Industries / Technology / Marketplaces and gig platforms
Fakes and scams
Keeping counterfeit and unsafe goods, fake sellers, fake orders and scams off a marketplace, and finding the people behind them.
6
Milestones achieved, from 2002 to 2026.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2002
Software screens listings for fakes
By May 2002 eBay ran a "fraud engine", software with rules and models that searched listings for breaches of its policies. It looked at signals such as the seller's internet address, problems with the seller's account and the feedback the seller had received, not only at keywords. Alongside it, eBay's Verified Rights Owner (VeRO) program let a brand owner file a notice against a listing it believed in good faith infringed its copyright or trademark; eBay's practice was to remove the listing within 24 hours, and it removed 70% to 80% of them within 12 hours. These facts are the federal courts' findings in Tiffany v. eBay (district court 2008, appeal court 2010).
Credit
eBay; findings of the US District Court for the Southern District of New York, affirmed by the US Court of Appeals for the Second Circuit
2016
Taobao returns to a blacklist
The US Trade Representative's Notorious Markets List names markets reported to deal in counterfeit and pirated goods; USTR says it is not a finding that the law was broken. USTR named Taobao, a consumer-to-consumer marketplace run by Alibaba, in its December 2011 list. On 13 December 2012 it removed Taobao for "notable efforts" to work with rights holders. In December 2016 it named Taobao again and called the levels of reported counterfeiting "unacceptably high", although Alibaba said it had removed more than 380 million listings and closed about 180,000 Taobao stores in the 12 months to August 2016.
LessonA platform's own count of takedowns did not settle whether fakes had become rarer; USTR judged by what rights holders still reported.
Credit
Office of the US Trade Representative
2020
Seeds nobody ordered
From July 2020 people in all 50 US states reported packets of seeds they had not ordered, often labeled as jewelry or toys and appearing to come from China, Taiwan or Singapore; people in Canada, the UK, France, Japan, Brazil and other countries reported the same. The US Department of Agriculture's Animal and Plant Health Inspection Service (APHIS) evaluated thousands of reports, and officials said they believed the unordered packets were part of a "brushing" scam: a seller ships cheap items to real addresses so it can record sales, which raises its rating and its place in search results. APHIS found no evidence of intentional harm and said it was working with e-commerce companies to remove sellers that illegally import seeds.
LessonFake orders made to game a marketplace's rankings can land on strangers' doorsteps, which makes ranking fraud a customs and plant-health problem too.
Credit
USDA APHIS, as reported by Food Safety News in February 2021
2021
A marketplace removed from search
Inspectors of France's consumer protection agency, the DGCCRF, analyzed more than 140 products sold on Wish and found 95% of toys non-compliant (45% dangerous) and 95% of electrical appliances non-compliant (90% dangerous). The agency ordered Wish's operator, ContextLogic, on 15 July 2021 to comply; when it did not, on 23 November 2021 the DGCCRF ordered the main search engines and app stores to stop listing Wish. The law firm CMS describes it as the first use of that power, created in December 2020. A Paris court refused to suspend the order on 17 December 2021. The DGCCRF lifted it in March 2023, after ContextLogic took measures on recalls and on keeping out products similar to ones found dangerous; the agency had found withdrawn products reappearing under other names.
LessonTaking a listing down does not keep a product off a marketplace if it can come back under a new name.
Credit
DGCCRF and the French Ministry of the Economy
2023
Big sellers must identify themselves
The US INFORM Consumers Act took effect 180 days after it became law on 29 December 2022, that is, on 27 June 2023. It requires an online marketplace to collect a bank account, contact details and a tax identification number from each "high-volume third party seller", one with 200 or more sales of new or unused consumer products and at least $5,000 in revenue in any 12 months of the past 24, and to verify them within 10 days. Sellers with $20,000 or more a year in sales must have their name, address and contact details shown to buyers, and a marketplace must suspend a seller who does not comply after notice.
Credit
US Congress (Consolidated Appropriations Act, 2023, division BB, title III)
2026
EU fines two marketplaces
The EU's Digital Services Act requires the largest online platforms to assess and reduce "systemic risks", including the sale of illegal products. On 28 May 2026 the European Commission fined Temu €200 million because its 2024 risk assessment "seriously underestimated" how often EU shoppers meet illegal items. Its evidence included mystery shopping by an independent testing organization, in which "a very high percentage" of the chargers selected failed basic safety tests and a high percentage of baby toys posed medium to high risks. On 20 July 2026 it fined AliExpress €550 million, the largest fine under the act so far, over illegal, unsafe and counterfeit products, and found that commitments AliExpress gave in June 2025 lacked measurable targets and outside checks.
LessonThe regulator tested the marketplace's own risk paperwork against test purchases by an outside lab.
Credit
European Commission
Watch for
Temu's repair plan
In the action plan that the European Commission accepted on 26 October 2026, Temu promises to redo its risk analysis and keep watch on its measures against illegal products, and the Commission says Temu will review how its recommender systems "may amplify the spread of illegal goods". Temu calls the plan "a clear testament" to its spending on compliance.
Achieved whenAn independent audit, or a repeat mystery-shopping test for the Commission, finds a lower share of failing chargers and baby toys on Temu than the test behind the May 2026 fine.
Claimed by
Temu, October 2026
Shein under investigation
On 17 February 2026 the European Commission opened formal proceedings against Shein under the Digital Services Act. It is examining Shein's systems for limiting the sale of illegal products, including child-like sex dolls, its rewards for engagement and other addictive design, and how open it is about its recommender systems.
Achieved whenThe Commission adopts a non-compliance decision against Shein or accepts binding commitments from it.
Claimed by
European Commission, February 2026
Open
01
Scam chats for testing
An open set of made-up chats between buyers and sellers on a second-hand marketplace, in at least five languages, each labeled with its scam pattern: a fake payment confirmation, an overpayment followed by a request to refund the difference, a fake courier or "protected payment" link, a deposit for an item that does not exist, or a request for the other party's account verification code. Half the chats are honest look-alikes, such as a buyer asking to pay through the platform's own checkout, so a detector is scored on false alarms as well as on catches. Requests to move a deal off the platform belong to the Matching data set above and are left out here.
The briefCredit · unclaimed02
A generator of fake orders
An open simulator that produces a marketplace's order, rating and refund records for made-up shops and shoppers, with known fraud mixed in: brushing (fake orders that lift a shop's rank), rings of accounts that buy from each other and rate each other, refund abuse, and seller accounts sold to new owners. Every record carries its true label, so detection methods can be compared on the same data.
The briefCredit · unclaimed
Open 01 · the brief
Scam chats for testing
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open set of made-up chats between buyers and sellers on a second-hand marketplace, in at least five languages, each labeled with its scam pattern: a fake payment confirmation, an overpayment followed by a request to refund the difference, a fake courier or "protected payment" link, a deposit for an item that does not exist, or a request for the other party's account verification code. Half the chats are honest look-alikes, such as a buyer asking to pay through the platform's own checkout, so a detector is scored on false alarms as well as on catches. Requests to move a deal off the platform belong to the Matching data set above and are left out here.
Why it is still open
The open synthetic sets found cover phone-call scams (BothBosu's multi-agent scam conversations) or multilingual fraud chats whose scam types are banking, investment, phishing, romance and tech support, with no marketplace or purchase type (firelibrary's multilingual fraud chat set). The one set built for marketplace scams, Kled's, holds real screenshots and is licensed commercially rather than open for download.
Built with what exists
Public scam warnings from consumer agencies and from marketplaces' own help pages, as the list of patterns; an AI model to write the chats, such as Mistral Large 4, whose weights are free to download (see the Note "A trillion open parameters"), or Claude; the Hugging Face Datasets library for release; a scikit-learn text classifier as a baseline.
First step
Write the list of patterns from ten public warnings, write one English example chat per pattern, and have two people label the examples without seeing the answers.
Done looks like
5,000 chats in five languages, half of them honest; two people check 300 chats picked at random and agree with the label on at least 95%; catch rate and false-alarm rate per pattern published for the baseline and for three AI models.
Open 02 · the brief
A generator of fake orders
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open simulator that produces a marketplace's order, rating and refund records for made-up shops and shoppers, with known fraud mixed in: brushing (fake orders that lift a shop's rank), rings of accounts that buy from each other and rate each other, refund abuse, and seller accounts sold to new owners. Every record carries its true label, so detection methods can be compared on the same data.
Why it is still open
The nearest public benchmark, Shopee's 2020 order-brushing challenge, covers one pattern defined by a fixed rule. The open agent-based fraud simulator found, published in August 2026, models generic honest and fraudulent users, about 500 agents, without marketplace fraud patterns or labels for scoring.
Built with what exists
Python with Mesa, an agent-based modeling library; NetworkX for account rings; scikit-learn and PyTorch Geometric for baseline detectors.
First step
Code the brushing pattern from Shopee's challenge rule into a generator of 10,000 orders, and check that the challenge's rule finds the planted shops.
Done looks like
Six fraud patterns and one million orders generated from a published configuration; three baseline detectors scored per pattern by precision and recall; a second team regenerates the data from the configuration and gets the same scores within 1 percentage point.