China’s Professionals Are Moonlighting for Their Replacements

ByPriscilla Li
Chinese professionals working on laptops in a modern office

Chinese AI platforms are paying teachers, engineers, architects, writers and other professionals to turn their real work into model-training examples, then retaining broad rights to reuse that work commercially. Alibaba’s Siriser and ByteDance’s Xpert recruit people whose expertise is under economic pressure, offering piecework that can take hours and pay 100 to 500 yuan per task.

The work is not ordinary click-labeling. A professional is asked to show an AI system how they would build a proposal, assess a company’s profitability, draft legal advice, write a judgment or assign work to a software agent. In other words: they sell the steps between a blank page and competent output.

China’s National Data Administration formally pushed this direction on June 3, 2026, calling for industry experts to participate more deeply in data annotation, including instruction tuning, examples that teach a model how to respond, and reinforcement-learning annotation, where people score outputs. The state wants richer training data; platforms have found a labor market willing to supply it.

Expert Workflows Become Model Training Data

Siriser is Alibaba’s specialist AI-training platform, launched in January 2026, while ByteDance launched Xpert in 2025. Both sit inside companies better known for consumer internet products: Alibaba operates shopping platforms such as Taobao and Tmall, while ByteDance owns TikTok and Douyin. Their newer pitch is workplace intelligence built from the people already doing workplace tasks.

Rest of World reported in September 2026 that Siriser recruits teachers, mechanical engineers and music composers. Xpert says it has recruited more than 50,000 experts, including writers and therapists, but that figure is the platform’s own statement, not an independently audited count.

The core transaction is straightforward. A contributor takes a plausible task from their profession, provides the relevant materials, and explains how an experienced person would approach it. One Shenzhen architect, identified as Cuicui, told Rest of World she logged into TalentsAI after her day job to show a model how to write building proposals and conduct profitability analysis.

“Every task has to come from my actual work.”, Cuicui, a Shenzhen architect, speaking to Rest of World

That distinction matters. A generic data-labeling job might ask someone to identify objects in photographs. Expert annotation asks an architect why a proposal is structured a certain way, a lawyer why a clause creates risk, or an engineer why one technical path beats another. It turns judgment accumulated over years into repeatable examples, how training data turns workers’ know-how into AI infrastructure, except here the raw material is white-collar process rather than physical movement.

A professional task is explained by an expert and converted into an AI training example

Workers interviewed by Rest of World described assignments that could consume several hours, with payments generally ranging from 100 to 500 yuan each. The price is highly variable because the useful asset is not time alone; it is credible domain judgment.

A Shanghai software engineer with two decades of experience, who began taking AI-training work in July, described the calculation bluntly:

“You have to keep evolving yourself to avoid getting completely replaced.”, a Shanghai software engineer, speaking to Rest of World

These accounts came from trainers who requested anonymity because they feared reactions from employers, colleagues or the platforms themselves. That fear is itself part of the record: the work pays badly enough to need a second job and is sensitive enough that naming yourself is a risk.

The labor-market backdrop is hard to miss. Cuicui said her income had halved over five years as government infrastructure spending weakened. China’s youth unemployment rate reached 17.9% in July 2026. For contributors, this is a hedge against unstable income, not a victory lap for the arrival of AI work.

The Platform Terms Put Rejection and Reuse Risk on Workers

The valuable part of these gigs is the contributor’s judgment. The risky part is that the platform decides whether that judgment qualifies for payment.

Siriser’s June 17, 2026 user agreement classifies participants as commercial collaborators rather than employees. That means no stated obligation to provide social insurance or employment benefits. The agreement also gives the platform final authority over whether a submission is accepted, and says it may decline to pay for work it rejects.

ByteDance’s Core Xpert Data Platform terms, dated July 16, 2026, set out a similar arrangement: the operator has sole authority to audit work, choose a payment coefficient and deny rewards when a task fails review.

The terms do not show how often platforms reject work or withhold payment in practice. They do show who holds the contractual lever when there is a dispute: not the person who supplied the expertise.

A platform reviews a worker submission, controls payment, and retains commercial reuse rights

Siriser’s terms go further than a normal short-term commission. They authorize Alibaba and its partners to adapt, publish and use submitted deliverables for model training, and to commercially sell those deliverables without additional consent or payment.

That is the asymmetry at the center of the arrangement. The worker may spend an evening explaining how to perform a specialized task and receive a one-off fee, if the platform accepts it. The platform can retain rights to package that explanation into a training-data product or a model capability that can be reused indefinitely.

Mirror Studio, republished by Jiemian in December 2025, reported that one Xpert participant doing basic image and text labeling earned 800 to 1,200 yuan per month after rejected work, which the report described as roughly 80 yuan an hour. That was one basic-labeling participant, not a measure of specialist-task earnings. The same report described specialist assignments in finance, law and medicine priced at 300 to 1,000 yuan.

Even access to tasks can be unstable. Mirror Studio reported that Xpert contributors could see several tasks listed, open the assignment page, and find the work had already disappeared because others got there first. Task competition and rejection are not incidental flaws in a marketplace model; they move idle time and quality-control risk onto workers.

China’s Expanding Market for Productivity-Focused Training Data

The National Data Administration’s June plan explicitly directs industries to build “high-quality data sets” and calls for deeper expert participation in annotation. Its emphasis on instruction tuning and reinforcement-learning annotation is a policy recognition that more raw text is not enough; models also need people who can distinguish a plausible answer from a professionally sound one. China’s rapid AI adoption drive is now reaching the less glamorous layer underneath the chatbots: the people making outputs useful at work.

IDC forecasts China’s AI-training-data market will reach 7.8 billion yuan in 2026, up 25% from 2025. The market is moving beyond consumer-content moderation and entertainment datasets toward data intended to make models draft documents, reason through business tasks and operate software.

What workers provide What platforms can build
Realistic professional assignments Benchmarks and training prompts
Step-by-step explanations Instruction-following examples
Corrections and judgments Reward data for model evaluation
Documents and task context Commercial training-data products

The companies pursuing this data have financial reasons to find useful AI products. Alibaba reported that its AI Labs and Applications unit made an adjusted EBITA loss of 13.861 billion yuan in the quarter ended June 30, 2026, compared with a 3.224 billion yuan loss a year earlier. Its AI Cloud and Compute Services revenue grew 45% to 48.437 billion yuan in the same quarter.

Alibaba attributes the wider AI unit loss to increased AI investment and higher Qwen-app inference costs; its filing does not tie that loss to Siriser. But the direction of travel is visible: better workplace data is an input into AI products that companies hope customers will pay for.

ByteDance faces a similar pressure to turn AI spending into durable capability. Caixin reported in April that ByteDance’s 2025 net profit fell by more than 70% amid aggressive AI investment, even as domestic revenue grew nearly 20% and overseas revenue nearly 50%.

The deal is concrete and it is one-way. Professionals whose incomes are already insecure are paid piece rates to hand over the working methods that make a professional AI system credible, and the platform keeps the power to reject the work, set the rate and reuse the result for as long as it likes. A method, once transferred, cannot be taken back. The teacher who has shown a model how to mark a paper has been paid once for something the platform can sell indefinitely, and nothing in the arrangement gives them a claim on what it earns.

That makes expert annotation less like consulting and more like piecework with unusually valuable raw material: the contributor’s own career.

This site’s reading of where that goes: nobody in this arrangement has a reason to stop it. The platform pays once for a method it can resell forever, the state has asked for exactly this kind of expert data, and the supply comes from people whose incomes are insecure enough to accept 100 yuan for it — a group that grows as the systems get better. Each task narrows the gap between what the professional can do and what the model can do, and it is the professional who is paid to narrow it. Nothing about that requires anyone to be replaced for the trade to be a bad one. It requires only that the work keep being worth more to the buyer than the price on the task, which is the one thing every party already agrees on.

Key Takeaways

  • Alibaba’s Siriser and ByteDance’s Xpert recruit professionals to create AI-training examples based on real workplace tasks.
  • China’s National Data Administration called for deeper expert involvement in high-quality AI data annotation on June 3, 2026.
  • Expert trainers reported payments of 100 to 500 yuan for tasks that can take hours to complete.
  • Siriser’s terms allow rejected work to go unpaid and grant broad commercial reuse rights over accepted deliverables.
  • China’s AI-training-data market is forecast to reach 7.8 billion yuan in 2026.

Further Reading