Ghost Work: The Hidden Labor Behind AI

“Ghost work” is a term for human labor that remains largely invisible in products marketed as automated. Workers may annotate images, transcribe audio, categorize text, compare generated responses, or evaluate search results. Their work can be assigned through microtask platforms, staffing companies, subcontractors, or in-house operations. Calling the task invisible does not mean the workers are anonymous in every arrangement; it highlights that users often do not see the labor or its conditions. Data labor contributes throughout an AI product lifecycle. Training sets can require people to define labels and resolve ambiguous examples. Reinforcement learning and preference training can rely on human ratings. Safety pipelines may ask people to classify harmful material, while evaluation teams score model outputs or check edge cases. Even after launch, workers may correct errors, moderate content, or update records. Automation therefore often redistributes work instead of eliminating it. The conditions vary considerably. Pay structures may be hourly, task-based, or mediated through contractors, and can depend on task complexity, location, qualification, and platform rules. Workers may have limited control over pace, quality standards, appeals, or how their annotations are used. Specific wages and conditions should be attributed to a study or reporting period rather than presented as universal facts. Worker status and labor protections are determined by applicable law and the details of the relationship. Responsible AI supply-chain governance should identify where human work occurs, who employs or contracts the workers, what tasks and exposures are involved, how compensation and quality are determined, and whether workers can raise concerns. Vendors should be asked for labor, safety, and grievance information, not only model-performance metrics. Transparency about data labor helps companies account for costs and risks that an “automated” product description can obscure.