Shallowfake insurance claims: what they are and how to stop them
The short version.
- A shallowfake is a real photo, video, or document edited with ordinary tools. No AI involved: Photoshop, a phone editing app, sometimes nothing more than a screenshot with a detail changed.
- It is already the volume problem. Allianz tracked a 300% rise in app-manipulated claim media between 2021-22 and 2022-23, and Verisk found 36% of consumers would consider editing a claim image.
- The strongest defense is intake design, not just detection. Live guided capture with no gallery uploads removes the window where the edit happens; sealing evidence on receipt makes later tampering visible.
- Written for counter-fraud and SIU teams at motor and home insurers who review claimant-submitted media every day.
A car sits on a salvage agent's website, photographed after being written off. Someone downloads those photos, edits a different registration plate onto the wreck, and files a claim for a vehicle that was never in an accident. The handler who opens the file sees a real car, real damage, and a plate that matches the policy.
That is a shallowfake, and it is quietly becoming the most common manipulated media in motor and home insurance claims. It needed no AI at all, and if you run counter-fraud or an SIU, that low barrier is the whole problem.
This post covers what shallowfakes are, the reported cases worth studying, why they outnumber deepfakes by a wide margin, and how to remove the editing window they all depend on.
We wrote this for the counter-fraud lead or SIU manager whose referral queue keeps filling with claimant photos nobody can fully trust, not for a general audience. Our team builds Venta Capture, a product of VentaVid, which exists because of exactly this problem. If edited media is on your risk register this quarter, start a free account and look at what sealed intake does to it.
What a shallowfake insurance claim is (and what it isn't)
A shallowfake is a genuine photo, video, or document that has been altered with everyday editing software. The fraudster doesn't generate anything from scratch. They start with something real (their own car, a photo from a salvage listing, a template invoice) and change the detail that matters: a dent painted in, a date moved, a registration plate swapped.
That puts shallowfakes in a different category from deepfakes:
- Tools. A shallowfake needs Photoshop or a free phone app. A deepfake needs generative AI.
- Skill. A teenager can produce a passable shallowfake in minutes. Convincing synthetic media still takes more effort.
- The starting material. A shallowfake abuses a real capture, which is precisely why it looks credible. Nothing about the lighting, the scene, or the vehicle is fake. Only the claim built on top of it is.
The barrier to entry is the whole story. In Verisk's 2026 State of Insurance Fraud study, 36% of consumers said they would consider digitally altering a claim image or document even knowing it breaks their insurer's rules, and among Gen Z that figure hits 55%. This is not a specialist crime anymore; it is a temptation sitting in every claimant's photo app.
Fully AI-generated claim media is a separate and growing problem with its own detection arms race, and it deserves its own article. This one stays on the low-tech version, because the low-tech version is where today's volume sits.
The fake plate on the written-off car
The clearest reported shallowfake scheme comes from Zurich UK, first reported in May 2024, and it is worth walking through slowly because every step defeats a control that most claims intake still relies on.
- Find a total loss. The fraudster browses a salvage agent's website for a written-off vehicle with dramatic, genuine damage.
- Transplant the plate. Using ordinary editing software, they place a registration number onto the wreck that matches a policy they control.
- File the claim. The photos arrive at first notification of loss looking like what they are: real pictures of a destroyed car.
- Collect on a total loss. Because the vehicle presents as beyond repair, there is no repairer visit, no engineer inspection, and often no physical touchpoint at all.
"People can now create a fraudulent claim entirely from behind their computer and extract significant sums of money because these cars are total losses." by Scott Clayton, head of claims fraud, Zurich UK, as reported in The Guardian, May 2024
Clayton's team called this an emerging threat for a reason: the handler assessing the claim has no obvious tell. The damage is real. Only the plate is not.
The raw material rarely needs stealing, either. Allianz UK found that 52% of drivers aged 18 to 34 have posted pictures of their car online, and 73% had no idea that exposes them to fraud. In one case Allianz highlighted, a tradesman's van photo was lifted from his own business page, edited to show a cracked front bumper, and submitted with a false repair invoice of around £1,000, for an accident that never took place.
The trend line matches the anecdotes. Allianz and Zurich reported a 300% rise in claims involving app-manipulated images, videos, and documents between 2021-22 and 2022-23. A 300% rise in a single year means the playbook has gone mainstream.
Why shallowfake insurance claims beat deepfakes on volume
Deepfakes make the headlines. Shallowfakes make the losses. Three structural reasons explain why, and none of them are going away.
The pool of capable offenders is the entire policyholder base. Nobody needs to learn a new tool. Opportunistic exaggeration (a real scratch stretched into a real-looking dent) sits within reach of anyone with a grievance and a phone. For scale: Allianz alone combatted more than 34,200 fraud cases worth nearly £174 million in a single year.
Small edits hide below referral thresholds. A wholly invented claim trips wires. A genuine £900 repair inflated to £1,400 with one edited photo and one adjusted invoice usually does not. Multiply a modest inflation across thousands of low-severity claims and the indemnity leakage dwarfs the occasional spectacular fraud.
Detection is losing the race on volume, not sophistication. In the same Verisk survey, 66% of insurers admitted digital media fraud slips through undetected often or very often. The backdrop is grim: Aviva alone reported detecting £233 million in suspected fraudulent claims in 2025, flagging more than 18,400 suspicious claims in the year. Those are the ones that got caught.
So the counter-fraud question shifts. When any claimant can produce credible fake evidence in five minutes, examining files harder after they arrive stops being enough. You have to ask where the fake gets made.
The editing window: where every shallowfake is made
Look at the path a standard "send us some photos" intake accepts: record, save to the gallery, edit, export, save again, upload. Every shallowfake in every case above was manufactured in the middle of that chain.
The photo of the salvage wreck sat on a disk before the plate went on. The van photo lived on social media before the bumper crack was painted in. The intake process saw only the final file, stripped of its history, and treated it as testimony.
That gap between recording and submission is the editing window. Upload-based intake was designed around it: the whole flow assumes the claimant prepares files first and submits them later. A gallery upload is, by definition, a file whose past you cannot see.
When we were mapping this problem while building our own capture product, the thing that struck me was not the cleverness of any single edit. It was how much time a normal claims flow politely hands the claimant between taking a picture and submitting it. Hours, days, sometimes weeks: all of it available for revision.
And the revisions convince. Among consumers who had already used editing tools on a photo, video, or document, 44% rated their own results as very realistic, per the same Verisk study. Close the window instead, and you stop hunting for edits after the fact because the time and place where edits happen no longer exist in your evidence path.
Provenance-first collection: how to close the window
This is the approach Venta Capture takes for remote claim inspection: treat the collection of evidence, not just the analysis of it, as the fraud control.
Capture replaces upload. The claimant gets a link by SMS, email, or QR code. It opens a guided capture flow in their phone's browser, with no app to install and no account to create, in any of 15 languages. Video and photos are recorded live inside that flow. There is no gallery upload, so a file that spent last week in Photoshop has no way into the case.
Your experts decide what gets shown. Instead of whatever the claimant chooses to send, the flow walks them through it: the whole vehicle, then the damaged area from a distance, then close up, then the surrounding context, with their spoken explanation transcribed into the case. Recycled or staged material has to survive a live, directed walkthrough rather than a one-angle snapshot.
Everything is sealed on receipt. The moment a submission arrives, the platform records a server-side receipt time, fingerprints every file with a SHA-256 hash, and wraps the submission in a digitally signed seal that can be verified outside the platform with a public key. The claims page documents 28 control points per submission: 10 automatic signals, 21 recorded session events, and 1 signed seal. If anyone alters a file afterwards, the fingerprint no longer matches. Months later, in a dispute, you can still prove the file you are looking at is the file you received.
Signals stay signals. Indicators like a virtual camera, a jailbroken device, or a location mismatch surface as graded flags (ignore, info, amber, red) that your team weighs against its own risk policy. The platform never returns a fraud verdict, because a signal is a reason to look, not a reason to reject.
For an SIU manager, the practical effect is a different starting position. Instead of asking "can I trust this JPEG?", your handlers review a structured evidence case: live-captured media, answers, transcript, receipt time, integrity fingerprints, and a session timeline, all in one place.
If you want to feel the difference, run one real damage scenario through it. Start a free account and send yourself a capture link; the walkthrough takes about ten minutes.
What provenance doesn't fix
Anyone who tells you a tool makes claims fraud impossible is selling something. The pressure is still rising (Admiral reported a 71% increase in detected fraud in 2025 versus 2024), and live capture closes one door, not all of them. Be clear-eyed about the limits.
- A genuine recording can still show a staged scene. Pre-existing damage presented as new, borrowed property, a misleading angle. Live capture proves the recording is real, not that the story around it is true.
- The timestamp proves receipt, not the event. A server-side receipt time tells you when the system received the submission. It does not tell you when the damage occurred, and treating it otherwise will burn you in a dispute.
- Documents need their own scrutiny. Invoices, receipts, and reports arriving through other channels remain classic shallowfake territory, and your document forensics still earn their keep there.
- People still make the call. Signals inform a human decision; they never replace it. Keep your referral criteria and investigator judgment at the center.
What a layered setup looks like in practice, from teams rethinking their media intake:
- Route FNOL media through live guided capture wherever the claim type allows, so the default evidence path has no editing window.
- Treat gallery uploads as lower-trust by policy, not as equivalent evidence that happens to look fine.
- Hash-verify on dispute. When a case reopens months later, check the file's fingerprint against the sealed original before anyone argues about its contents.
- Write signal thresholds into your referral rules, so an amber device flag or location mismatch triggers a consistent second look instead of an ad-hoc one.
- Keep detection tools on documents and legacy channels. Provenance-first collection and post-submission forensics cover different ground; you want both.
The same guided flows carry inspection work well beyond claims, from field service and site checks to property and equipment captures, so the intake fix is not a single-use tool. If you'd rather see it before touching it, book a 15-minute demo and bring your ugliest closed file.
Frequently asked questions
What is a shallowfake insurance claim?
A shallowfake claim uses a real photo, video, or document that has been edited with ordinary software to support a false or inflated claim. Typical examples are added damage on a genuine vehicle photo, a swapped registration plate, or an altered invoice. No AI is involved, which is what separates shallowfakes from deepfakes.
What is the difference between a shallowfake and a deepfake?
A deepfake is synthetic media generated by AI; a shallowfake is authentic media manually edited with tools like Photoshop or a phone app. Shallowfakes are far cheaper and faster to produce, which is why insurers currently see many more of them.
How do insurers detect shallowfake photos?
Post-submission tools analyze metadata, compression artifacts, and cloned pixel regions, and reverse-image search catches recycled photos. The complementary approach is provenance-first collection: requiring media to be recorded live inside a guided flow and sealed with hashes and signals on receipt, so there is no editing window in the first place.
Does live guided capture make shallowfake fraud impossible?
No, and nobody should claim it does. A live recording can still show a staged or misleading scene, so investigator judgment stays central. What live capture removes is the ability to submit media that was edited between recording and submission, which is where the shallowfake itself is made.
What does the timestamp on a sealed submission actually prove?
A server-side receipt time proves when the insurer's system received the submission, independent of the claimant's phone clock. It does not prove when the damage occurred. That distinction is exactly what makes the timestamp reliable in a dispute: it states something narrow and defensible.
Do claimants need to install an app to submit guided capture evidence?
No. A capture link opens directly in the phone's mobile browser, with no app download and no account, which matters because every extra installation step costs completion rates at FNOL. The claimant records, answers the flow's questions, and submits in one session.
Try Venta Capture on your own process
Build one flow for your highest-volume case type. Free, no credit card.
Close the editing window
Shallowfakes work because standard claims intake accepts files with an invisible history. You can keep scanning those files harder, and you should. But the bigger win sits upstream: collect evidence through a channel where the edit has no room to happen, and seal it the moment it arrives.
Start a free account (no credit card required) and run a test claim through a guided flow this week, or have the team behind the platform walk your counter-fraud workflow through it live.
The VentaVid team builds Venta Capture, guided visual capture that turns claimant smartphones into a sealed evidence channel.