Short answer: Most photo fraud in claims and inspections is not sophisticated forgery — it is reuse. A photo of another property, an old photo of the same property, an image pulled from a listing, a photo of a screen, or now an AI-generated image. Verifying at the moment of capture, rather than inspecting the file afterwards, is the only control that reliably blocks all five.
The five patterns
| Pattern | What it looks like | What defeats it |
|---|---|---|
| Wrong property | Real damage, different address | Location captured at the moment of capture and bound to the file |
| Wrong time | Pre-existing or post-repair damage submitted as new | Trusted capture timestamp, not file metadata |
| Downloaded image | Listing, social, or stock photo | Live session capture; no upload path for unverified images |
| Rebroadcast | Photo taken of a screen showing another photo | Screen-detection signals: moiré, refresh banding, flat depth response |
| Synthetic | AI-generated damage | Provenance chain from a real camera sensor; generative-artifact scoring |
Why EXIF checks are not enough
EXIF is user-writable. Free tools rewrite GPS coordinates and timestamps in seconds, and most messaging and email paths strip EXIF entirely, so its absence proves nothing either. Any control that depends on trusting the metadata inside an uploaded file is checking a claim the submitter controls.
Capture-side verification
The alternative is to make the photo trustworthy at the instant it is created, inside a live session, and to record signals the submitter cannot rewrite afterwards:
- Session binding — the image exists because a live session existed, with both parties on the call.
- Device-reported location at capture, compared against the expected site anchor. See GPS geofencing.
- Server-side time, not the device clock.
- Cryptographic hash written at upload so any later change is detectable.
- Content credentials attached per the C2PA standard, giving downstream reviewers a verifiable provenance chain.
- Signal scoring for rebroadcast and synthetic-image indicators. See fraud and tamper detection.
Score, do not just block
Hard blocking creates false positives — a legitimate photo can lack camera metadata because of a browser path, and a warehouse interior can lose GPS lock. The workable design is a per-category score with advisory warnings for missing signals and hard flags only where a signal actively indicates manipulation, plus a reviewer view that shows why the flag fired. That keeps honest field users moving while giving SIU something concrete.
What reviewers should see
Every image in the file should answer four questions on one screen: when was it captured, where, by whom, and has it changed since. That is the practical definition of chain of custody, and it is what turns a photo into evidence in a dispute. See audit trail and admissibility of verified media.
The four ways photo evidence goes wrong
| Method | What it looks like | What defeats it |
|---|---|---|
| Recycled imagery | Damage photos from a prior loss, another property, or the internet | Session-bound capture with server-side time and location |
| Edited pixels | Retouched or composited damage | Ingest hashing plus signed provenance |
| Metadata manipulation | Rewritten EXIF date or coordinates | Never trusting client metadata as authoritative |
| Generated imagery | Synthetic damage from a diffusion model | Provenance, because detection is an arms race |
Note that three of the four are invisible to a reviewer looking at the picture. This is the core problem with photo-based intake: the file arrives with no way to interrogate where it came from, and reviewers are asked to judge authenticity by eye.
Why detection alone loses
Forensic detection — noise analysis, compression artefacts, model fingerprints — improves, and so do the generators. Every detection technique is a rule that the next model learns to satisfy. Provenance inverts the problem: instead of asking "can I prove this is fake?", you ask "can you show me this is real?" A signed record produced at capture time does not degrade as generators improve. That is the practical case for content credentials in claims intake, and it is why carriers are moving verification upstream to the moment of capture rather than downstream into review.
Designing intake so fraud has nowhere to enter
- Capture, do not collect. Photos originate in a live session opened from a texted link; nothing is uploaded from a camera roll. Virtual Inspection Pro guests join in the browser with no app install, which removes the usual excuse for falling back to email attachments.
- Authoritative time. Server-side, not the device clock.
- Anchored location. Position plus accuracy radius compared against the loss address — see how GPS verification works and where it fails.
- Integrity at rest. Hash at ingest, keep originals immutable, hold annotations as a separate layer.
- Complete logging. Every view, download, and export recorded in the audit trail.
- Human review of contradictions only. Missing signals are advisory; a precise fix hundreds of miles away is a flag.
What "verified" should mean on a report
A verification badge is worthless if nobody knows what it asserts. State it plainly on the document: this image was captured in session X at server time T, at coordinates within N metres of the loss address, and its bytes match the signed original. Anything the system cannot assert should be shown as unverified rather than quietly omitted — an honest partial record is defensible, a badge that overclaims is not. See tamper detection for the signals behind the state.
Handling a suspicious file without accusing anyone
Most anomalies are innocent: denied permissions, an indoor fix, a photo taken before the claim was opened. Build the process so the first response is a targeted retake request — one specific shot, on a link, with a reason — rather than a referral. Genuine claimants complete it in minutes; the ones that do not are the signal. That single step keeps false positives from becoming customer-service incidents while still narrowing the funnel to real cases for the SIU.
FAQ
Can AI-generated damage photos be detected reliably?
Detection alone is an arms race. Provenance is the durable answer: prove the image came from a real sensor in a real session at a known place and time, and detection becomes a secondary check.
Does verification slow the field user down?
No, when it runs inside the capture flow. The user takes a photo; the verification happens around it.
What about photos the customer already has?
Accept them, but label them clearly as unverified supporting material so they are never mistaken for session evidence.
Related: how fraud prevention works, warranty claims, insurance claims.
How common is photo fraud in insurance claims?
Insurance fraud is a long-standing, well-documented cost across the industry — the Coalition Against Insurance Fraud publishes ongoing research on its scale — but reliable figures specifically isolating manipulated photo evidence are scarce. Treat it as a real and growing exposure rather than a quantified one.
Can you tell if a claim photo is AI-generated?
Sometimes, and less reliably every year. Provenance at capture is durable in a way detection is not.
Does EXIF data prove when a photo was taken?
No. EXIF is user-writable and routinely stripped in transit. Use server-side capture records instead.
What is verified capture?
Photos taken inside a live session by the application, bound to session identity, server time, location, and a content hash, so origin and integrity can be checked independently.
How should adjusters handle a flagged photo?
Request one specific retake with a stated reason before escalating. Most flags resolve immediately, and the ones that do not are far better qualified for investigation.
Keep reading
- C2PA content credentials — signed provenance explained.
- GPS-verified photo capture — accuracy, spoofing, and advisory design.
- Remote claims adjusting with live video — the workflow around the evidence.
- Chain of custody for inspection evidence — the full record.
- Virtual home inspection for insurance — how carriers use live video.