By Kris Li
The views expressed below are those of the author and do not necessarily reflect those of the Carr Center for Human Rights Policy or Harvard Kennedy School. These perspectives have been presented to encourage debate on important public policy challenges.
While the Chinese government aims to win the AI war, the global community must recognize and reject a harmful, worker-resistant approach to AI advancement.
Right before China’s 2025 International Workers’ Day holiday, the artificial intelligence race among Chinese tech giants reached a new peak: Alibaba released Qwen 3, a large language model (LLM) family claimed to outperform OpenAI and Google’s best offerings; Xiaomi launched its first LLM, Mimo, now competing with OpenAI’s o1-mini; and DeepSeek unveiled their new model, Prover V2.
Just as DeepSeek shocked the world during the 2025 Spring Festival, releasing new products before International Workers’ Day is a strategy, providing users more time to experiment with these models while preempting rivals. In China, May Day is no longer perceived as a day for worker solidarity, but rather as a holiday not shared by the most vulnerable communities, including approximately 300 million rural migrant workers. When users first encounter DeepSeek’s censorship by testing it with questions about the Tiananmen Massacre or Xi, they often overlook that this AI also ignores ongoing social movements in China, including worker protests. Examining how the rise of Chinese-developed models in the AI landscape impacts workers’ rights has become increasingly urgent.
During the first 4 months of 2025, approximately 540 worker protests in China were documented by China Labor Bulletin, the leading NGO in Chinese labor rights advocacy. However, when asked to provide a list of labor protests in China in 2025, DeepSeek’s response began with “China is a country governed by the rule of law, where the rights and interests of workers are highly protected by the government.”, asserting that China continues to maintain social harmony under the leadership of the Chinese Communist Party and denying the existence of widespread protests. When presented with information on real cases, such as a small protest that occurred on April 28 in Liaoning Province, DeepSeek again responded with similar sentences, only emphasizing harmonious labor relations and worker protection under Party leadership. When the author sent these same questions in Chinese Mandarin, the AI cautioned against the risk of spreading “false information”.
Facing the same questions regarding labor protests in China, Alibaba’s newly released Qwen 3 model explained that “disturbing the social order was prohibited.” When testing the Qwen 3-235B-A22B demo in English, the model could still cite organizations typically censored by the Chinese government, such as Human Rights Watch, to address questions about labor conditions and worker rights in China. While it identified worker rights violations, including the absence of independent trade unions, suppression of collective bargaining and strikes, inadequate enforcement of labor standards, and exploitation of migrant workers, Qwen 3 consistently stopped itself before generating content about forced labor in Xinjiang, and its responses in Chinese were notably more conservative. Although Qwen 3’s detection mechanisms currently censor only critical topics like Xinjiang and Tibet, Chinese developers typically require merely weeks to “adjust” newly released models regarding prohibited content, as the political risks are exceedingly high.
While actively covering worker rights violations and the suppression of independent trade unions and worker-led organizations in China, models like DeepSeek are also enhancing the Chinese government’s capacity to monitor and disrupt labor protests and other civil movements through mass surveillance. In February 2025, GoLaxy, the state-owned enterprise leading China’s big data industry, launched its Comprehensive Social Listening System in collaboration with DeepSeek. With local authorities integrating DeepSeek into social monitoring systems, China’s already stringent online censorship could become even more precise for labor-related content and dissidents, affecting both public platforms like TikTok and Kuaishou favored by the working-class people, and semi-private platforms like WeChat, which is essential for daily life.
With DeepSeek now adapted for police operations in regions such as Uxin Banner, Inner Mongolia, AI deployment by police departments for potential persecution is also concerning, since local police in China are frequently implicated in political violence against labor activists, dissidents, and ordinary citizens. This includes cases like that of Wuyi (乌衣), who was targeted only for attempting to visit the victim of the Xuzhou chained women incident, being perceived as a threat to local authorities who protected women trafficking.
Furthermore, while “petitioning” (上访) represents a Chinese administrative system designed to address grievances from individuals mistreated by local authorities, petitioners often face imprisonment and abuse. With DeepSeek now being integrated into petition case analysis in places like the Airport Economy Zone of Zhengzhou City, AI-enhanced surveillance may intensify the persecution of petitioners, particularly workers, considering that many petition cases come from labor disputes.
With projected investments in the Chinese AI industry expected to reach $100 billion by 2028, the working conditions being reshaped also merit attention. Since 2020, worker exploitation within Chinese food delivery platforms such as Meituan, which allocated approximately $650 million to technology development alone in the third quarter of 2021, has drawn public attention. Under rigid algorithmic systems and punishment mechanisms, delivery workers operate in hazardous conditions to satisfy routing algorithm requirements, with average delivery times compressed to 28 minutes by 2017. As rural migrant workers without formal contracts or proper insurance, their vulnerable circumstances reflect both algorithmic oppression and the lack of bargaining power for workers in China. CHEN Guojiang, a leading advocate of these workers, was arrested in 2019 for organizing strike actions and again in 2021 for exposing scandals of the Ele.me delivery platform.
Additionally, data labeling—an industry similarly dependent on unskilled labor and forming the foundation of China’s AI sector—emerged between approximately 2017 and 2019, reaching its peak in 2025. Characterized by extended working hours and opaque wage structures, this field manifests social disparities, including urban-rural dynamics, social class divisions, and gender inequalities that warrant further examination.
In 2025, China continues to be classified as a country with no guarantee of rights by the International Trade Union Confederation. While the Chinese government aims to win the AI war, the global community must recognize and reject a harmful, worker-resistant approach to AI advancement.
Kris Li, Student Ambassador, Carr-Ryan Center for Human Rights
Matheus Bertelli - Pexels
