Shallowfake
Shallowfake, defined: what a shallowfake is and how it differs from a deepfake
A shallowfake is a real photograph, video or document that someone has altered using ordinary editing tools: a phone photo app, Photoshop, or a free browser-based editor. No AI is involved. The file starts out genuine, and something in it is added, removed, or swapped.
The name suggests a lesser deepfake. Operationally it is the opposite. Shallowfakes are the ones claims teams meet every week, because making one takes a phone and about five minutes.
Shallowfake vs deepfake: is a shallowfake a deepfake?
No. They are different things with different tells, and treating them as one problem is how claims operations end up defended against the rarer of the two.
- Shallowfake: a real file, edited with conventional software. The pixels of the car are that car's actual pixels. The crack in the bumper is painted on.
- Deepfake or generated image: a synthetic file, produced by AI. There may be no real car, no real scene, and no original photograph at all. See the deepfake and AI-generated insurance fraud side of it.
- Skill required: a deepfake needs a model and some intent to learn a tool. A shallowfake needs a finger and a smartphone.
- Detection: generation artefacts and impossible physics point at synthetic media. Edit residue, metadata gaps and image reuse point at shallowfakes. Tools tuned for one will pass the other straight through.
That accessibility is why shallowfakes carry the volume today. Anyone who has removed a stranger from a holiday photo already has the technique.
What does shallowfake mean in a claims file?
- Added damage: a dent, crack or scrape edited onto an undamaged vehicle, then submitted with a repair estimate.
- Swapped registration plates: a plate number implanted onto a photo of a different vehicle, so a claim can be made against a car the claimant does not own.
- Recycled images: a photo lifted from social media or a salvage listing, lightly edited so it no longer matches the original on a reverse image search.
- Edited paperwork: an invoice total, a date, or a garage name changed on an otherwise genuine document.
- Cropped context: nothing altered at all, but everything that would explain the scene removed from the frame.
Shallowfake example: the salvage plate swap
Zurich UK has described the pattern in public. Fraudsters find written-off vehicles listed on salvage agents' websites, implant a different registration number onto those photographs, and claim for the damage as though it were their own car.
Scott Clayton, Zurich's head of claims fraud, called it an emerging threat and made the point that a handler assessing the file takes the image at face value, because the damage in it is real. It just belongs to somebody else's car. The fraud used to require two vehicles and a collision. Now it requires a laptop.
Why shallowfakes are the volume problem
Allianz UK and Zurich UK both reported that cases where apps were used to distort real images, videos and documents rose by around 300% between 2021-22 and 2022-23. Both flagged it early as a pattern that would keep growing, and the reason is the economics rather than the technology: the skill floor is on the ground and the payout is unchanged.
How claims teams spot them, and where they go wrong
- Ask about provenance before appearance: where the file came from is more answerable than whether it looks edited. Gallery uploads carry a record, edit, export, upload path that live capture does not.
- Check metadata and file structure: EXIF gaps, re-save artefacts and mismatched device signatures earn a closer look. None of them is proof.
- Reverse image search: recycled photos are the most common variant, and the original often sits on a public salvage listing or social profile.
- Cross-reference the plate: check the registration against write-off and salvage records before valuing the damage.
- Collect through guided capture: when an image is recorded inside the claims process rather than chosen from a camera roll, the editing window closes.
Two mistakes cost the most. The first is buying detection built for synthetic media and assuming it covers this, when most of a shallowfake file is authentic. The second is going quiet: a flagged claim parked for three weeks punishes the honest claimant far more than the dishonest one, who has already moved on to the next insurer.
None of this makes fraud impossible. A genuine live recording can still show a staged scene, an unrelated vehicle, or missing context. Harder to manipulate is not the same as impossible to deceive, and every signal is a reason to look rather than a reason to decline. The operational detail is in shallowfake insurance claims and in what makes photo evidence hold up.
Venta Capture, a product of VentaVid, works on the provenance question directly: the claimant records in the browser at the moment of capture rather than uploading from a gallery, and the submission arrives with server-side receipt time, a SHA-256 fingerprint, session events and integrity signals attached for the reviewer to weigh.
