AI photos in refund and insurance claims

Photo evidence is no longer proof

Refunds and claims run on photos. A customer reports a crushed box, an undercooked meal, a cracked screen or water coming through the ceiling, and a photo backs it up. Most of those photos are honest. A growing number are not: food-delivery platforms in India and South Korea have described customers sending AI-made pictures of spoiled food to get refunds, and insurers now treat generated damage photos as a fraud risk of their own.

It shows up wherever a photo can turn into money:

• Food delivery: raw or undercooked meat, mould, a hair, a spilled or crushed order.
• Parcels and online shopping: a crushed box, a broken item, a missing part, the wrong colour.
• Travel and hotels: a dirty room, a broken fitting, a stained bed.
• Motor and property insurance: a dent, a cracked windscreen, water damage, a fallen tree.

Each fake is usually small. Getting the response wrong is not: pay every claim and you fund the fakes; doubt every photo and you fail the honest customers who make up most of your claims.

Two kinds of fake damage photo

It helps to separate them, because they need different checks:

• Fully generated: the whole picture comes from a prompt such as "a burger with raw meat in a delivery box" or "a crushed parcel on a doorstep". This is what Perspicax is built for.
• Edited: a real photo of the real order, car or room, with damage added by an AI edit tool: a crack painted onto a screen, mould added to bread, a dent added to a door. Most of the pixels came from a camera, so a detector may answer likely real or inconclusive. See how to tell if a photo was edited with AI.

A third kind needs no AI at all: an old photo, or one copied from the internet, sent again. A reverse image search and your own claim history catch most of those.

Real photoReal · camera photo

AI-generatedAI-generated · Nano Banana Pro

Food is exactly what refund fakes are made of, and generators are good at it. Steam, sauce, crumbs and reflections on a metal tray all come out right, and a little damage is one more line in the prompt. Neither of the photos above gives itself away by looking.

Process checks that work

The best defence is a claim flow in which a real photo is easier to submit than a fake one:

1. 1

   Capture the photo in your app

   Ask customers to take the photo in your app rather than upload one from their gallery. A live capture comes with the time, and with permission the place, and it rules out most files made somewhere else.

2. 2

   Ask for more than one angle

   Two or three photos of the same damage from different sides. A generator can make one convincing picture; a consistent set showing the same object from several angles is much harder.

3. 3

   Add something only the claimant could show

   The order receipt in the frame, a reference object for scale, or a short code shown in your app and written on a piece of paper. It ties the photo to this claim, now.

4. 4

   Ask for a short video for larger claims

   A few seconds walking around the item or the car. For high-value insurance claims, an adjuster can direct a live video call: "show me the other side", "move closer to the corner".

5. 5

   Check the records

   Compare with the order, the delivery time and the courier's own photo, the policy and its start date, and the claimant's history. An account that claims every week is a pattern, whatever the photos look like.

What looking closely can still tell you

Looking at a claim photo still helps, as long as you know which signs carry weight:

How Perspicax helps

Perspicax combines proven computer vision techniques with TrueEye 2.3, our state-of-the-art detector. It examines every region of each photo, down to the pixels, and answers in one of three bands:

It is built and measured on fully generated, photo-like images. In testing, it caught 98.6% of images from the newest OpenAI, Google and ByteDance models on average. Just as important for claims, it rarely flags real photos: fewer than 1 in 2,000 original camera photos were wrongly flagged (0.05%), and about 1 in 100 real photos shared on the web and social media (1.2%).

Route claims, don't auto-deny

A detector is a triage tool. Use its three answers to decide how much attention a claim gets, not whether the customer is lying:

• Likely real: carry on with your normal flow. Most honest claims land here and can move fast.
• Inconclusive: ask for another angle, a live capture or a short video, then decide.
• Likely AI-generated: send it to a person who checks the records and asks for more evidence. Don't refuse on the detector's answer alone.

The false-alarm figures above are small, but a large claims queue turns about 1 in 100 into real people. Give every customer a way to answer with more evidence, keep a person in the loop for any refusal, and never close a claim automatically on a single signal.

Claims, refund and trust teams can run the check on every uploaded photo through the Perspicax API, and use the answer to route claims in the systems they already have. To talk through a set-up, contact us.

Extra steps for insurance claims

Insurance claims are larger and slower than refunds, which leaves room for checks a delivery app can't make:

• Compare with photos taken when the policy started, if your process collects them.
• Check the date and place against outside records: weather reports for storm or flood damage, police or incident reports for road accidents.
• Ask for the repair estimate or invoice, and check that the garage or contractor exists and did the work.
• Look at the claimant's history across policies, and at claims that share photos, addresses or repairers.
• For high-value claims, inspect in person or by adjuster-directed live video.

Take care not to overcorrect. Slowing every claim down, or refusing on suspicion, harms honest customers and invites complaints. The aim is a fast lane for photos that hold up and a closer look for the rest.