Purpose: Explain how fraudsters in China are using AI-generated images and videos to win refunds, why current checks fail, what cases reveal about scale and motive, and what sellers, platforms, and regulators can do next. Interrupt: fake photos, fake videos, real refunds. Engage: honest sellers, honest buyers, and platform trust are on the line.
What happened — quick facts
Scammers began submitting AI-generated photos and videos as evidence for refund claims on Chinese ecommerce platforms. RedNote posts collected by WIRED show multiple merchants and customer service reps reporting doctored images. One buyer sent a photo of bed sheets supposedly torn to pieces; the shipping-label characters were gibberish. Another showed a coffee mug with cracks that looked like paper tears, not ceramic breaks. A live-crab seller on Douyin received fabricated videos of dead crabs; local police found inconsistencies—leg positions, sex ratios changing between clips, and one crab with nine legs—and detained the buyer for eight days. Outside China, Forter reports a 15% rise in AI-altered images used for refund claims since the start of the year. Organized groups have pushed this at scale, submitting large batches of fraudulent claims using rotating IPs and tight timing to overwhelm systems.
How the scam works — the mechanics
Fraudsters exploit the standard refund workflow: platforms ask customers for photos or videos as proof, and frontline reviewers often make quick decisions. AI tools now generate plausible damage images without needing real objects. The scammers focus where the process is weakest: fresh groceries, low-cost beauty items, and fragile goods like cups—items sellers often refund without return. The AI does not need perfect realism; it only needs to look convincing enough for a rushed reviewer. Fraudsters add operational tricks: rotating IP addresses, burst submissions to create system stress, and synthetic metadata to mimic real phones. The result: refunds paid out where no real damage or return occurred.
Why standard checks fail
Two forces collide here. First, platforms optimize for user experience: fast refunds increase conversion and reduce complaints. Second, frontline staff are overloaded and lack time or tools for forensic checks. AI images can be good enough to pass casual inspection while still bearing telltale signs that require careful scrutiny. Watermarks and EXIF data are fragile defenses—easy to strip or fake. That gap is where fraud grows.
Case study — the Douyin crab incident
Gao Jing’s family has farmed crabs for decades. She received videos claiming most crabs arrived dead and two had escaped. Close inspection found biological inconsistencies: legs pointing upward (unnatural for dead crabs), change in sex ratios between videos, and a crab with nine legs. Police confirmed fabrication and detained the buyer. This case matters because it led to enforcement action—proof that the fraud can meet real legal consequences. It also shows what careful human review can catch: biological and contextual details that AI-generated fakes often mishandle.
Scale and organized fraud
Forter’s data shows the trend is not limited to a few bad actors. AI-enabled image fraud spread rapidly after mid-2024 and is accelerating. Organized groups have submitted claims totaling over a million dollars in a short window for home goods, using tactical timing and IP rotation to hide identity and exploit automated refund flows. That level of coordination pushes this from opportunistic scams to industrialized fraud.
What sellers are doing now
Some merchants use AI to fight AI—running suspect submissions through detection tools or chatbots to flag doctored photos. Others add manual steps: asking for additional angles, time-stamped videos, or a photo with the order number written on a piece of paper. A few require returns for certain categories. Still, these measures have limits. AI detectors can be fooled; extra steps frustrate honest buyers; and platform policies don’t always favor the seller even when fraud is likely.
Why simple fixes won’t work
There’s no single magic fix. Tightening return rules reduces fraud but also hurts trust and conversion for legitimate customers. Adding manual review increases costs and slows response. Forensic metadata and watermarking are useful but fragile. AI detection tools lower some false claims but raise false positives that penalize honest shoppers. Any change creates trade-offs between fraud loss and customer experience.
Practical steps sellers can test today
Take action that keeps honest customers moving while raising the bar for fraudsters:
- Require time-stamped, multi-angle photos or a short video with the order number visible. This raises the difficulty for mass-fake campaigns.
- For high-risk categories (live food, fragile items, low-cost beauty), require return or partial refund unless the seller accepts instant refund. Make the rule public—consistency reduces disputes.
- Use velocity checks: flag multiple refund claims from the same region or IP cluster within a short period.
- Keep simple forensic checklists for CS reps: look for gibberish labels, repeating artifacts, inconsistent object counts, or biological impossibilities (like the crab leg/sex ratio mismatch).
- Share fraud indicators across sellers on the platform—social proof of fraud patterns helps everyone detect coordinated attacks.
Platform-level changes that balance risk and experience
Platforms must act with surgical precision, not blunt force. Options to consider:
- Risk-tiered policies: auto-refund small, low-risk orders; require returns or higher-proof for high-risk categories.
- Specialized fraud review teams with enough time to inspect suspicious claims flagged by AI detectors—combine machine speed with human judgment.
- Introduce friction for suspicious flows: short delay with request for live verification or live video chat. The friction should be targeted, not blanket.
- Strengthen collaboration with payment providers to halt large, coordinated refund requests and trace funds where possible.
- Invest in cross-platform intelligence sharing to spot coordinated campaigns that hit multiple sellers.
Technical defenses worth exploring
Technology can help—but thoughtfully. Suggestions that matter:
- Use ML models trained on known fakes and real damage photos from your category. False positives fall when you train on your own data.
- Require metadata from approved capture apps that cryptographically sign photos at the time of capture. This is stronger than EXIF data, which can be stripped.
- Add behavioral signals to fraud scoring: account age, past refund ratio, delivery timestamps, and device fingerprinting.
- Use randomized physical markers in packaging—simple QR or numbered slips inside parcels that buyers must show in photos. That raises the cost for fraud rings.
Legal and regulatory levers
The Douyin crab case shows enforcement can deter fraud. Platforms and sellers should work with authorities to classify mass AI-enabled refund schemes as economic crime when evidence supports it. Regulators can also mandate minimal standards: validated capture apps for evidence, penalties for organized abuse, and protocols for cross-border tracing of payment flows used in fraud.
Trade-offs and the human factor
We must accept trade-offs. Saying “No” to blanket easy refunds is sometimes right—sellers can refuse a quick refund when evidence looks fake. Saying “No” sets boundaries and forces more dialogue. But every added barrier costs honest customers time and trust. Platforms must choose where to place friction so that honest shoppers still get a smooth experience while fraudsters face higher costs and slower schemes.
Questions to start a real conversation
What will you try first? Which of these checks can you add without killing conversion? Which product categories in your catalog match the scam patterns—fresh food, cheap cosmetics, fragile items—and how would you change their return rules? How do you balance fast service and fraud prevention? Repeat: fast service and fraud prevention. Fast service and fraud prevention.
Final thought — keep trust alive
Ecommerce depends on trust. AI makes it easier to fake evidence and cheaper to run fraud at scale. That forces a rethink: better detection, smarter policies, targeted friction, and smarter collaboration between sellers, platforms, and law enforcement. If we refuse to accept fraud as inevitable, we must act—practically, measurably, and together. Will you test one small change this week and report what happens?
#EcommerceFraud #AIFraud #RefundFraud #ChinaEcommerce #FraudDetection #SellerTools #PlatformPolicy
Featured Image courtesy of Unsplash and Oxana Melis (-iKLhCsr1v8)
