A Brown Undergrad Mapped AI's Hidden Data Labor Across 31 Countries

Brown University published an interactive map on October 26 that makes the least visible part of the AI supply chain traceable: built by Cunyan Ma, a senior concentrating in social analysis and research, working with the university’s Center for Technological Responsibility, Reimagination and Redesign, it connects 13 data-labeling companies to 141 corporate customers and worker sites across 31 countries. The firms involved include the industry’s known names in annotation outsourcing, and the customers span social media platforms, AI labs, retailers, and defense contractors. Her advisor, Suresh Venkatasubramanian, who leads the center and formerly worked in the White House Office of Science and Technology Policy, summarized the thesis in one line: “It’s all humans underneath.”

What the map actually shows

The methodology aggregates existing databases of labeling firms and their clients, adds an original database of delivery centers as a proxy for worker locations, and layers in news reporting on working conditions, with power-mapping used to trace the connections. The tasks being traced are the foundation of every model this site covers: labeling objects in photos and parts of speech in sentences, transcribing speech, rating chatbot responses, drawing tumor boundaries in medical scans, and reviewing violent or abusive content so that AI systems can learn to filter it. The last category is the one that gives the map its moral weight; content moderation work, outsourced to Kenya, Indonesia, Colombia, and elsewhere at wages the map describes only qualitatively as low, is the labor that makes every safety claim in every model card possible, and its psychological cost is documented in the reporting the map aggregates.

Why a map, and why now

The absence of this map was itself a design choice by the industry. AI companies publish model cards and, increasingly, energy and compute disclosures, but the labor supply chain has stayed deliberately opaque to avoid accountability for wages and conditions; Ma’s framing calls it “AI colonialism,” the extraction of low-wage labor from the Global South for Western profit, and notes the opacity is what prevents workers at dispersed delivery centers from even knowing they share an employer. The map is explicitly illustrative rather than comprehensive, which is the right humility: 13 firms and 141 customers is a floor, not a census. But as a public-education artifact it changes what a journalist, a procurer, or a worker can see: the same tracing exercise that Anthropic’s S-1 does for compute commitments, this does for the humans in the pipeline.

What to watch

Three developments would give the map consequence beyond documentation. First, procurement: if enterprise buyers of AI begin asking model vendors for labeling-labor disclosures the way they now ask for carbon accounting, the map becomes an audit standard rather than an advocacy tool. Second, worker organizing: Ma’s stated hope is that dispersed workers can use it to find each other, and cross-border content-moderator organizing has precedent but no shared infrastructure. Third, the semiconductor extension: the center plans to map chip-industry labor next, which would extend the same tracing to the hardware layer whose geopolitical walls, export controls and licensing regimes, dominated this month’s news. The AI industry publishes what it builds; this map is an early attempt to publish who builds it, and the gap between those two disclosures is where the next several years of AI-labor accountability will be fought.