Refund request RF-2026-771940Photo 1 of 2
Cardboard parcel on a doorstep, one corner torn open
IMG_2207.JPG, 18/03/2026, 21:52Capture data: edited
AI tracesManipulation traces
Refund request RF-2026-771940Photo 1 of 2
IMG_2207.JPG, 18/03/2026, 21:52Capture data: edited
What we readWhat the eye sees
Logged to the requestAnalysis kept, reason written
ConclusionSuspect
Justifications
AI tracesHigh
Manipulation traces, torn corner
Capture data edited
No generator signature found

A steady flow of refund requests,
and one photo to decide each.

For each “damaged on arrival” request, the agent has one photo, an order history and little time. They have to choose between refunding straight away and asking for the item back, with a reason to record either way.

What it does

Checks the photos and receipts sent with a refund request and tells the agent whether AI produced or edited them, in words they can paste into the ticket.

Why it matters

A crushed box or a cracked screen now takes a few seconds to generate, and a refund granted on a picture is money that rarely comes back.

Who it’s for

Customer service teams at retailers, marketplaces and carriers handling damaged, missing or not-as-described claims, and the outsourced support teams who run them.

The problem

Refund rules judge the customer and leave the photo unchecked

Order history, return rules and abuse scores look at who is asking and how often, and a convincing photo of a damaged parcel sent from an account with a clean history gives them no reason to hold the refund.

The box that arrived intact

A genuine photo of a delivered parcel, with a torn corner generated into it. The label, the doorstep and the doormat are all real, which is what makes the picture so convincing.

The product that was never broken

A photo of the item itself, with a crack, a stain or a missing part added by a model. The order is real and the product was delivered, and only its condition has been rewritten.

Smartphone with a cracked screen in an open delivery boxWhat your eyes see
What we see

The parcel arrived, the phone was intact, only the crack across the screen was generated.

The series

Accounts, addresses and products all differ, yet the same model sits behind every picture. Each request looks ordinary on its own, and the link appears once you can read how the photos were produced.

In figures

Refund abuse has become a cost line of its own

up to£5.76bn

a year is what refund fraud could cost UK online retailers, according to the first estimate of its kind.

Source: Cifas and University of Portsmouth, May 2026
9%

of returns in the United States are fraudulent, on nearly $850bn of goods expected back in 2025.

Source: NRF, October 2025
44%

of UK businesses say they are affected by returns and refund abuse.

Source: Forter, reported by Retail Gazette, June 2026

Our answer

Where Label4.ai fits in your process

Label4.ai adds one question to the checks you already run, and leaves the rest of your process as it is.

Request received

Help centre form, chat, e-mail or marketplace dispute, whatever your intake already is.

Your existing checks

Order history, return rules, the abuse score you already run. Unchanged.

Label4.ai analysis

Is this picture what it claims to be? Checked automatically as it arrives, in seconds, with its reasons.

Decision

Refund, ask for the item back, or pass to the abuse team, with the reason written down.

A result anyone on the team can read

The agent sees the conclusion first, then three sentences written in the words of the job, like these.

  • This photo carries traces of a generative model, and they are strongest on the torn corner.
  • The corner of the box was modified after the picture was taken, which is where the damage appears.
  • No generator signature was found, so the conclusion rests on forensic analysis alone.

How we do it

What sits behind each conclusion

  • An analysis shaped to your process Each project is set up for the way you work, with the checks that matter for your content, your threshold and the action we recommend. Every analysis looks first for a signature left by the generator, then for the traces a model leaves and for signs of editing.
  • Detection that keeps the process running Most requests are genuine and should be refunded quickly, so before the contract we run your own content and agree with you on the false alarm rate that suits your process. That is why we quote no generic detection rate: we set up a project and tune it to give you the best detection rate on your own use case.
  • A conclusion that holds up The same content, analysed twice, gives the same report twice, which turns a series of requests from suspicious into provable.

It runs in our cloud or on your own servers, through the Anomaly Engine console or the API.

See the technology in detail

Your content

See what your own requests reveal

You send a batch of refund photos whose outcome you already know, including the frauds you caught and the ones you found out about too late. We come back with a result for each picture, your false alarm and missed fraud rates, and a plain statement of what we did not see.

We reply to every request within two working days, under NDA if you prefer.