Alibaba’s marketing arm hands merchants an image generator and advertises the result as production cost they no longer have to spend. Taobao, inside the same company, now pays engineers to find the pictures that tools of that kind produce. Catalogue automation was sold as the deletion of a cost line; it deleted a one-off expense and installed a standing one.

The scale of the generation is not in dispute, because Alibaba publishes it as a selling point. Alimama, the group’s marketing platform, put its Wanxiang creative tools at more than 190 million assets produced for merchants, in Double 11 figures carried by PingWest last November. That is a volume boast rather than an audit, which is exactly what makes it useful: it states what the company believes it has industrialised. The same set of figures claims over 3.5 billion yuan in merchant production costs saved. Savings land on the merchant’s side of the ledger, and the images land in the catalogue.

Taobao’s answer to that output arrived on 27 March 2025, and it was not a ban. The platform launched a governance programme against AI 假图, “fake images”, defined by degree rather than by method: material or style that does not match the goods, exaggerated effects, obvious cut-and-paste. Polish must stay “moderate and realistic”, in the platform’s own phrasing reported by the South China Morning Post. Close to 100,000 images had already been intercepted by the time the rules were published. A prohibition on AI imagery would have been one classifier’s work and then finished. Judging degree never finishes, because someone has to sit down every day and decide where the line falls.

A ban is a rule you write once; a boundary is a payroll.

Thirteen months later the same company shipped a second detector, aimed the other way. Taobao and Tmall introduced a model in April 2026 that identifies AI-forged evidence inside refund claims, the photographs of holes and stains buyers submit to get their money back, separating pure generation from edited real photographs and from watermarked fakes. Access went first to merchants rated 4.8 and above, which is an admission in itself: detection capacity is being rationed rather than distributed. Two models in thirteen months, pointed at opposite sides of the same transaction, is the shape of the expense. Generation is cheap on every surface, so inspection has to be funded on every surface.

Product photography was the rate limiter, and nobody costed it that way. A studio day, a model booking and a retoucher’s hours capped how many images a catalogue could absorb in a quarter, which capped in turn how many a moderation team ever had to look at. Marketplace trust and safety was quietly subsidised for two decades by the expense of making the thing being inspected. Generation lifts the cap on production without touching the obligation to inspect. What read as a saving on one line was a transfer to another, and the receiving line renews annually.

This is an operating function with a published governance report now, not a press release about a cleanup. Douyin’s e-commerce arm intercepted more than four million risk merchants over the preceding year, in its first annual governance report, reported by Tencent News in February. Companies do not staff and publish yearly totals for a one-off; they do it for a department, even on the first outing. A separate line in the same report covers 11,000 accounts handled for impersonation carried out using AI or special effects, an enforcement category that had no reason to exist while synthetic video was expensive.

Western retailers bought the same generation capability on a cost case and have not yet been handed the second invoice. As of a September 2024 disclosure, more than 900,000 Amazon selling partners had taken up its generative listing tools, and Amazon’s own account of how they get used is the part to read twice: sellers accepted the generated content with little or no edits roughly 90% of the time. That describes the human review step the business case assumed would catch errors, declining to happen. The EU’s Article 50 transparency obligations began to apply on 2 August, requiring synthetic image output to be marked in machine-readable form and detectable as artificially generated. The Commission finalised a Code of Practice on that article in June, which tells you the marking problem was hard enough to need one. The compliance line exists; the budget line mostly does not.

The strongest case against all this is that detection is capital expenditure dressed as operating expenditure. Build the classifier once, run it at trivial cost per image, and wait for provenance to make forensics redundant: China’s labelling measures and Article 50 both oblige generators to embed implicit, machine-readable marks, so the detector should decay into a metadata lookup. For that to hold, the mark has to survive the journey from generator to listing page. It does not. Screenshot it, crop it, re-encode it, and the implicit label is gone while the picture arrives intact. China’s rules took effect on 1 September 2025; eight months on, reporters at China Jiangsu Net found AI traces in more than half of a twenty-image sample drawn at random across platforms and categories.

Someone pays for inspection, and the only open question is who. A Guangzhou intermediate court has answered it once already, for a merchant who assured a shopper the product photographs were real: the images were generated and unlabelled, the court called it consumer fraud, and the bill was a refund plus treble damages. Platforms can carry the cost, as Alibaba is doing, and recover it through take rates. Merchants can carry it through disclosure and verification duties almost none of them have staffed. Buyers carry it through returns, which is the most expensive option and the one that applies by default when nobody chooses. If catalogue automation keeps being underwritten on cost per image, the line that grows is inspection, and it is the one nobody put in the deck.