The Low-Wage Data Labor Hidden Behind AI Model Training

For core AI model tasks like image classification, text labeling, and evaluating algorithm responses, 430 million data processing workers are being deployed worldwide. According to World Bank data, this figure has surged sharply from a previous 150 million.

Key characteristics are as follows:

• Low-wage outsourcing: Tech companies source labor through developing countries (India, Bulgaria, Kenya, Venezuela, etc.) where labor costs are low and social protection systems are inadequate.
• Concealment of labor: In order to maintain a narrative of technological autonomy, the massive scale of human labor actually invested in model training tends to be deliberately hidden from the public.
• Poor working conditions: On certain platforms (e.g., Mindrift), users are paid about $9 a day for completing 12 tasks, while inmates in Finnish prisons receive a daily wage of only 3 to 4.62 euros.

Economists point out that the current existence of AI models would be impossible without the human labor that fills data repositories.