Low-Wage Labor Exploitation in the Development of Generative AI Generative AI tools are not built simply by training models through large-scale data computation. In many cases, data labeling is required to organize raw data into forms that can be meaningfully used for training. Moreover, because trained models inevitably reflect biases, errors, or harmful content present in their training data, additional fine-tuning is necessary before deploying them as real-world services in order to minimize inappropriate or harmful outputs. Fine-tuning itself is also a form of data labeling and typically takes the shape of large-scale microwork, involving the labor of many people. Data labeling labor is characterized by the instability inherent in microwork, the psychological burden of repeatedly encountering harmful or hateful content, and the frequent outsourcing of tasks to low-wage regions in the Global South. These labor processes are often obscured by complex subcontracting chains and corporate secrecy, making it difficult to accurately assess their scale and conditions. In this sense, the production of generative AI tools rests on multiple layers of labor exploitation, raising serious ethical concerns about how generative AI is developed and used. Automation, Job Displacement, and Productivity Generative AI is often perceived as a technology that enhances productivity by automating and restructuring work. However, this also carries the risk of job displacement and the reduction Generative AI Guide for Civil Society