Look closely at the photo attached to that refund request. The shattered tablet screen. The "melted" blender base. The receipt with the exact item, price, and date. All of it perfectly coherent. All of it wrong. Welcome to the newest job description for Stable Diffusion, Midjourney, and DALL-E: freelance evidence fabricator for the retail returns counter.
Image generators have become so good that producing a convincing "proof" photo now takes a single prompt and about five seconds. No design skills. No photography background. No expensive software. Just a text box and a grudge against return shipping fees. Security experts have started flagging this as a genuine retail risk: fake return photos generated by the same tools artists use to make surreal landscapes and anime portraits.
Retailers have leaned on photos and paperwork to validate returns for decades, and for a long time the logic held. Manufacturing believable fake evidence required skill, time, and money, which kept casual abuse in check. As an analysis published this week put it, the dam is breaking. A single prompt can now produce realistic receipts and damaged-product images that look identical to the honest photos legitimate customers send every day.
The numbers back the metaphor. The Merchant Risk Council found that 57% of merchants reported rising refund and policy abuse over the past year, and retail's global fraud problem already runs into the billions. Earlier research from Riskified concluded that nearly half of consumers now lean on AI when making return claims. A ReBound analysis of one million real orders flagged roughly 29 million pounds in potentially fraudulent activity, an industry-wide blind spot hiding in plain sight.
This is not a small-time grift. Fraud-detection firm Forter says it broke up a coordinated returns-abuse operation during the 2025 holiday season that used AI-altered damage images to push through fraudulent refunds worth $1.5 million. The playbook was simple: claim the item arrived broken, collect the refund, quietly resell the perfectly fine merchandise. When a single fraudulent claim is not caught, the abuser gets paid twice, once by the retailer and once by the next buyer.
Two Aisles in the Fraud Hardware Store
Retailers are now squeezed from two directions at once. On one side, organized fraud rings treat policy abuse as a scalable business. They spin up multiple accounts, coordinate claims across dozens of merchants, and test variations until something sticks. AI-generated images make every fake complaint look like the work of a model customer, and the rings can tailor a claim to a specific merchant's policies in minutes.
On the other side sits your neighbor. Friendly fraud, the consumer who returns something they damaged themselves or invents damage that never happened, used to require commitment. You had to actually break the thing or keep a closet of dead gadgets as props. Now the prop department is a graphics card. A shopper who wants to dodge a return fee, recover the cost of a spilled drink, or exploit a generous policy can create realistic supporting evidence in seconds.
The effects land hardest on small shops. Independent retailers told reporters this week that they trust proof photos more than ever, precisely because they cannot afford to question every claim. One convincing image can drain the goodwill, and the cash, out of a store already running on thin margins. The picture in front of them can be a lie, and they know it, but they also know that disputing a photo is a bad way to keep a customer.
The modern return-fraud playbook, courtesy of a diffusion model:
- Fake damage shots: a cracked screen, a dented laptop, a blender that caught fire. None of it real, all of it photorealistic.
- Invented receipts: documentation for products that were never bought, priced exactly right, dated exactly right.
- Cross-store testing: the same fake photo submitted to multiple retailers to see who blinks first and refunds fastest.
- Resale funnel: refund collected, merchandise resold, profit taken twice on a product that was never damaged.
The images do not even need to be good. Retail return desks rarely run forensic analysis on a blurry photo of a cracked phone. They check a box, process the refund, and move to the next queue, which is exactly what the prompt engineers are counting on. Speed and volume are the fraudster's friends, and AI hands them both.
So what is a retailer supposed to do? The obvious moves, shorter return windows, paid returns, slower refunds, punish honest customers more than they stop the fraudsters. A merchant who tightens every policy simply hands its best customers to the competitor across the street. Static rules and manual review cannot operate at the speed or volume that AI-powered abuse now demands.
The realistic answer is machine learning on both sides of the counter. Frauds are caught by context, not pixels. A damaged-phone photo from a five-year customer with one prior return means one thing; the same photo from a three-day-old account filing its fifth claim across eight retailers means another. Identity, device behavior, purchase history, and claim patterns tell the story the image tries to hide.
There is a grim poetry to the arms race. The same generative models that fabricate fake evidence are pushing stores toward detection models that judge the claimant instead of the claim. AI versus AI, with your refund in the middle. The tools that democratized creativity also democratized counterfeiting, and retail is now the testing ground.
For everyone using image AI honestly, none of this changes the fun. Generate all the art you want. But the next time a refund form asks for a photo of the damage, double-check that the damage exists in this timeline first. The algorithm approving your refund might not be able to tell the difference yet. It is definitely learning.
Comments