Following the links behind data labeling
A Brown University undergraduate has built an interactive map tracing some of the companies and workers involved in preparing data for artificial intelligence systems. The project connects 13 data-labeling businesses with 141 corporate customers and worker sites across 31 countries, offering a public view of a supply chain that is often difficult to examine.
Cunyan Ma, a Brown senior working with the university’s Center for Technological Responsibility, Reimagination and Redesign, assembled the tool from existing data-worker databases, news reports and original research into labeling companies’ delivery centers. Those centers serve as a proxy for the location of workers rather than a complete census of individual jobs. Ma describes the map as illustrative, not comprehensive.
Data workers perform tasks that help train and evaluate AI systems. Their work can include identifying objects in images, tagging parts of speech, transcribing recordings, rating chatbot answers and marking boundaries in medical scans. Safety work can also require people to review disturbing text or imagery so that a system can learn to recognize content that should be filtered.
The map is designed to show how technology companies frequently contract with specialist labeling firms, which in turn recruit workers in lower-cost labor markets. Ma uses the term “AI colonialism” to describe the extraction of difficult, low-paid work from countries in the Global South for companies and consumers elsewhere. The visualization presents that argument through connections among clients, intermediaries and locations rather than treating AI as a purely automated product.
A public-education tool with room to grow
Ma said the immediate goal is to give students and the public an accessible way to understand the economic and political structure beneath AI services. Suresh Venkatasubramanian, a Brown computer science professor and Ma’s adviser, said the work helps counter narratives that portray AI as an independent intelligence by emphasizing the accumulated human effort used to construct it.
The project could also have practical value for workers. Because labeling jobs are distributed across employers and countries, workers may have little visibility into who performs similar tasks elsewhere in the same supply chain. Ma hopes a later version could help them identify one another. That possibility remains an aspiration rather than a feature or documented outcome of the current map.
Ma plans to expand the data and explore other material foundations of computing, including labor in the semiconductor industry. Any such expansion will face the same transparency challenge as the current project: outsourcing chains change, company relationships can be opaque, and delivery-center locations do not reveal every worker or employment condition.
The map’s present contribution is therefore a structured starting point. Its figures should not be read as the full scale of global data labor. They document a set of visible relationships and give researchers, workers and users a tool for asking who supplies the human judgment embedded in everyday AI products.
