AI-generated insurance fraud: why it beats first review
This post in 30 seconds.
- The shift: editing or generating a convincing claim photo now takes minutes on a phone and costs nothing. In Verisk's 2026 study, 98% of insurers agreed AI editing tools are fueling digital media fraud.
- The numbers: Admiral's detected fraud rose 71% in 2025. Aviva stopped £233 million. Belgian insurers proved €181 million and estimate the real total near €800 million a year.
- The gap: detection tools inspect a file after it arrives, when its history is already gone. Provenance-first collection controls how the evidence gets created in the first place.
- Written for heads of counter-fraud and claims directors who have to catch the manipulated few without slowing the honest many.
The claim file looks clean. Twelve photos, sensible angles, a readable plate, an invoice for the repair. Five years ago that file was your strongest evidence. Now it might be the least verifiable thing on your desk, because AI-generated insurance fraud has made the photograph the easiest part of a claim to fake.
The fraudsters did not get smarter. The tools got free. Every insurer that has published fraud numbers for 2025 is describing the same curve, and this post walks through what they disclosed, which kinds of fake are driving it, and why the structural answer sits at the moment of capture rather than in the review queue.
A note before the data. We build Venta Capture, a product of VentaVid: guided evidence capture where the claimant films damage on their own phone through a link, no app and no account, and the submission arrives as a digitally sealed case. If you head counter-fraud or run the claims book at an insurer, and manipulated images are getting past first review while your cycle-time targets stay exactly where they were, start a free account and test it on a single claim type.
Faking a claim used to cost effort. Now it costs a prompt
An exaggerated motor claim used to need physical commitment: a staged collision, real damage, sometimes an organised crash-for-cash ring. Convincing photo edits needed desktop software and someone who knew how to use it.
That barrier is gone. Phone-native editors remove objects, add damage, and relight scenes in seconds, and image generators will produce a photorealistic cracked bumper or a water-stained ceiling from one sentence of text.
The people watching this land in real queues are blunt about it. In Verisk's March 2026 State of Insurance Fraud study, which surveyed 300 claims professionals and 1,000 consumers, 98% of insurers agreed that AI editing tools are fueling a rise in digital media fraud, and 76% said claim submissions have grown more sophisticated in the past year.
The consumer half of the same study shows where the volume comes from. 55% of Gen Z respondents and 49% of millennials said they would be at least somewhat likely to make a small, rule-bending edit to a claim photo or document. Among baby boomers, 12%.
Opportunistic manipulation at population scale, not specialist crime, is the growth segment. Allianz UK saw claims built on app-manipulated images, videos, and documents rise 300% from 2022 into 2023, before image generation even went mainstream. One case from its files: a claim was pursued in a van owner's name for an accident that never took place, using a photo lifted from his own business's social media page with a cracked bumper edited on, plus a false repair invoice for just over £1,000.
The numbers: AI-generated insurance fraud in 2025 and 2026
A sourcing note first, because fraud statistics attract inflation of their own. Every figure in this section comes from an insurer's own disclosure, an industry federation, or a named study, with the source in the row. Where a number is an estimate, it says so.
Read the Belgian row twice. Assuralia can prove €181 million, yet estimates the true total at up to €800 million. Detected fraud is a floor, not a ceiling, and the gap between the two is the honest measure of the problem.
Aviva says it plainly in its own release: it is seeing "a growing number of claims supported by AI-generated images and manipulated documents", particularly in motor. Reinsurers agree on the direction. Swiss Re's SONAR 2025 report lists deepfakes and disinformation enabling insurance fraud among its high-impact emerging risks.
The industry's answer so far is mostly detection spend. In Deloitte's 2025 survey, 35% of insurance executives put fraud detection in their top five priorities for generative AI investment, and Deloitte projects AI across the claims lifecycle could save P&C insurers between $80 and $160 billion by 2032. Spending is not the same as confidence, though: in a 2026 ACFE and SAS survey, only 7% of anti-fraud professionals called their organisation more than moderately prepared for AI-driven fraud.
Shallowfakes, deepfakes, and generated images are three different problems
Counter-fraud teams tend to inherit "deepfake" as a catch-all label. It isn't one, and the distinctions decide which of your controls fail.
Each type defeats a different control.
- Shallowfakes beat the visual sniff test. Most of the pixels are real, so nothing looks off. Zurich UK's head of claims fraud, Scott Clayton, describes fraudsters locating total-loss vehicles on salvage agents' websites and editing a target registration onto them, so a handler assesses a wreck the claimant never owned. We break this category down further in shallowfake insurance claims.
- Generated images beat reverse-image search. The picture has no history because it never existed before the prompt. There is no original to trace, no source photo to match against a database.
- Deepfakes attack identity rather than evidence. Cloned voices and synthetic video target the conversations around a claim: the "policyholder" on the phone, the "witness" on a video call. Different threat, different control surface.
The common thread is that all three arrive through the same door: an upload field that accepts whatever file it is given. That door is the real weakness, and it is the subject of our companion piece on photo evidence in insurance claims.
Why detection-only counter-fraud keeps losing ground
Detection tooling matters, and it is improving. It is also structurally behind, for three reasons no vendor roadmap removes.
- The arms race resets itself. Detectors learn the artifacts of today's generators: lighting inconsistencies, repeated textures, impossible shadows. Each new model release erases yesterday's tells, and the detector starts over while the fakes already in your queue were made with the newest tools.
- Base rates punish false positives. Roughly 1 in 10 P&C losses involves fraud, per the Coalition Against Insurance Fraud. The other nine claimants are honest people having a bad week. A detector that wrongly holds even 5% of them delays thousands of genuine customers per manipulated claim caught, and your cycle-time and NPS numbers absorb the damage.
- Forensics starts after the facts are gone. A gallery upload arrives with its history amputated: recorded, saved, edited, exported, re-saved, uploaded. You can examine the pixels. You cannot examine the process that produced them.
Detection asks whether a file looks fake. Provenance asks what you know about how the file came to exist. The second question turns out to be far more answerable, which is why the wider industry is moving toward provenance standards like C2PA content credentials for cameras and editing software.
For a claims organisation, the most useful control point sits at the moment of capture, well before the review queue. Provenance cannot be retrofitted onto a file someone hands you, but it can be built in when the evidence gets created.
Provenance-first collection: decide how evidence gets created
Flip the collection model. "Send us some photos" invites files with unknowable histories into your claim system. Guided live capture replaces the upload with a process you control.
It works like this. The claimant receives a link by SMS, email, or WhatsApp and opens it in their phone's browser, with no app to install and no account to create. The flow walks them through exactly what your assessors need: the full vehicle before the damage close-up, the room before the water stain, a spoken explanation while recording. Everything is filmed in the moment, inside the session. Gallery uploads never enter the case.
This is the model Venta Capture runs for claims teams, and the collection step is only half of it. The other half is what arrives: not loose files but a sealed case.
- Server receipt time. The system records when it received the submission, independent of the phone's clock. This proves the moment of receipt. It does not prove when the damage happened, and a tool that implies otherwise is overclaiming.
- A SHA-256 fingerprint per file. If anyone asks in month eight whether the video in the case is still the exact file you received in week one, the fingerprint answers it.
- A session timeline. Every step from link opened to submission, logged on server time, so a reviewer sees the process around the material, not just the material.
- Automatic signals. A virtual camera in play, a jailbroken device, a device or IP already seen in earlier claims, GPS location verified to within roughly 11 metres.
- A signed seal. The whole submission carries a signature verifiable with a public key, so integrity checks do not depend on trusting anyone's dashboard.
The live product page quantifies it as 28 control points per submission: 10 automatic signals, 21 recorded session events, one signed seal.
Now the caveat, stated plainly because your SIU will raise it anyway. Harder to manipulate does not mean impossible to deceive. A genuine live recording can still show a staged scene, a neighbour's car, or damage that predates the policy. Live capture narrows the manipulation problem sharply; it does not settle the truth of a claim, and nobody should tell you it does.
When we designed the evidence layer, one rule sat above all the others: signals, not verdicts. A flagged submission is a reason to look. On its own it is never a reason to reject, and the decision stays with your handlers and your investigators rather than with a black-box score.
Want to see what a sealed case looks like against one of your own claim types? Start a free account and build a capture flow in an afternoon, or book a 15-minute demo and we'll walk one through together.
What provenance-first looks like at FNOL
On paper this is a big shift. In the workflow, it is one added step at first notification of loss.
- Claim reported. The handler sends a capture link instead of asking for photos.
- The claimant films, guided. Overview, damage from distance, damage close up, spoken explanation. They do it when it suits them; nobody schedules a video appointment.
- The case arrives sealed. Video, photos, answers, transcript, receipt time, fingerprints, and signals land together as one reviewable case.
- Signals sort the queue. Clean submissions move straight to desktop assessment. An amber signal earns a closer look. A red pattern goes to the SIU with context attached instead of a hunch.
- Retakes replace restarts. Missing the rear quarter panel? Request that one shot. The addition joins the same case with its own timestamps.
- The decision stays human. The handler assesses, remote-first rather than remote-only, and escalates to a physical inspection when the case warrants it.
Remember the base rate while reading that list. Aviva flagged more than 18,400 suspicious claims in 2025, and behind that number sits a far larger set of genuine claims that had to keep moving. Structured capture serves them first: fewer photo chase-ups, faster desktop assessment, and a documented file if a dispute surfaces later. That is how claims cycle time comes down while scrutiny goes up.
And because the capture is asynchronous, it scales without headcount. Five hundred claimants can film independently on the same afternoon while your team reviews by priority on the same platform. No staffed video sessions, no calendar alignment.
Frequently asked questions
How big is AI-generated insurance fraud right now?
Nobody has a global total, but insurer disclosures show the direction. Admiral's detected fraud rose 71% in 2025, Aviva stopped £233 million in suspect claims, and Assuralia proved €181 million in Belgium while estimating the real figure at up to €800 million a year.
What is the difference between a shallowfake, a deepfake, and a generated image?
A shallowfake is a real photo or document altered with ordinary editing tools. A generated image is fully synthetic, created from a text prompt with no original behind it. A deepfake is synthetic video or voice of a real person, typically used to fake identity rather than damage.
Can detection software spot AI-generated claim photos?
Sometimes, and the tools are improving, but detection is an arms race: every new generation model erases the artifacts detectors were trained on. That is why only 7% of anti-fraud professionals told the ACFE and SAS they feel more than moderately prepared. Detection works best as one layer, not the whole defence.
What is provenance-first evidence collection?
Instead of accepting uploaded files and testing them for manipulation afterwards, the insurer controls how evidence gets created: guided live capture on the claimant's phone, recorded in the moment, then sealed on receipt with timestamps, cryptographic fingerprints, and session data. The question shifts from "does this look real?" to "what do we know about how this was made?"
Does guided live capture make claims fraud impossible?
No, and you should distrust any vendor who claims it does. A genuine recording can still show a staged scene or misleading context. What sealed capture does is make manipulation much harder, make integrity verifiable afterwards, and give reviewers signals that say where to look.
Will extra verification slow down honest claimants?
Done as guided capture, the opposite. The claimant films once, in minutes, in their browser, instead of trading photo requests over email for a week, and clean submissions move to desktop assessment faster because the documentation arrives complete.
Are insurers using AI to fight AI fraud?
Heavily. Deloitte found 35% of insurance executives rank fraud detection among their top generative AI priorities, and projects up to $160 billion in P&C savings by 2032. The strongest programmes pair that detection layer with provenance-first collection, so fewer manipulated files enter the pipeline at all.
See it on a real claim
Reading about sealed evidence is one thing. Watching a guided capture arrive from your own phone, with the signals and seal attached, is what usually convinces a counter-fraud team.
Low-risk to try.
- Free account, no credit card. Build a capture flow for one claim type and send yourself the link.
- Nothing for the claimant to install. The capture runs in the phone browser, in 15 languages.
- A demo on your scenario. Bring a real claim type and we'll build the flow live on the call.
Venta Capture is built by the VentaVid team.
Start a free account or book a demo if you'd rather see it before you build.