Document fraud detection

Detect a doctored claim photo

Water damage, broken glass, car bodywork: a photo is often the first evidence of a claim. With editing and artificial intelligence tools, creating or exaggerating damage in a picture has never been easier. Docusan analyzes every photo submitted and flags manipulations.

What a claim photo proves, and what it does not

A photo answers only one question: what does this image show? It does not say by itself when it was taken, whether the damage comes from the reported incident, or whether it affects the insured item. It is a clue to weigh against the rest of the case file, not stand-alone proof.

Rules vary by country and policy. In France, for example, policyholders have the period set by the contract to report a loss, at least 5 working days, or 2 for theft. They gather supporting evidence, including invoices and photographs, and the insurer can appoint an expert to check the circumstances and assess the damage.

The stakes are real. ALFA, the French association against insurance fraud, reports €946 million of fraud identified by its members for the 2025 financial year, including about €627 million in property and casualty insurance. These figures measure what insurers detect rather than the true extent of fraud, and they do not separate photos from other evidence.

What the file can tell you: metadata, provenance, AI marking

Exif metadata. A photo can record date and time, camera settings and even a geographic location. You can compare this with the claim. It remains fragile: it can be missing or altered, and messaging apps and screenshots often remove it.

C2PA provenance. Some files carry Content Credentials, a signed history of the image. Validating them tells you whether that history is intact and who signed it, not whether the image is true. Their absence alone means little: this marking is optional and can be removed.

AI marking. Since 2 August 2026, the EU AI Act has required providers of generative systems to mark their output in a machine-readable format, with systems already on the market given until 2 December 2026. A missing mark still proves nothing: older content need not be labeled, and standard editing functions are excluded.

In doubt about a document?

Analyze it for free: Docusan flags the anomalies and locates them on the document.

Try it for free

How to detect manipulated images in damage claims: a 6-step method

  1. Ask for the original file

    A forwarded copy or a screenshot loses some of its information. Ask for the photo as it came out of the device, along with several views of the same damage: close up, from a distance, with a landmark from the surroundings.

  2. Compare the image with the claim

    Do the place, the item, the season and the time of day fit? Number plate, model, room, furniture: these details should match the policy, the accident report and the other documents on file.

  3. Look closely at the scene

    Are the shadows and reflections consistent with each other? Does the shape of the damage fit the stated cause? Is the grain the same around the damaged area? In a generated image, text and number plates are often approximate, but these flaws are fading: our guide on AI-generated documents explains the limits of the eye.

  4. Read the file’s properties

    Our free tool reads an image’s software, device, dates and any C2PA marking in your browser, without sending anything. A date before the reported incident, or an image editor, raises a question rather than a conclusion.

  5. Check whether the photo already exists

    A reverse image search finds photos already published online. Compare the image with your own files too: has the same photo, or the same damage, already been used on another claim or by another policyholder?

  6. Have it verified if the stakes justify it

    Serious doubt calls for a new photo, an expert assessment or a site visit. Cross-check the photo with the repair invoice, which should match the damage shown.

Water damage, broken glass, vehicles: typical inconsistencies

None of these signs is proof on its own; several together justify a closer check:

  • Vehicles: an impact on the opposite side from the one described in the accident report, a number plate or color that does not match the insured vehicle, rust in a scratch that is supposed to be recent.
  • Water damage: a mark unrelated to the stated source, a stain already dry and old on a photo supposedly taken that day.
  • Broken glass: an impact in the wrong place for the stated cause, a reflection that crosses the crack without distorting.
  • Theft: a photo proves at best that an item existed, not that it disappeared. Compare it with the invoice and the police report.

Honest edits, experts, privacy: special cases

Not every edit is fraud. Phones correct the light on their own, and a policyholder may crop or brighten an image to make it legible. The right question is whether the change alters what the photo proves.

A policyholder’s photo and an expert’s report do not carry the same weight. In France, the expert is appointed by the insurer, and the policyholder can obtain a copy of the report; if the contract allows it, they can commission a second assessment at their own cost. You must therefore be able to explain your conclusions about an image.

A photo sometimes shows more than the damage: a face, the inside of a home, a number plate, and its metadata may include a location. Limit access to the people handling the claim.

What to do if a claim photo looks doctored

Do not reject a claim over a single anomaly: ask for other views, the original file or an expert assessment, and record what you checked. Our guide on verifying a document details this approach.

According to ALFA, the French Insurance Code sets no explicit penalty for claims fraud, but most contracts provide for loss of cover in the event of an intentional false statement, and the insurer can reclaim a payment already made. Under French criminal law, prosecutions mainly rely on fraud, punished by five years in prison and a €375,000 fine, and forgery, punished by three years in prison and a €45,000 fine.

ALFA also gives its members access to a network of certified investigators. What happens next is for your claims and legal departments, under the rules of your own country. Our insurance page shows these checks within claims handling.

Analyzing claim photos as volumes grow

The eye and the file’s properties have their limits: a careful edit can leave no visible trace. Docusan, a French document fraud detection service, analyzes the photo you receive, as a JPEG, PNG or TIFF, along with the PDFs and scans in the same file, and gives a verdict: No detection, Warning or Alert. The anomalies are located on the image, with a comment to understand them, and the decision always stays with your claims handlers.

Docusan analyzes the file: it does not contact the policyholder or query government databases, and it does not replace an expert. You can use it from the platform or integrate it into your claims management tool with its API, with the Enterprise plan. Documents are hosted in the European Union and are never used to train our models (security and privacy).

Sources

Typical frauds

The most common forgeries

01

Added or exaggerated damage

A scratch, crack or water stain is added or enlarged by retouching a genuine photo.

02

AI-generated image

A damage photo is entirely created with an image generation tool, without any real incident.

03

Repurposed photo

An image from another context is presented as evidence of the declared claim.

Our detection approach

What Docusan analyzes

Four levels of analysis to spot forgeries, whether the document is a PDF, a scan or a photo.

01

Metadata

  • Missing metadata
  • Inconsistencies in the file properties
02

Pixel-level anomalies

  • AI-generated images, deepfakes
  • Inserted content (splicing, inpainting, copy-move…)
03

Structural anomalies

  • Edits and annotations
  • Recovery of previous versions
04

Customizable detectors

  • Consistency with document templates
  • Data cross-checking
  • Watchlists

Relevant industries

The result

What a report looks like

Each document gets a verdict: No detection, Warning or Alert. The anomalies are located on the document, with a comment to understand them.

Alert

  • Modified area of the document
  • Inconsistent file properties
  • A comment for each clue
Illustration with a fictitious document. The clues depend on the document analyzed.

FAQ

Your questions

How do you verify photos in insurance claims?

Ask for the original file and several views, compare the image with the claim, look at the shadows, reflections and grain around the damage, read the file’s properties and check whether the photo has already circulated. Serious doubt calls for an expert assessment.

How do you prevent the same image being used on several claims?

Compare each photo you receive with your earlier files, for the same policyholder and for others, and run a reverse image search for photos that may have circulated online. Keep the original files too, as they are easier to compare than copies.

Is a claim photo without metadata suspicious?

Not in itself. Messaging apps and screenshots often remove it, and some devices write very little. Ask for the original file, then judge the photo together with the rest of the claim.

Is an edited claim photo always fraud?

No. An automatic light correction, or a crop to make the image legible, is not fraudulent. Fraud requires an intent to deceive: what matters is whether the change alters what the photo proves.

Does Docusan detect AI-generated photos?

Yes. Docusan identifies images generated by artificial intelligence. Like any detector, it provides a verdict to interpret, not absolute certainty.

Can a photo sent through a messaging app be analyzed?

Yes. For the most complete analysis, ask for the original photo rather than a copy forwarded through a messaging app.

How do claims handlers see the anomalies?

The viewer highlights the suspicious areas on the photo and shows the verdict, so the handler can quickly decide whether an expert assessment is needed.

Try it for free

High volumes?

Automate your checks

A subscription with included pages from the Pro plan; the API, dedicated infrastructure and customized detectors with the Enterprise plan.