Bridging the Human-AI Gap at Whose Cost? Skill and Organizational Alignment in AI Data Annotation Work

• In Event: 20806 - Artificial Intelligence in the Workplace (Co-sponsored by Section on Organizations, Occupations and Work)

Abstract

This study explores the evolving relationship between human labor and AI, with a particular focus on the skill dynamics of AI data annotation work in China. Data annotation, where human workers label and categorize data to train AI systems, is critical for enabling AI to interpret human norms and values. Drawing on 4.5 years of ethnographic research and 192 interviews with stakeholders from leading tech firms such as Alibaba, ByteDance, Tencent, and Baidu, the paper examines the shifting demands for human skills in two distinct stages of AI development: Vision AI and Large Language Models (LLM). The study introduces the concept of "skill alignment," reinterpreting it as a socio-organizational process in which human labor and AI technologies evolve together. Through this lens, the research identifies two trajectories: downward alignment in Vision AI annotation, where automation and platformization reduce skill requirements, and upward alignment in LLM annotation, where the increasing complexity and the experimental nature of technology heighten the need for skilled human input.

This research challenges the prevailing narratives in AI labor studies that characterize annotation work as low-skilled. It reveals instead the multifaceted and dynamic nature of the skills required in this sector. It critiques the binary models of deskilling and reskilling, proposing that skill transformation in the AI industry follows a cyclical pattern driven by technological evolution. By contributing to Human-Computer Interaction (HCI) and labor studies, it highlights how organizational processes and power dynamics shape the evolution of labor skills in AI industries. Furthermore, it underscores that the costs of these skill transformations—whether through deskilling in Vision AI or upskilling in LLMs—are disproportionately borne by human workers. The study calls for reevaluating how human skills complement AI models and are valued in AI-driven industries, challenging assumptions of inevitable job displacement in the AI era.

Author

• Tongyu Wu, Zhejiang University