DiPLab is pleased to announce the release of our first policy memo in a new series examining critical issues at the intersection of AI development and labor practices. The complete memo, authored by Antonio A. Casilli (DiPLab, Institut Polytechnique de Paris, France), Thomas Le Bonniec (DiPLab, Institut Polytechnique de Paris, France), and Julian Posada (DiPLab, Yale University, USA), can be downloaded here.
Can We Believe DeepSeek’s Impressive Metrics?
China’s latest AI sensation DeepSeek has captured global attention with bold claims: matching ChatGPT’s capabilities at just 1% of the cost, while consuming only a fraction of the energy. Marketing itself as an open-source alternative to U.S. tech giants, DeepSeek presents an appealing narrative of technological efficiency and accessibility. However, our research at DiPLab reveals a strikingly different reality lurking beneath these impressive metrics.
The Hidden Workforce Behind the Algorithm
While public discourse has centered on DeepSeek’s technical achievements, our investigation uncovers a crucial factor in its success: an extensive network of government-subsidized data labor. Chinese policy initiatives have strategically established data-annotation hubs throughout the country’s “tier 3” cities, offering substantial tax breaks and financial incentives to companies willing to maintain large workforces of data labelers.
The company’s public narrative presents these workers as expert researchers, even suggesting that the CEO personally participates in data labeling activities. DeepSeek claims to operate with a mere 32 annotators—a number that raises significant questions when compared to the scale of data processing required for large language model development.
Questioning the Marketing Narrative
Our research indicates a stark disconnect between DeepSeek’s public claims and documented evidence. This pattern bears striking similarities to earlier revelations about ChatGPT, where initial promises of pure artificial general intelligence (AGI) were later tempered by discoveries of extensive human annotation networks supporting the system. The remarkably low costs and high efficiency metrics touted by DeepSeek deserve closer scrutiny. Our analysis suggests these figures may obscure significant human labor contributions—work that remains largely invisible in public discussions about the platform’s capabilities.
Policy Implications and Recommendations
Based on our findings, we propose several key policy recommendations:
- Decent Work for Data Annotators: DiPLab advocates for the fair treatment of data workers is both ethical and practical—better conditions yield better annotation quality and AI systems.
- Data Protection and Privacy: DiPLab maintains privacy authorities must evolve to protect both AI users and workers.
- AI Act Implementation: DiPLab advocates for extended impact assessments beyond high-risk systems to scrutinize working conditions and data handling throughout the AI lifecycle.
Looking Forward
The case of DeepSeek highlights a broader pattern in the AI industry: the tendency to emphasize technological capabilities while minimizing or obscuring the human labor that makes these systems possible. As we continue to evaluate and regulate AI technologies, it’s crucial to consider not just their technical specifications, but also their social and labor implications.
The true cost of AI development cannot be measured solely in computational resources or energy consumption. We must also account for the human labor that powers these systems—labor that often remains hidden behind marketing narratives of pure artificial intelligence.
This post summarizes findings from Casilli, A. A., Le Bonniec, T., Posada, J. 2025. “The Human Cost of DeepSeek. Hype the technology, hide the workers,” DiPLab Policy Memo, vol. 1, n. 1.
DiPLab is a research program at the Institut Polytechnique de Paris (France), focused on studying the social and labor implications of digital technologies.
