Data Labelers: The Invisible Workers Earning $400 a Month at the Heart of the AI Boom

Beneath the Boom, Someone Has to 'Feed' the AI

Since 2024, the large model race has continued to intensify, with giants like OpenAI, Google, Baidu, and Alibaba rushing to release next-generation AI products. Yet at the foundation of this technological frenzy, a group of people is quietly performing the most fundamental — and most easily overlooked — job in the AI supply chain: data labeling.

Their daily tasks involve drawing bounding boxes on massive volumes of images, tagging text, and ranking the quality of AI-generated responses. Without their labor, large language models cannot complete RLHF (Reinforcement Learning from Human Feedback), and image recognition models cannot tell a cat from a dog. They are AI's "teachers," yet they earn assembly line wages — roughly 3,000 yuan (about $400) per month. In some data labeling centers in China's third- and fourth-tier cities, this is considered "normal."

Cyber 'Screw-Tightening': Repetitive, Tedious, and Highly Replaceable

The daily work of data labelers bears a striking resemblance to that of screw-tightening workers on a factory assembly line.

One worker at a labeling center in Guizhou province described a typical day: clocking in at 9 a.m., opening the labeling platform, and using a mouse to draw boxes around pedestrians, vehicles, and traffic lights in image after image — each requiring over a dozen annotations — completing hundreds of images a day. For voice labeling tasks, the same audio clip must be replayed over and over, with transcription text proofread word by word. The work is extremely repetitive; a momentary lapse in concentration can lead to errors, and an excessive error rate means pay deductions or even losing the project.

"At first, I felt like I was participating in cutting-edge technology. Later, I realized I was just doing manual labor for the digital age," one labeler summarized.

What's even more anxiety-inducing is that the job offers virtually no growth curve. Unlike programmers or algorithm engineers, labelers face extremely low skill barriers, with training periods typically lasting only one to two weeks. When projects arrive, they work overtime; when projects dry up, they face unemployment. There is no clear career advancement path, no accumulation of technical expertise, and barely even a sense of belonging to an industry.

'AI Consumables' at the Bottom of the Supply Chain

From an industry perspective, data labeling sits at the very bottom of the AI value chain. Upstream are chip manufacturers and cloud computing platforms; midstream are model development companies; downstream are applications across various industries. The labeling sector itself is highly fragmented, with numerous small and mid-sized labeling companies winning projects through low-price bidding, then subcontracting tasks to labeling factories or freelancers on crowdsourcing platforms.

This multi-layered subcontracting model further squeezes the income of frontline labelers. According to industry insiders, by the time a labeling project passes from the client to the final executor, the price may be compressed to one-third of the original quote or even less. Labelers are paid per piece, with compensation for a single data entry often ranging from mere fractions of a cent to a few cents. To earn 3,000 to 4,000 yuan per month, they need to work at high intensity for more than 10 hours a day.

Meanwhile, AI technology itself is cannibalizing the industry. As large model capabilities improve, the "auto-labeling plus human review" model is gradually replacing purely manual labeling. Some simple image classification and text classification tasks can already be completed automatically by AI, requiring only minimal human quality checks. This means data labelers face the ironic prospect of being replaced by the very AI they helped "feed."

The Deep Contradictions Behind the Predicament

The plight of data labelers reflects several deep contradictions behind the AI industry's prosperity.

First, a severe disconnect between technological value and labor value. High-quality labeled data is one of the key elements of large model performance — OpenAI invested heavily in manual labeling resources when training GPT-4. Yet in market pricing mechanisms, labeling work is treated as "low-skill labor," and its value contribution to the final product goes largely unacknowledged.

Second, a new form of labor exploitation in the digital economy. The rise of the data labeling industry has been embraced by some local governments as a new pathway for industrial transfer and job creation, with provinces like Guizhou, Henan, and Shanxi establishing data labeling centers. But if these positions remain permanently mired in low pay, low skill, and high replaceability, then so-called "digital employment" is merely a repackaged form of cheap labor export.

Third, the 'invisible hand' of the AI industry. In the public imagination, AI is created by genius scientists and top-tier engineers. In reality, behind every intelligent system lies the invisible labor of thousands of labelers. This labor is systematically erased from the narrative, and labelers become "people who don't exist" in the story of technological progress.

Where Is the Way Out?

In the face of industry challenges, some changes are emerging.

Some leading companies are beginning to prioritize labeling quality and are willing to pay higher fees for high-quality labeling in specialized domains. Vertical fields such as medical imaging labeling, legal text labeling, and autonomous driving scenario labeling demand greater professional knowledge from labelers, with correspondingly higher pay — some specialized labelers can earn 8,000 to 12,000 yuan (approximately $1,100 to $1,650) per month.

At the same time, a trend of "labelers transitioning to AI trainers" is emerging within the industry. Some companies are training labelers to learn prompt engineering, model evaluation, and other skills, upgrading them from simple data processors to AI system "coaches." While this pathway has yet to become mainstream, it at least offers the possibility of upward career mobility.

From a policy perspective, the data labeling industry urgently needs standardized labor protection mechanisms and industry standards. Issues such as working hours, minimum piece-rate pay, and social insurance coverage for labelers need to attract regulatory attention.

Conclusion

The brighter AI shines, the more we must remember those who "feed" it frame by frame, word by word, behind the screen. Data labelers are the true "infrastructure" of this technological revolution — but infrastructure should not be treated as disposable. How to distribute the dividends of technological progress more equitably among all participants is an ethical question the entire AI industry must confront. When we marvel at the intelligence of large models, perhaps we should also ask: How are the people who trained the AI doing?