A Shenzhen architect's second shift teaches AI to write her building proposals

The pay rule is where the pitch and the work part company. Xpert's user terms say a trainer whose submission fails review does not get paid ^[12]. Siriser told Rest of World that rejected submissions can be revised or appealed ^[13]. Trainers gave the practical version: quality control teams reject outputs, and hours of work earn nothing ^[11].

A task runs a few hours and pays 100 to 500 yuan ^[10]. Take three hours as "a few" and the floor is about 33 yuan an hour, the ceiling about 167 ^[1], a fivefold band ^[2] on work whose payment is settled after delivery. Trainers told Rest of World the work is precarious, demanding and lower-paid than in America ^[7], where Mercor, Surge AI and Handshake have already signed tens of thousands of filmmakers, lawyers and scientists for the same kind of task writing ^[8]. That is a different account from the one MeetChances posted on RedNote in June, which asked what middle-aged people can do after being laid off and suggested they try becoming AI trainers ^[19]. Youth unemployment was 17.9% in July, near a three-year high ^[23].

The task design requirement should interest anyone buying the output. Trainers have to invent assignments the model cannot already solve, and that has got harder as models improve ^[14]. They are also barred from using AI while doing it ^[15]. The billable supply is therefore defined by the model's current weak spots, and it thins as the model gets stronger.

Then there is the upload. The instruction is to bring real documents from the day job and write down your own thought process over them ^[9]. "Every task has to come from my actual work," Cuicui told Rest of World ^[4]. Every trainer the publication interviewed asked for anonymity, partly to avoid being identified by their employer ^[5]. If you run an architecture practice or a tax team, your document controls are now competing with a few hundred yuan a task, and nobody inside the firm signs the release.

Tiezhen Wang, a Sydney-based researcher who previously led Hugging Face's Asia-Pacific ecosystem work, told Rest of World that Chinese labs have closed much of the gap on general reasoning and coding and are scaling specialized data collection because they expect work tools to be the next contested ground ^[20]. What has shipped so far, Tencent's WorkBuddy and ByteDance's Doubao Work, generates slide decks and builds websites; medicine, law and finance need more data than that ^[22].

Two axes sort the data behind those products. First, did the task come from an artefact someone was already paid to produce at their job, or was it invented for the platform. Second, was the expert paid for the attempt, or only for the accepted result. Invented tasks plus contingent pay produce submissions tuned to satisfy a reviewer, which is not the same as work getting done. The terms on the public record describe the contingent column ^[12]. A vendor that cannot tell you what share of submissions its reviewers threw out is asking you to assume its reviewers and your job agreed on what good looks like.