Alibaba’s merchant tools generated 150 million images and video assets during Tmall’s Double 11 last November, a number the company published as an achievement. The claim is fair: the platform did not buy those pictures, it built the machinery that makes them and handed the machinery to its sellers. China’s marketplaces armed their merchants with generative imaging and are now absorbing the cost of what the merchants did with it. Six months on, a reporter at Shanghai’s Shangguan News pulled twenty product images at random from Chinese platforms and found generation traces in more than half. The trust signals these companies sell advertising against are the collateral: the photograph, the review, the dispute record.

The capability was distributed before the market asked for it. A June white paper from Taotian’s merchant platform, written with the trade title Tianxia Wangshang, surveyed close to a thousand merchant decision-makers and reported that more than 95% already use AI tools in daily operations. Read the provenance before the finding: a marketplace polled its own customers about a capability the marketplace supplies, which makes that number a distribution metric in a lab coat. Alimama, Alibaba’s marketing arm, puts the production cost merchants no longer incur at more than 40 million yuan a day. Photography was the last fixed cost inside a listing, and it has been removed.

Apparel absorbed the change first, because in clothing the photograph is the specification. A buyer reading a dress listing is reading the image for colour, drape, cut and finish, none of which the copy carries. Alibaba’s export-facing Pic Copilot let over 10,000 sellers produce more than 80,000 model images during a single Black Friday trial, collapsing the model, the studio and the shoot into one step. A generated model wearing a generated garment in a generated room is a photograph of nothing, and the listing does not say so.

A Guangzhou court spent this spring establishing what the platform’s own systems had not. A buyer paid 411 yuan for two tops and asked the shop directly whether the images showed the actual goods; the merchant replied that everything in the store was shot from the physical product. Every image on the page had been generated, no photograph sat behind any of them, and nothing carried a label. Guangzhou Intermediate People’s Court called it consumer fraud and ordered a refund plus 1,233 yuan, three times the price. The sum is trivial and the sequence is the point: the buyer had to ask, the merchant had to lie, and a court had to rule, with the marketplace absent at every stage.

Generation is priced by the thousand; detection is priced by the dispute.

Alibaba’s answer has been to build the second half of the machine and publish its throughput as evidence of control. At its merchant conference in January, Taobao and Tmall announced a forthcoming after-sales fake-image detection model and opened a one-click channel letting sellers flag a buyer’s photograph of damaged goods, with a confirmed flag feeding straight into refund and appeal decisions. That product exists because buyers acquired the same tools: photograph the real item, ask for a plausible seam split or stain, attach the result to a refund claim. Jing Daily reported the scheme hitting beauty hardest, and the mechanism transfers cleanly to apparel, where damage is a question of stitching and fade rather than function. Image evidence stopped working as evidence on both sides of the transaction.

The more revealing move was the retreat from judging any of it. In April 2025, Taobao announced a rule change under which the platform would no longer step into refund-without-return claims on its own initiative, leaving sellers to negotiate with buyers directly — one of several platforms reported to be making comparable moves around the same time. Merchants had asked for precisely this, and the grievance behind the demand was real. It also meant the platform stopped producing the one record that could have priced the problem, which is an adjudicated finding about which party lied. A marketplace that declines to arbitrate keeps the placement revenue and sheds the evidence base. The state has stepped partway into that gap: the national anti-fraud app now carries its own AI-content verification feature, and eight e-commerce platforms have signed a joint self-regulation pledge coordinated by Beijing’s consumer council. Neither substitutes for a marketplace ruling on the listings it profits from.

The strongest case against this argument is that nothing structural has happened. Merchants have flattered their goods since catalogues existed, and retouching, borrowed studio shots and generous colour correction all predate diffusion models. On that reading, generation and detection are two curves rising together, and the platform holding the larger dataset wins. For it to hold, one condition must be true: generated and photographed images have to stay separable categories. They do not. Liu Xiangdong (a pseudonym the outlet used for a researcher who builds detection tools) told Shangguan News that no model is fully accurate, and that a large share of commercial images are now real shots with AI retouching applied, which places them in ambiguous territory by construction. China has required visible and embedded marks on generated content since 1 September 2025, and the Shanghai sample was drawn nine months into that regime.

The exposure is measurable, and it sits in customer management revenue: the fees merchants pay Alibaba for placement. That line ran to 82.5 billion yuan in the June quarter alone, and every yuan of it assumes a buyer who trusts a listing enough to click and then to keep what arrives. If the share of generated imagery keeps climbing while adjudication stays devolved to sellers, then the platforms will be selling ranked access to a catalogue their own buyers have learned to discount, and the discount will surface in what merchants will pay per click. None of this arrived from outside. A marketplace can make the picture cheap or it can make the picture credible; Taobao made it cheap, published the number, and is now buying credibility back one dispute at a time.