Brown student creates a map tracing AI’s data-labeling labor across 31 countries

Cunyan Ma wants to make AI’s hidden workforce visible. The Brown University undergraduate has created a free interactive map connecting corporate customers, data-labeling companies and worker locations across 31 countries. It offers a wider view of those relationships, but not a complete census: company delivery centers serve as a proxy for where workers are based.

Brown described the project on October 26. Ma, a senior working with the university’s Center for Technological Responsibility, Reimagination and Redesign, built it to connect detailed accounts of individual workers with a larger picture of the industry. Her stated goal is public education: placing AI within the economic and political structures that supply its labor.

Following the chain through intermediaries

The map combines existing databases of data-labeling firms and their customers, Ma’s own data collection, and news reports describing workers’ conditions. Her original database tracks labeling companies’ delivery centers. The interactive tool draws connections among corporate customers, the firms handling their data work, and countries containing worker sites.

That approach follows an outsourcing structure. Corporations ranging from technology and social media businesses to retailers and defense companies hire specialists to label data. Those firms, in turn, seek inexpensive labor. Ma points to Kenya, Indonesia and Colombia as examples of countries where that work is outsourced.

The location method matters when reading the result. A delivery center is a stand-in for workers’ whereabouts, rather than a direct record of each person doing the work. Ma describes the map as illustrative, not comprehensive. Its current coverage is a set of linked organizations and places, not the full global AI workforce.

13Data-labeling companies

Labeling firms represented in the map.

141Corporate customers

Customers linked to those labeling firms.

31Countries

Countries containing data-worker sites in the map.

The human work behind model behavior

Data work supplies the human judgments used to train AI systems. It is not one uniform task, and it extends beyond attaching simple tags to pictures. Brown’s account describes work ranging from routine identification to assessments requiring people to interpret language, evaluate responses or examine medical images.

• Labeling objects in photographs or identifying parts of speech in sentences.
• Transcribing speech and rating chatbot responses as good or bad.
• Marking tumor boundaries in medical scans or identifying disturbing material for content filtering.

The last category carries a particular burden. Teaching systems to recognize and filter disturbing content can require workers to view that material themselves. Ma describes people repeatedly reading violent or sexually abusive content to help train safer AI. She uses the term “AI colonialism” for the extraction of difficult, low-paid labor from the Global South for Western profit.

From public education to worker connections

Ma calls her approach power mapping: tracing how global power structures shape the distribution of labor. She argues that AI’s labor chain is deliberately opaque to avoid accountability. The map is her attempt to make those relationships easier to see, rather than leave public understanding dependent on isolated accounts of working conditions and pay.

Her adviser, Brown computer science professor Suresh Venkatasubramanian, frames the project as a corrective to accounts that treat AI as an independent, living intelligence. He emphasizes the human knowledge, understanding and effort beneath the algorithms: systems constructed by people working around the world.

Ma also hopes a future version could help workers find one another. She says the workforce’s scattered nature leaves people unaware of fellow workers, making collective organizing difficult. That is an ambition for a later iteration, not a current organizing feature. For now, she plans to keep expanding the map.