Claim review
How to spot AI damage photos in an insurance claim
A damage photo used to be the strongest evidence in a claim: only the policyholder was there. Image generators have removed that assumption. Insurers report dealing with AI-assisted fraud almost daily, from generated damage photos to invented invoices (BBC News; SAS). That calls for targeted review — and for restraint, because most submissions are genuine and no detector may accuse anyone of fraud.
Direct answer: five checks on a damage photo
- Find the oldest file and its source. A forwarded app compression, screenshot or chat export is not an original. Ask for the file as it left the camera or the claims app, with date and time.
- Ask for several angles with a reference object. One object of known size (coin, card, key) in frame makes scale and consistency checkable. Whoever really owns the object or was at the damage can supply that within minutes.
- Compare the damage with the vehicle or object. Does the damage pattern fit this model, colour and construction? A crumpled edge where no crumple belongs, missing or misaligned panels, badges or seams are strong indicators.
- Check details, reflections and background. Zoom into small text, number plates, mirrors, glass and repeated patterns. Impossible reflections, warped letters and details that shift between photos point to synthesis or editing.
- Ask for the original file or fresh captures. A short continuous video of the damaged part from another angle cannot be produced on request. A refusal is not proof, but it is a reason to escalate the file to a fraud investigator.
Why this is growing
Claims fraud is not new, but the cost of producing a convincing photo has fallen to almost nothing. Insurers and researchers also report the technique being used to inflate existing damage — exaggerating scratches, dents or water damage so a repair estimate rises (Debevoise Data Blog). Every claim paid on fabricated evidence is paid by the premium pool, so the pressure to check runs in both directions.
What a detector does and does not do
The image detector measures whether a file carries patterns common in synthetic imagery. That is triage: it tells you which photo deserves attention first. The score proves nothing about the cause, extent or existence of the damage, and nothing about the sender's intent. Genuine photos can draw a strong warning because of compression, HDR processing or a modern phone pipeline; edited or generated photos can score low.
Check the source chain, not just the image
- Provenance: ask for Content Credentials or original metadata before a screenshot or forward breaks the chain.
- Consistency: compare the damage photo with earlier photos of the same object in earlier policies or files.
- Proportionality: does the damage match the reported event, place and time?
- Process: who took the photo, when, and is there a purchase record, inspection report or earlier assessment?
How to handle a suspicion
Do not accuse anyone on one score. Document the five checks, record which files and versions you requested, and let a qualified claims specialist or fraud investigator take the decision. If fraud is suspected: preserve the whole file, request additional evidence and follow the internal fraud and complaints process. The claimant always retains the right to a substantive assessment of the claim.
If your own claim is disputed
Provide the original file, several captures from different angles and one photo with a reference object. Ask on what grounds the claim is disputed and which rule or signal was used. More on reading those signals is in the guide on how to tell if an image is AI generated; practice with the free checkers. Dutch readers find the same workflow on AI-schadefoto herkennen in een verzekeringsclaim.
Practical: what to do first
- Keep the incoming file unchanged and work from a copy.
- Record file metadata (camera, date, time, resolution) before converting anything.
- Request the extra captures and the original file in one message, with a concrete instruction.
- Use the detector as triage, not as a verdict, and document the outcome in the file.