Chinese Tech Workers Are Training AI Doubles — and Pushing Back

Inside Chinese technology companies, a quiet but consequential shift is underway. Workers in engineering, product, and operations roles are being asked to document their workflows, annotate their decisions, and contribute the behavioral data needed to train AI systems designed to replicate what they do. The framing from management is typically collaborative — helping the company scale, improving efficiency, building internal tools. The workers involved are often more clear-eyed about the trajectory.

The practice represents a specific and increasingly visible phase in enterprise AI deployment: the extraction of tacit professional knowledge from human workers to construct agent-like systems that can eventually operate in their place. What makes the Chinese context notable is both the pace at which this is happening and the fact that workers are beginning to organize informal resistance — declining to participate fully, submitting low-quality training data, or raising objections through internal channels.

This dynamic is not unique to China, but the scale and directness with which major Chinese technology firms are pursuing it offers a preview of tensions that will emerge more broadly as AI deployment moves from productivity tooling into workforce substitution.

The mechanics vary by company and role, but the general pattern is consistent. Workers are asked to narrate their processes, record how they handle edge cases, explain their reasoning on ambiguous decisions, and sometimes interact directly with prototype AI systems to evaluate and correct outputs. This data — structured and unstructured — feeds into fine-tuning pipelines or is used to build internal agent systems scoped to specific job functions. In some firms, participation is framed as voluntary but tracked for performance evaluation purposes, creating implicit pressure to cooperate.

The resistance that has emerged is notable for its practical rather than ideological character. Workers are not broadly opposing AI development. They are specifically concerned about being asked to perform unpaid knowledge labor that accelerates their own redundancy, with no protections, disclosures, or compensation attached to the contribution. Some describe deliberately incomplete participation — providing surface-level documentation while withholding the judgment-based knowledge that would make the resulting AI system genuinely capable. Others express uncertainty about whether their cooperation is even legally required, given the absence of clear Chinese labor regulations governing AI training data extraction from employees.

For companies operating in or adjacent to China's technology sector, this development carries several immediate implications. First, the quality of internally trained AI systems is directly contingent on worker cooperation, meaning coercive or opaque approaches to data collection risk producing systems that are structurally incomplete. Second, the growing awareness among workers of their role as training data sources introduces a new variable in workforce management, particularly as the practice becomes more visible through informal networks and press coverage. Third, the absence of regulatory clarity in China — and most other major markets — means firms are operating in a policy vacuum that will not remain stable.

From a longer-term perspective, what is happening in Chinese tech firms is the first widespread instance of a predictable inflection point: the moment when AI deployment shifts from augmenting workers to formally ingesting them as inputs. The human knowledge embedded in experienced workers is, for many enterprise use cases, the primary remaining bottleneck for capable AI systems. Accessing that knowledge requires either extracting it from the workers who hold it or accepting the limitations of systems trained on public or synthetic data. Companies that resolve this tension well — through genuine reciprocity, clear disclosure, and appropriate compensation structures — are likely to build more capable systems and retain the workforce stability needed to maintain them.

The firms that treat it as a data acquisition problem with humans as a reluctant resource will likely encounter the same degraded cooperation now visible in parts of China's tech sector.

Sources: — MIT Technology Review (https://www.technologyreview.com/2026/04/20/1136149/chinese-tech-workers-ai-colleagues/)