarial network fills one or more gaps in the 3D point cloud while generating the gap-filled semantically segmented 3D point cloud" (USPTO grant, US 12,705,712 B2). Two operations happen in one forward pass through the network: the GAN reconstructs the missing surface and simultaneously assigns a semantic label, roof, window, wall, to every point in that reconstruction. A conventional pipeline would run these as separate steps, first patching the point cloud with an interpolation or inpainting method, then running a segmentation model on the patched result. Claiming both in a single trained network is the part State Farm is trying to own.

Dependent claims spell out what the segmentation is meant to extract. Claim 5 covers identifying "a number of windows of the structure, a size of the portion of the structure, a number of stories of the structure, a roof composition of the structure, or a roof type of the structure," and claim 6 extends the same logic to appliances located inside the structure (USPTO grant, US 12,705,712 B2). Those are not incidental outputs. Roof composition, roof type, window count and story count are precisely the structural characteristics an underwriting model uses to set eligibility and rate a homeowners policy, which means the patent claims the imputation step as an input pipeline for rating variables, not merely a visualization tool.

The Same Model Feeds Underwriting and Claims

Claim 8 is where the patent crosses from feature extraction into loss adjustment: it describes "comparing the gap-filled semantically segmented 3D point cloud against the historical customer data to detect damage to the structure" (USPTO grant, US 12,705,712 B2). In practice, that means a fresh scan of a property, reconstructed and labeled by the GAN, gets diffed against the customer's prior scan on file. A shifted roofline, a missing section of shingles, a collapsed section of fencing, anything that reads as a structural delta between the two reconstructed point clouds becomes a candidate damage flag, potentially before a homeowner has filed a claim or an adjuster has been dispatched.

The abstract frames this as a single system serving three functions at once: a method "for using a trained generative adversarial network to improve underwriting, claim handling and retail operations" that "includes receiving a 3D point cloud; and generating a gap-filled semantically-segmented 3D point cloud using a trained generative adversarial network" (USPTO grant, US 12,705,712 B2). One imputation model, three downstream consumers. That consolidation is the commercial logic of the patent: State Farm does not need to build and validate a separate reconstruction model for pricing, a separate one for claims triage and a third for whatever comes next in retail operations. It builds one gap-filling, labeling GAN and lets every business line that touches property geometry draw from the same output.

State Farm's aerial-imagery program is not hypothetical background here. A Santa Ana, California homeowner was told in early 2026 that her policy would lapse unless she spent roughly $20,000 on roof repairs the carrier identified through an unannounced aerial assessment, a demand State Farm reversed only after she pushed back publicly (ABC7 Los Angeles, 2026). That dispute involved observed imagery from a third-party aerial vendor, not machine-invented geometry. It still shows the trust deficit an insurer is starting from before any synthetic reconstruction enters an eligibility decision, and it is the baseline this patent's imputation step has to be measured against.

Where Imputed Geometry Sits in the Actuarial Chain

Every underwriting and reserving model already relies on assumptions, that is the nature of the work, but assumptions are conventionally applied to observed data: a rating factor takes a known roof age and known territory and produces an expected loss cost. Patent 12,705,712 changes what counts as the observed input itself. When the GAN fills a gap in the point cloud, the roof composition or window count that eventually reaches a rating engine did not come from a measurement of that specific structure. It came from the network's statistical prediction of what a plausible roof or window arrangement looks like, conditioned on the visible portion of the scan and on patterns learned from the training set of other structures. The rating input is synthetic before the rating model ever sees it.

That is a different category of model risk than the one actuaries are used to managing. A pricing actuary can audit a rating factor by tracing it back to an observed variable and a documented relationship. Auditing an imputed structural characteristic requires a second layer: not just "is the rating relationship correct," but "how much of this structure's roof was actually scanned, and how much of what the model reports as roof composition was manufactured by the GAN filling a gap." A patent that segments and labels in the same pass as it imputes makes that distinction invisible downstream unless the carrier deliberately preserves it. Nothing in the claim language requires the output to flag which points in the "gap-filled semantically segmented 3D point cloud" were observed versus generated; the claims describe a single unified reconstruction, presented to underwriting and claims as though it were one consistent measurement.

The governance framework carriers are already operating under does not cleanly anticipate this. As of the second quarter of 2026, 25 states and the District of Columbia had adopted the NAIC's AI Model Bulletin in full or substantially similar form (NAIC AI Bulletin adoption tracker, Q2 2026), and that framework's documentation expectations, training data provenance, validation testing, disparate-impact review, were written primarily with predictive rating and underwriting models in mind: models that transform observed inputs into a score. A GAN that generates the input itself is a step upstream of that framework's usual scope. If a carrier's rate filing or model documentation describes the roof-composition variable without separately validating the imputation sub-model that may have partly invented it, the filing has documented the wrong layer of the pipeline.

A Sixth Branch of the Same 2020 Patent Family

Patent 12,705,712 is not a standalone filing. Its priority chain runs back to Provisional Application No. 62/967,315, filed January 29, 2020, through a continuation filed September 24, 2020 and a further continuation filed November 7, 2022, before splitting into a set of applications that each apply the identical GAN gap-filling method to a different business domain (USPTO Patent Public Search). The pattern is a single core invention, one GAN trained to impute 3D point cloud data, spun into a continuation family that now spans at least six application domains, filed and granted at different points between 2022 and 2026.

Patent	Application domain
US 11,508,042 / 11,995,805	Imputation of 3D data using GANs (base filing)
US 12,039,706	Facilities management and operations
US 12,051,179	Vehicles and transportation
US 12,056,859	Peril modeling
US 12,243,199	Construction and urban planning
US 11,983,851 / 12,705,712	Underwriting, claim handling and retail operations

That breadth is a deliberate patent strategy, not an accident of the continuation process. By filing the same core imputation method against separate application domains, State Farm forces a competitor who wants to build a comparable capability in any one of those areas, peril modeling, facilities management, claims, to design around the same underlying invention repeatedly rather than once. The August 2026 grant extends that strategy into the domain with the most direct pricing and claims exposure: a property carrier's own rating and loss-adjustment pipeline.

The timing lines up with a related grant from the same summer. US Patent 12,694,432, covering a system where a GPT model detects a pricing miss and a separate code-writing model tests and deploys the fix, issued July 28, 2026, two weeks before 12,705,712 (State Farm Newsroom, patent portfolio). One patent claims a method for generating synthetic structural data that can feed a rating engine; the other claims a method for a pricing model to rewrite its own logic with no human sign-off required by its core claim. Granted weeks apart to the same assignee, they point at the same underlying shift: State Farm patenting AI systems that generate or alter their own inputs and outputs, faster than the governance layer built for human-reviewed models was designed to track.

State Farm's overall AI patent position gives that pairing scale. Evident's Insurance AI Patent Tracker counted 326 total granted AI patents for State Farm since 2014, ahead of USAA's 218 and Allstate's 136, with the three carriers together controlling 77% of all insurer AI patent filings (Evident Insights, Insurance AI Patent Tracker). This grant adds to a property-imaging and IoT cluster that already includes a claims-trained water-sensor placement algorithm and a live smart-building sensor scoring method, both granted within the past several weeks, plus a self-updating pricing model patent from earlier in 2026. The point cloud imputation patent slots into that same cluster as the layer that decides what the sensors, cameras and drones are actually looking at when the raw capture is incomplete.

What the Portfolio Signals About the Data Race

Third-party property-data vendors compete by licensing observed imagery, EagleView, Cape Analytics and Zesty.ai each sell carriers access to structure-level attributes pulled from aerial or satellite capture, with Zesty.ai alone reporting coverage of nearly one million previously uninsurable properties across roughly 200 regulatory approvals (Phidea, ZestyAI carrier roster analysis, 2026). That business model has a hard floor: a vendor can only sell what its sensors actually captured. Patent 12,705,712 claims a way past that floor. A GAN trained to impute missing point cloud data does not need a complete scan to produce a complete-looking output; it needs enough of a structure visible to infer the rest with acceptable confidence. If State Farm can generate a usable rating and claims input from a partial scan that would otherwise be unusable, it reduces its dependence on any single imagery vendor's capture completeness, and it does so under patent claims a competitor cannot simply license around by signing a different vendor contract.

That is also why the governance question matters more here than in a typical vendor-imagery dispute. A carrier that buys observed imagery from EagleView or Cape Analytics can, in principle, point to the underlying pixels if a rating decision or a claims denial is challenged. A carrier relying on GAN-imputed geometry has to be able to say, for any specific structural characteristic driving a decision, whether that characteristic was observed or generated, and current model-governance frameworks do not yet require that distinction to be preserved and disclosed. State Farm's own recent AI expansion, including its February 2026 partnership with OpenAI's Frontier program supporting AI tools across 96 million policies, makes the scale of that exposure larger with each new deployment, not smaller (actuary.info coverage of State Farm's OpenAI Frontier partnership).

Documenting Reliance on Data Nobody Observed

The practical response for a pricing or reserving actuary working with a system built on this patent is not to reject synthetic inputs outright, imputation is a well-established statistical practice with a long history in missing-data methods. It is to insist that the provenance flag survive the pipeline: every structural characteristic feeding a rating engine or a damage-detection trigger should carry a marker for whether it was directly observed in the scan or reconstructed by the GAN, and the reconstructed share should be tracked as its own quality metric, not folded silently into the same field as an observed measurement. A rate filing that documents the downstream rating factor without documenting the imputation confidence behind it has audited the wrong layer.

The claims side raises a version of the same problem with sharper stakes. If claim 8's damage-detection comparison flags a structural delta between two reconstructed point clouds, and part of either reconstruction was invented rather than observed, a coverage decision built on that flag needs to be defensible on exactly the same terms the Santa Ana dispute already tested with observed-only imagery, except now with an added question the earlier case never had to answer: was the disputed measurement even real. Carriers moving imputed geometry into underwriting and claims decisions should expect that question to surface in dispute resolution well before it surfaces in a regulatory filing.

There is also a development-pattern consequence worth flagging for reserving, distinct from the damage-detection dispute itself. A carrier running claim 8's comparison systematically, at renewal or on a periodic rescan cycle rather than only after a homeowner files a first notice of loss, opens a claim-reporting channel that does not originate with the insured. Structural deltas the model flags as damage would generate claim records on a schedule tied to the rescan cadence rather than to when a loss actually occurred, which decouples the reported date from the occurrence date in a way a standard loss-development triangle does not expect. A reserving actuary working a property triangle that increasingly mixes insured-initiated claims with model-flagged claims should treat the two as separate reporting patterns rather than assume a single tail factor still describes both, since a periodic-rescan-driven claim can surface months after occurrence in a way an insured-reported claim typically does not.