Data-labeling work—categorizing, testing and moderating text, images, audio and video for AI training—is performed by a global digital workforce serving tech companies and industries including banking, insurance, health care and government agencies. Yet this labor largely occurs under exploitative conditions with little public scrutiny.
Interviews with 10 workers in China and Australia revealed stark inequalities across the sector. Specialized workers with Ph.D.-level qualifications or STEM certifications in the Global North earn up to A$400–A$800 per hour. Most workers, however, perform repetitive general tasks such as drawing bounding boxes for images or annotating audio, earning as little as A$6 per day. The low pay cannot cover daily expenses, and long hours produce chronic eye strain and back pain.
Workers are classified as "users" rather than employees and have no formal contracts. They are not entitled to appeal performance assessments and often do not know which companies' data they label or whether machines or humans evaluate their work. User agreements primarily protect companies and require non-disclosure of information workers encounter.
All interviewed workers reported receiving less work as AI advances, leaving more difficult and time-consuming tasks. Workers expressed resignation about eventual job replacement. "If I don't make this money, someone else will, and I will be replaced eventually anyway," one noted. Companies recruiting increasingly from vulnerable groups—people with disabilities, recent graduates and pregnant workers—who are less likely to leave because few other employment options exist.
Payment practices vary by geography. U.S. platforms pay workers upon task submission. Chinese platforms pay only after work is assessed and confirmed to meet standards, leaving workers unpaid for hours of effort. Workers in China cannot access U.S. platforms without risking account suspension through VPN use.
