On a humid weekday afternoon in May, Chao Teng-Hung, a 20-year-old computer science sophomore at National Taiwan University of Science and Technology, sat in a bustling cafe near his campus, his face illuminated by an AI prompt window on his laptop screen.

To any casual onlooker, he looked like a typical student browsing the web. In reality, he was commanding an income that peaked at $10,000 a month—roughly 320,000 New Taiwan dollars—all while working fewer than 40 hours a week on his own terms.

Mr. Chao belongs to a specialized, highly educated cohort of data annotators known as language model "trainers." His work centers on a sophisticated methodology called Reinforcement Learning from Human Feedback, or RLHF. Put simply, his job is to review responses generated by two competing AI models and determine which output reads as more accurate, nuanced, and fundamentally human.

Because of his background, Mr. Chao is primarily assigned tasks in Traditional Chinese. The corrections and feedback he provides are packaged as proprietary data assets by data-labeling platforms and sold directly to Silicon Valley giants, including Google, Meta, and OpenAI. Over the past 21 months, he has pulled in $78,000 from just one platform, DataAnnotation—and it is only one of four platforms he routinely cycles through.

The Shift to High-Skilled Cognitive Labor

The explosion of interest in RLHF stems from a structural shift in how AI is built. While tech firms pitch AI as fully automated, the systems are remarkably dependent on human labor. In the industry's infancy, outsourcing companies hired workers in Southeast Asia and Africa for pennies to handle low-stakes classification tasks, like identifying stop signs for self-driving cars.

But the launch of ChatGPT in late 2022 upended that model.

"Today’s AI models require highly precise, real-time human data to anchor them," said Chang Chien-Wen, vice president of marketing at E-Land Information, a Taiwanese big data analytics firm. "Without this high-level verification, large language models cannot overcome the hurdle of 'hallucinations'."

This hunger for high-fidelity data has turned Silicon Valley labeling startups into Wall Street darlings. Scale AI, the industry's undisputed juggernaut, saw its valuation soar to $29 billion last year following a funding round backed by Meta. Its founder, Alexandr Wang, is just 29. Similar meteoric rises have followed firms like Surge AI, founded by tech veteran Edwin Chen, and Mercor, a platform specializing in matchmaking tech firms with "white-collar annotators."

Alexandr Wang, 29, is the founder of Scale AI, the global data-labeling juggernaut. (Source: AP)

Industry experts estimate that the global RLHF workforce now numbers anywhere from hundreds of thousands to millions. John Winsor, an open-talent specialist, notes that global demand for individuals capable of training AI systems is surging at an annual rate of 25 to 30 percent.

While platforms guard their regional enrollment numbers as proprietary secrets, the trend is vividly clear in Taiwan. Data from 104 Job Bank, a major local employment platform, reveals that monthly postings for AI-related part-time gigs averaged 750 in 2025—a 16 percent year-over-year increase that outpaced the broader part-time job market. On local student forums like Dcard and Threads, posts dissecting RLHF strategies are ubiquitous.

Why Silicon Valley is Hunting in Taiwan

Julian Posada, an assistant professor at Yale University who tracks the global digital gig economy, told CommonWealth Magazine that the demand for data annotation has moved heavily into new domains. "The shift has focused mainly towards language, be it textual language or spoken language, or even coding language and math," he said.

This has forced platforms to target countries with specific skills and higher education levels. Instead of traditional global outsourcing, "they are hiring people more in-site," increasing local recruitment in regions like Japan, South Korea, and the Chinese-speaking world.

Taiwan presents a uniquely attractive arbitrage opportunity. The island offers an elite tier of higher education and a workforce fluent in Traditional Chinese, paired with local professional salaries that have remained stagnant for decades.

Tseng Yen-Hsuan, 37, is an example of the overqualified talent entering this ecosystem. Holding a master’s degree in economics from National Taiwan University, Mr. Tseng has a robust portfolio of published statistical research. Last March, he was headhunted via LinkedIn by Outlier, a prominent labeling platform owned by Scale AI.

Boasting elite higher education and fluency in Traditional Chinese, Taiwanese professionals have become prime targets for data-labeling platforms. Chao Teng-Hung, a student annotator, converses with AI to send refined data back to tech giants. (Photo: Chien-Ying Chiu)

Working from a dimly lit apartment room, his long hair pulled back, Mr. Tseng spends his days playing schoolmaster to unreleased neural networks. His task is to formulate complex statistical problems, evaluate how multiple models attempt to solve them, and pinpoint logical fallacies. Occasionally, he infuses his love for German and poetry into the prompts. "It feels a bit like raising children," Mr. Tseng said with a wry smile. "You can actually witness them getting smarter over time."

The work is strictly piecework, with rates hovering between $25 and $45 an hour. At his peak, Mr. Tseng cleared more than $23,000 in a single month. The financial windfall and absolute autonomy prompted him to quit his traditional job as an AI consultant to do RLHF full-time.

For younger annotators like Mr. Chao, the work also doubles as an accidental form of cultural preservation. Beyond testing for accuracy, his prompts enforce safety guardrails—forcing the AI to reject instructions on how to commit crimes or generate hate speech.

He also rigorously grades the AI on its regional vernacular. "I constantly have to catch mainland Chinese phrasing and force the model to adopt Taiwanese usage—changing terms like shipin [video] to yingpian," Mr. Chao said. In a way, these workers are the ones ensuring local culture isn't erased by the models.

The 'Ghost Work' Precarity

Yet, beneath the allure of outsized paychecks lies a profound sense of professional dislocation.

"Most of the tasks are mind-numbing," Mr. Chao admitted. "You aren't acquiring any new skills. At the end of the day, it's just a dead-end gig."

For Mr. Tseng, the existential dread is tied to the technology itself. "The more meticulously I nurture these models, the faster they outgrow me," he said. "Eventually, they won't need me to train them anymore. What happens to my value then?" To combat this growing sense of futility, he is considering transitioning back to a traditional full-time job.

While Professor Posada does not believe human experts will be easily automated away—noting that "not even AI today has been able to completely replace humans"—he stated frankly that it is "very tough, very complicated to call it a long-term career." Platforms lack basic training and health insurance, and abrupt account deactivations are the norm.

More realistically, the lucrative pay may be a temporary trap. "Once they achieve a critical mass of workers, then they remove those incentives to make a profit," Professor Posada warned.

Mary Gray, a senior principal researcher at Microsoft Research who coined the term "ghost work," offers a biting assessment of the situation.

"Taiwan isn’t just benefiting from AI demand—it’s being folded into its hidden labor infrastructure," Dr. Gray said. "Short-term pay can be attractive, but it can mask long-term tradeoffs. We’re seeing globally valuable expertise routed through platforms that treat it as disposable."

She warned of a broader societal impact: "The risk is normalizing a future where even expert work comes without stability."

When pressed on these criticisms regarding precarious employment, Scale AI defended its corporate model. In a statement to CommonWealth Magazine, the company noted: "Outlier was designed to provide flexible, supplemental earning opportunities for people looking to apply their expertise on a part-time basis. Most contributors use the platform alongside full-time employment, studies, caregiving responsibilities, or other professional pursuits."

This corporate defense starkly illustrates the reality of the gig economy: platforms offer no long-term safety nets. So why do Taiwanese professionals continue to flock to it?

Lin Thung-Hong, a sociologist at Academia Sinica, notes that the phenomenon is driven less by domestic low wages and more by the aggressive pull of American capital. "This is the Silicon Valley model expanding globally at an unprecedented speed: high compensation with short-term results," Dr. Lin said.

While Dr. Lin acknowledges that allowing Taiwanese youth to plug directly into the global AI supply chain is a form of success, he urges caution. "The subtext of the Silicon Valley model is an endless cycle of churning through people," he warned. Annotators must recognize the reality of the bubble and prepare to return to the traditional job market when it eventually bursts.

Back in the Taipei cafe, Mr. Chao is already preparing for that exact scenario. He spends his evenings studying Japanese, hoping to secure a traditional job in Tokyo after graduation.

"As a university student, you eventually have to plan for a real career," he said quietly, before closing his laptop.

Have you read?

• Why France Chose Foxconn to Build Europe’s Most Advanced Packaging Plant
• How Machine Manufacturer Ta Liang Became a Hidden AI Champion
• Inside the AI Rack War: MediaTek and Foxconn Race Toward the Copper-to-Optics Era

Uploaded by Ian Huang