NewThe detectors that scored perfect collapsed the hardest under attack.
Back to publications
Fraud story

The AI-edited claim photos insurers are already catching.

Insurers are not waiting for this threat. Admiral's own fraud examples are AI-edited damage on real photos, and a national study puts the gap in one line: a third of consumers would alter a claim image, and only a third of insurers are confident they could catch a deepfake.

Article · September 2026 · 5 minute read

A claim photo used to be evidence. With a phone and a free editing tool, a claimant can now add damage to a real photograph that never happened, and insurers already know it. The numbers they and their researchers publish are the clearest picture of the threat, and the shape of what they are catching is not what most detectors are built to find.

The gap in one line

Surveying a thousand US consumers and three hundred claims professionals, Verisk's 2026 State of Insurance Fraud study produced two headline numbers that carry the whole argument. On the consumer side, 36 percent said they would consider digitally altering a claim image, rising to 55 percent of Gen Z. On the insurer side, only 32 percent were very confident they could identify a deepfake.

So a third of your claimants would edit a photo, and only a third of your fraud teams are confident they would catch it. Nor is that a forecast: in the same study, 98 percent of insurers said AI editing is fuelling fraud and 99 percent had already encountered AI-altered documentation. It is the current state, described by the people paying the claims, and reporting on motor claims describes the same thing from the outside, with AI images used to fake evidence of vehicle damage.

What Admiral caught

The UK insurer Admiral put concrete examples on the record. It detected 86.8 million pounds of fraudulent claims in 2025, up from 50.9 million the year before, and its claims handlers attribute part of the rise to AI editing. Read together, the examples it published amount to a taxonomy of the attack:

  • the same damaged car edited with a different number plate to claim twice
  • damage to the rear of a vehicle added in Photoshop
  • damage added to a pair of designer shoes in Photoshop
  • a designer watch made to look damaged with an AI filter
  • photos of damaged luggage sourced online and submitted as the claimant's own

One caveat is worth stating plainly, because most coverage gets it wrong: the 71 percent rise is total detected fraud, not AI fraud, and Admiral does not publish an AI-versus-traditional split. The point of the list is not a growth rate. It is the shape of what is arriving.

Edited, not invented

Look again at Admiral's examples, and notice that almost none of them is a wholly synthetic image. They are localized edits: damage added to a real photograph, a plate swapped, a filter applied to a genuine object. That matters because the public conversation about "AI fraud" assumes fully generated pictures, and, assuming that, teams buy detectors tuned to spot a generated image rather than a real one quietly altered in one region.

Asked only whether the whole image was generated, such a detector passes a real photo with a single edited dent every time. The question that actually holds is where an image was changed, not just whether it is synthetic, and it is the same lesson that shows up across detection: reading the wrong property, a control passes the attacker, however well it scored on a clean test. See how detectors collapse once the test stops being clean.

What it means for claims

If your claims flow accepts an uploaded photo as proof of damage, this is your threat model, not a curiosity. A genuine-looking image is no longer evidence on its own, and vendor benchmarks age faster than the editing tools your claimants already carry on their phones. Retailers are meeting the same attack from the other direction, in refund photos of damage that never happened, and their tell is the one an insurer can use too: the edit looks right and behaves wrong. The only honest answer to whether your checks would catch an edited claim photo is a measured one, run against current tooling on imagery that looks like your real claims.

Margen does not sell claims software or a detector. We are an independent third party that red-teams these checks under adversarial, realistic conditions, and reports where they hold and where an edited photo gets through, so the fix goes to the team that owns the model.