What claims fraud looks like now
Fabricating an entire scene is harder than it looks and rarely necessary. A claimant photographs their own kitchen, then adds water staining to the ceiling. The room is real, the address is real, the timestamp is real. One region is not.
The same pattern appears across motor, property and contents claims: an existing dent deepened, a crack extended, an item placed into a photograph of a room it was never in. Each takes under a minute and each defeats a check that produces a single number.
Two adjacent tiles above the line, the rest well below. That adjacency is typical: an inserted object rarely respects tile boundaries.
A workflow that fits a claims desk
-
Collect originals, not forwards
Ask for files from the device that took them rather than images pasted into an email or sent through a messaging app. Recompression removes the evidence before you see it, and the request itself is informative.
-
Score every image, not a sample
Batch the whole set. Fraud is rarely confined to the one photograph a handler happens to open first, and a set where one image behaves differently from the others is worth more than any single score.
-
Read the region map on anything above the line
Decide whether you are looking at a generated scene or a real scene with an alteration. Those are different conversations with the claimant and different referral routes.
-
Check the file's own credentials
Where a phone or an editing tool has left Content Credentials, that record is cryptographic rather than statistical and carries more weight than any score.
-
Record the result in the file
Export the report so the score, the bands, the region map and the date sit in the claim record. A decision made months ago needs to be explainable later.
Why the upload question matters here
Claim photographs contain other people's homes, vehicles, injuries and documents. Sending them to a third-party service for scoring turns a review step into a data processing arrangement, with a contract, a retention policy and a breach obligation attached.
A tool that scores in the browser avoids that entirely. The file is read from disk by the page and never transmitted, so there is no processor to add to a register and nothing to disclose in a privacy notice.
Where a score must not decide
A detector result is a reason to look harder. It is not a basis for declining a claim, and treating it as one creates a serious problem for the insurer as well as the claimant.
| Action | Score alone? | What is needed |
|---|---|---|
| Flagging a claim for closer review | Yes | The score plus the region map |
| Requesting original files or a site visit | Yes | A stated, specific concern |
| Referring to a fraud team | With care | The pattern across the whole set, not one image |
| Declining a claim | No | Independent evidence and a qualified examiner |
| Alleging fraud to a third party | No | A forensic report and legal review |
The false positive risk in this setting
Claim photographs are among the worst cases for a detector. They are taken in poor light on phones, often at night, frequently indoors, and then sent through a messaging app. Night mode, noise reduction and recompression all push genuine photographs upward.
Expect a meaningful rate of high scores on entirely honest claims. Build the process so a flag routes to a human who asks for the original file, rather than to a letter that accuses somebody of fraud.
What to record, and for how long
A claim decision may be reopened by an ombudsman, a regulator or a court years after it was made. Whoever reviews it will ask what the handler knew and when they knew it. A score with no surrounding evidence answers neither question.
Record the result rather than the judgement. The exported report carries the score, the band, the decision line, every tile value and the date, which together let somebody reconstruct the reading without needing the original photograph or the model that produced it.
Keep it with the claim, under the same retention period as the rest of the file. Do not keep a separate register of scores against named claimants, which turns a review step into a profiling exercise and creates an obligation you did not intend to take on.