Expert AI Training Gigs Spread Across China’s White-Collar Workforce

A Shenzhen architect’s second shift now starts after the office day ends: Rest of World found that Chinese professionals are turning their own white-collar routines into training material for AI models, often for low-paid gig work that helps cover household bills.

The work turns specialist judgment into task instructions.

Cuicui, a Shenzhen-based architect in her 40s, had designed public stadiums, schools and hospitals, but declining government infrastructure spending over the past five years cut sharply into her employer’s business and halved her income.

TalentsAI tasks gave her a way to convert building proposals and profitability analyses into model-training prompts after her regular workday.

That individual workaround is becoming a wider labor channel.

China’s National Data Administration backed the building of “high quality data sets” in June and called for more industry experts to be pulled deeper into annotation work to raise the “knowledge density” of AI training data.

The same push also framed annotation as flexible employment for college graduates, linking model development to a weak job market rather than only to research spending.

Platforms are recruiting beyond the old image-labeling workforce.

Alibaba’s Siriser, launched in January, is seeking teachers, mechanical engineers and music composers.

ByteDance’s Xpert, launched in 2025, lists more than 50,000 recruited experts, from writers to therapists.

TalentsAI and MeetChances are adding journalists, tax accountants and headhunters, turning professional workflows into data products for office, legal, finance and engineering tools.

The assignments ask trainers to create realistic work problems, upload documents from their actual jobs and write out the reasoning a model should learn.

Trainers are told not to use AI themselves, and a task usually takes a few hours for pay of 100 to 500 yuan, or about $15 to $74.

That makes the role materially different from the higher-paid expert-training market in the U.S., where companies including Mercor, Surge AI and Handshake have recruited filmmakers, lawyers and scientists.

The risk sits inside the workflow.

Trainers say tasks must be hard enough that models cannot solve them on their own, a bar that rises as AI improves.

Submissions can also be rejected by quality-control teams, leaving workers unpaid for hours.

Xpert’s terms state that trainers will not be paid if submissions fail requirements or reviews, while Siriser allows revisions or appeals after rejections.

China’s shift toward expert data comes as AI labs narrow the gap with Silicon Valley in general reasoning and coding.

Office products are moving in the same direction: Tencent has WorkBuddy, ByteDance has Doubao Work, and both target tasks such as slide generation or website building.

Tiezhen Wang, a Sydney-based AI researcher and former head of the Asia-Pacific ecosystem at Hugging Face, connected the next competition to better performance in real work settings.

Medicine, law and finance require more specialized data than general consumer applications.

The labor pressure is visible outside the platforms.

Youth unemployment reached 17.9% in July as a record cohort of graduates entered the market, and IDC expects China’s AI training data market to reach 7.8 billion yuan, or $1.1 billion, in 2026, up 25% from 2025.

Data annotation companies once leaned on student interns, rural residents and stay-at-home mothers for self-driving and voice-recognition work; productivity software is now drawing in middle-class professionals with degrees and industry experience.

Researchers and workers both describe the appeal as defensive.

Julie Yujie Chen, an associate professor at the University of Toronto who studies data labor in China, tied middle-class interest in the gigs to a tough labor market and government endorsement that gave the occupation more legitimacy.

MeetChances also pitched the work on RedNote in June as an option for middle-aged people after layoffs, showing how platforms are marketing AI training as a supplement to unstable careers.

For workers such as Liu, a Shenzhen legal AI employee with a master’s degree in law, the choice is defensive as much as opportunistic.

After work, she tests models at Moonshot AI, Tencent and TalentsAI by asking them to draft legal advice or judgments, then evaluates the answers.

Jessica, a Shanghai software engineer with two decades of experience, took gigs in July after AI disruption slowed her company and contractors replaced full-time engineers.

Her reason for trying the work was direct: she wanted “one more option.”