Can You Trust AI to Assess Insurance Claims?

The Problem

A major insurer asked its AI to "enhance" a photo of a severely dented rear bumper. The AI erased the damage entirely and returned a pristine, fictitious image of an undamaged car. This is the now-infamous "Pristine Bumper" incident. The insurer had embedded a generative AI tool — built on the same type of technology behind Stable Diffusion and DALL-E — into its mobile claims app. When a policyholder uploaded a photo after a collision, the AI interpreted the crunched metal as visual noise. It digitally smoothed the dent away, replacing it with the clean lines of a perfect fender.

The automated claims engine then denied the claim. Its reason: zero visible damage. The policyholder, staring at a wrecked vehicle in the driveway, sued for bad faith. The insurer was left holding a digitally altered record that directly contradicted physical reality.

This was not a bug. The AI did exactly what it was designed to do. These models learn from billions of images that a "car" is a smooth, symmetrical object. A dent looks like an error to fix, not evidence to preserve. If your claims workflow touches generative AI for image processing, this same failure mode exists in your system today. The question is whether it has already fired — and whether you would know if it did.

Why This Matters to Your Business

The financial and legal exposure here is not hypothetical. It is structural. Every time a generative AI tool alters a claim photo, your organization risks what the legal system calls spoliation of evidence — the alteration or destruction of records relevant to a legal proceeding. Courts can impose sanctions, issue adverse inference instructions to juries, or grant summary judgment against the spoliating party.

Here is what that means for your balance sheet and your boardroom:

• Litigation multiplier. The whitepaper identifies a $7.2 billion litigation risk tied to AI-driven evidence failures in insurance. A single bad-faith lawsuit from one altered photo can cost millions. Scale that across thousands of daily claims, and you see the exposure.
• Regulatory liability you cannot delegate. The NAIC Model Bulletin on AI use by insurers now mandates that carriers maintain a written AI governance program. Crucially, you cannot outsource accountability to your AI vendor. If a third-party tool hallucinates or discriminates, the insurer is liable. Multiple states have already adopted this guidance.
• EU AI Act classification. For insurers with global operations, the EU AI Act classifies AI used for insurance risk assessment involving individuals as High Risk. That triggers strict requirements for data governance, automatic event logging, and human oversight.
• Fraud vulnerability. The same generative AI that deletes real damage also enables fraudsters to manufacture fake damage. Criminals now use text-to-image tools to add smashed bumpers or fire damage to photos of pristine vehicles. Human reviewers perform no better than random chance at detecting high-quality deepfakes. Your current defenses are likely inadequate.

The "Pristine Bumper" incident cost one insurer its credibility and a lawsuit. Your exposure is the same every time a generative tool touches your evidence chain.

What's Actually Happening Under the Hood

To understand why this keeps happening, you need to understand one key difference. There are two fundamentally different kinds of AI: generative and discriminative. Generative AI creates new content. Discriminative AI analyzes what already exists. The insurance industry needs analysis, but it keeps buying creation tools.

Think of it this way. A generative AI is like a portrait artist who paints what a car should look like. A discriminative AI is like a forensic examiner who measures what the car actually looks like. You would never ask a portrait artist to serve as an expert witness. But that is essentially what happens when you embed generative image processing into your claims pipeline.

The technical failure mode is called inpainting. Generative models work by learning statistical patterns. In their training data, cars overwhelmingly appear with smooth, unbroken surfaces. When the model sees a dent — a high-frequency disruption in the expected pattern — it treats it as noise. Its mathematical objective is to produce an output that looks like a "normal car image." So it fills the dented pixels with smooth ones. It heals the damage.

This process runs on diffusion mathematics, where the model reverses the introduction of noise to recover a "clean" signal. A dent looks like noise to these models. They remove it. In art restoration, that is a feature. In insurance claims, it is automated evidence destruction. The model replaces the messy, expensive reality of a crash with a plausible, inexpensive fiction.

The problem compounds when you rely on third-party API wrappers. Most InsurTech tools do not build their own AI. They send your data to general-purpose models that have no understanding of indemnity, collision physics, or forensic standards. If the model provider updates its weights to be more "aesthetic," your damage assessment tool might start erasing dents more aggressively — and you have zero control over that change.

What Works (And What Doesn't)

Let's start with what fails in practice.

Generic generative AI wrappers: These tools treat damage as visual noise and remove it. They alter pixels, creating spoliation risk. They are vulnerable to deepfake fraud because they evaluate content, not the physics of how an image was made. A 99% failure rate is referenced in the context of these tools' unsuitability for forensic work.

Simple image classifiers: Standard classification models (like ResNet) give you a binary answer — "damaged" or "not damaged." They cannot tell you where the damage is, how deep a dent goes, or whether the photo is a deepfake. That level of vagueness does not survive a courtroom challenge.

Human-only review at scale: Humans are poor at detecting AI-generated fakes. They also introduce subjective variability. Two adjusters looking at the same dent may give different severity scores. You cannot audit subjectivity.

Here is what works — a deterministic forensic approach that analyzes without altering:

1. Input — Evidence locking. The original photo is hashed with SHA-256 encryption the moment it arrives. The AI reads the image but never writes to it. Metadata — timestamp, GPS coordinates — is locked to the file. This creates a chain of custody that holds up in court.

2. Processing — Three-layer analysis. First, semantic segmentation — a technique where the AI classifies every single pixel — identifies and maps damage boundaries. It can tell you a scratch is 14 centimeters long and a dent covers 45 square centimeters of surface area. Second, monocular depth estimation — calculating 3D geometry from a 2D photo — measures dent depth and volume. A steep depth gradient indicates a sharp crease requiring part replacement. A shallow gradient means paintless dent repair may suffice. Third, specular reflection analysis — studying how light bounces off the car's surface — detects invisible damage like hail dents or previous repairs that standard AI misses entirely.

3. Output — Structured, auditable report. The system produces a JSON report listing damaged parts, severity scores, and repair-versus-replace recommendations based on your configured business rules. All analysis is saved as separate sidecar files linked to the original image hash. Nothing in the original evidence is changed.

This is what your compliance team needs to hear. The analysis overlay can be toggled on and off. The adjuster remains the human in the loop with final decision authority — a critical safe harbor under the EU AI Act. Every processing step is logged. You can explain exactly why a claim was approved or denied, down to the pixel measurements. That is the difference between a system that survives regulatory scrutiny and one that creates your next lawsuit.

When your current AI vendor says they use "AI for claims," ask what kind. If the answer involves generation, enhancement, or any process that writes new pixels into your evidence, you are carrying risk you may not have priced.