A customer emails a home goods brand three days after delivery. The ceramic lamp arrived cracked, the message says, and two photos are attached: the base with a jagged line through the glaze, and the box beside it. The person handling support that morning, often the owner or one of two staff, has a few minutes to decide. The lamp is one of the brand's better sellers, its policy promises a refund or replacement for anything damaged in transit, and asking for the broken lamp back means paying return freight on a heavy, fragile item that will go straight in the trash. So the refund goes through, the photos are filed, and the queue moves on. Until recently that was a reasonable call. The trouble is that the crack may never have existed.
Generative image tools now let anyone add a tear, stain, dent or shattered corner to an ordinary photo of an intact product. For a small online brand, every "arrived damaged" claim that is decided from photos alone is now a judgment about whether the evidence is real, made by people who were never trained for it and who are measured on how fast they close tickets.
Photos were the cheap way to settle a claim
The photo-based damage claim exists because it is efficient for everyone. Shoppers expect easy remedies: in research from the National Retail Federation and Happy Returns, 82% of consumers said free returns are an important consideration when shopping online, and online purchases came back at a rate of 19.3%, well above stores. For bulky, fragile or low-margin goods, many brands decided long ago that asking for a picture was cheaper than paying to ship a broken item back for inspection. The photo was treated as proof because faking one well took skill and time.
That assumption has quietly stopped holding. In early 2026, Modern Retail reported on direct-to-consumer brands receiving claims with doctored images. The chief executive of bedding brand Boll & Branch described photos of torn sheets where the rip did not look like real cotton fraying, and one image still carried an AI watermark. Bogg Bag received a photo of a supposedly ripped tote that turned out to be an earlier customer photo with the damage added, and the tear did not match how the bag's material actually fails. A fraud-prevention executive quoted in the piece said such claims had gone from occasional to daily across the merchants that company serves. The same reporting describes AI being used to produce fake shipping receipts and police reports for packages claimed as lost, and language models drafting pushy messages meant to wear down support agents.
Scale is harder to pin down than the anecdotes. The NRF survey found retailers classify 9% of all returns as fraudulent, a figure that covers every kind of return fraud, not just image manipulation. Experts told Modern Retail that faked photos remain less common than older schemes such as doctored return tracking. But fraud platform Forter told Retail Gazette in June that AI-generated damage claims are now the fastest-growing form of return abuse, and the article describes fraud rings selling "returns-as-a-service" for a cut of each refund.
More orders are judged at a distance
The exposure grows with the share of sales that happen online, because every online order that goes wrong is adjudicated remotely, through a screen, on whatever evidence the customer chooses to send. That share has kept climbing since the pandemic spike faded.
Source: U.S. Census Bureau, Quarterly Retail E-Commerce Sales (via FRED)
View data
| E-commerce share | |
|---|---|
| 2019 | 10.3% |
| 2020 | 16.3% |
| 2021 | 14.6% |
| 2022 | 14.2% |
| 2023 | 15.3% |
| 2024 | 16.1% |
| 2025 | 16.3% |
| 2026 | 17.1% |
For a brand doing most of its revenue through its own site, nearly every claim arrives this way. And the incentives inside a small company push toward approving. Support is often staffed by generalists who also handle wholesale orders or social media, the cost of a single wrongly approved claim looks small next to the cost of an angry review, and nobody has the time to compare today's photo with a claim filed under a different name two months ago.
Where the standard defenses give out
Owners and sellers who have run into this describe a familiar set of responses, and each has a real cost. The most common is to require the item back before refunding. It works against the faker, who has nothing broken to send, but it penalizes honest customers and can cost more than the product. Bogg told Modern Retail it requires a return only when a defect looks suspicious, precisely because its bags are large and expensive to ship. Boll & Branch asked one claimant to show the damage on a live video call, and the claim went quiet, but a video call is not something a three-person team can offer on every ticket.
Platform protections are uneven. Marketplace sellers have described buyers sending altered photos through direct messages rather than opening a formal claim, which sidesteps the platform's return process and leaves the seller negotiating alone. Brands on their own storefronts have no platform at all, only whatever their helpdesk and returns app record.
AI-detection tools are the obvious next step, and they help at the margins, but they produce false positives and may become less reliable as image generators improve. A wrong accusation is its own cost. In one public seller forum thread, a long-time seller asked whether a buyer's damage photos were fake, and most replies pointed instead to thin packing. Real damage still happens, and a brand that starts treating every claim as suspect will lose customers it wanted to keep. The Boll & Branch chief executive put the tension plainly: "We can't assume everybody's guilty."
What has made the checks practical
The useful signals were always there; they were just too scattered for a small team to use. A claim can be compared against the customer's order and refund history, the carrier's delivery scan and package weight, the photos the warehouse took at packing, and the images submitted in earlier claims. Reviewing that by hand for every ticket was never realistic. What has changed is that the same class of models making fakes easy can also read a helpdesk ticket, pull the matching order and shipment records, check a submitted photo against past submissions for reuse, and flag inconsistencies such as a box shown as crushed when the carrier recorded normal weight and no exceptions. Helpdesk, storefront and shipping platforms now expose most of this data through standard connections, so the checks no longer depend on someone remembering to open five tabs.
A custom-built system can sit inside the existing support workflow and do that gathering for every damage claim: auto-approving the routine ones that match clean histories and plausible evidence, and routing the small set with conflicting signals to a person with the evidence already laid out. The decision on a borderline claim stays with someone who knows the customers and the product, and the brand can keep the generous policy that made the photo-based claim worth offering in the first place.