Hidden labor behind AI: a Brown student's map links 13 data-labeling firms to 141 corporate customers and worker sites in 31 countries

The hidden labor behind AI now has a map. On October 26, 2026, Brown University described an interactive map by senior Cunyan Ma that traces the human work of labeling data for AI systems: which companies buy it, which intermediary firms supply it, and where those firms' workers sit. In its current form, Brown said, the map links 13 data-labeling companies to 141 corporate customers and to worker sites in 31 countries. It's a small, openly partial dataset, but it targets a part of AI and work that official statistics barely capture. It also complements a very different map, Mercor's attempt to measure which jobs AI benchmarks cover, which our explainer on Mercor's benchmark map examines.

This story is based on Brown University's October 26, 2026 article on the project, the International Labour Organization's Convention No. 193 and its February 2026 research brief on platform-worker numbers, and TIME's 2023 investigation of data labeling in Kenya, checked October 27, 2026. The map itself, at ai-materiality-map.org, loads its content with JavaScript; its headline figures come from Brown's description and were checked against the three public data files the map loads, as they stood on October 27, 2026.

The short version

• What it maps: the chain from tech companies to the data-labeling firms they outsource to, to those firms' delivery centers around the world.
• Key numbers (current version, per Brown, and matching the map's data files): 13 labeling companies, 141 corporate customers, worker sites in 31 countries.
• Where the data comes from: existing databases such as dataworklandscape.org, Ma's own database of delivery centers, and news reports on working conditions.
• What it doesn't show: how many people work at each site, what they're paid, or which customer's work happens where. Ma calls it illustrative, not comprehensive.
• Why it matters: the ILO says current data "remain insufficient" to count platform workers at all.

What does the Brown map of AI data workers show?

The map visualizes a three-step supply chain. At one end are corporate customers: social media, tech, retail and defense companies that, Brown says, typically outsource data labeling to firms seeking inexpensive labor. In the middle are the 13 labeling companies. At the other end are the places where the work is done, which Ma approximates using each firm's delivery centers.

Labeling is the work of tagging images, text, audio and video so models can learn from them, and of rating model outputs. Brown notes that some of it requires workers to view disturbing content so AI systems can learn to filter it. Much of it, Ma says, happens in the Global South: "companies outsource that work to the Global South — Kenya, Indonesia, Colombia and elsewhere."

Brown's article doesn't list the 13 companies, but the map's public data files do. As of October 27, 2026, they name Centific, CloudFactory, Cognizant, DesiCrew, Digital Divide Data, iMerit, Impact Enterprises, IngeData, Innodata, Keymakr, NextWealth, Sama and TaskUs, and place 93 delivery centers in 31 countries, from Argentina to Zambia. By 9to5AI's count of those files, TaskUs accounts for 31 of the 93 sites. The customer file names 172 customers; 141 of them have map coordinates and are drawn on the map, which matches Brown's count.

Where does the map's data come from?

Brown names four inputs:

• Existing databases of data workers and labeling firms, including Data Work Landscape (dataworklandscape.org), which describes itself as shining a light on a "shadowy industry" of people who "develop, train, test, and maintain AI systems."
• Ma's own database of labeling firms' delivery centers, which supplies the locations.
• News reports describing working conditions.
• Her own data collection, which Brown says fills gaps left by anecdotal reports.

The method has a built-in limit. A delivery center is a place a firm says it delivers work from, not a headcount, and Brown's description doesn't say whether people working remotely for these firms are captured. So the map shows where firms have a presence, not how the work is distributed.

Who made it, and why?

Ma is a senior concentrating in social analysis and research. She works with Brown's Center for Technological Responsibility, Reimagination and Redesign (CNTR), led by computer science professor Suresh Venkatasubramanian, as part of an ongoing project on "the material underpinnings of the digital world."

Her framing is explicitly critical. "My primary motivation is to visualize the idea of AI colonialism in an impactful and immediately accessible manner," Ma said, defining it as "the extraction of difficult, low-paid labor from the Global South for profit in the West." She argues the supply chain is opaque by design, which limits accountability. Venkatasubramanian put it this way: "It's all humans underneath. That's the real story behind the generation of AI."

Ma also sees a practical use. Because the work is distributed, she says, "People don't even know about their fellow co-workers," and "maybe they can use a future iteration of the map to identify each other." Readers should take the map as an advocacy-minded research project with a stated point of view, not a neutral census.

What are the map's limits?

Ma names the main one herself: it is not a comprehensive view of AI labor chains but an illustration. Beyond that:

Last verified: October 27, 2026

Ma also notes that most of what is known about data workers comes from anecdotal reports by authors and journalists, detailed but limited in scope. Her map aggregates those sources rather than replacing them.

What do official sources say about data-labeling work?

Mostly that it's hard to count. In a research brief published on February 16, 2026, ILO researchers Uma Rani, Michael Frosch and Morgan Williams updated estimates of active platforms and workers, but concluded that "current data sources remain insufficient to accurately estimate the total number of individuals who engage in digital platform employment," and called for internationally agreed statistical standards.

Regulation is starting to reach this work. In June 2026 the International Labour Conference adopted the ILO's first convention on the platform economy, the Decent Work in the Platform Economy Convention, 2026 (No. 193). Its definition of a digital labour platform applies "regardless of whether that work is performed online or in a specific geographic location," and it covers platform workers "regardless of their classification of status in employment." That plausibly includes the online microtasks behind AI training, although the convention doesn't mention data labeling by name. It enters into force "twelve months after the date on which the ratifications of two Members have been registered."

The best-documented single case remains TIME's January 18, 2023 investigation into Kenyan workers employed by Sama to label toxic text for OpenAI. TIME reported take-home pay of "between around $1.32 and $2 per hour depending on seniority and performance," while OpenAI's contracts set a rate of $12.50 an hour. Sama disputed the figures, saying workers could earn $1.46 to $3.74 an hour after taxes and that "the $12.50 rate for the project covers all costs." OpenAI told TIME that filtering harmful content "is a necessary step." Sama ended the work in February 2022, eight months early. Kenya is one of the countries Ma names, and the map lists a Sama delivery center in Nairobi.

What should builders and buyers take from it?

If you buy model training or evaluation data, someone labeled it, and you probably can't see who. The map is a reminder that procurement questions, such as which firms, which countries and what pay and support for workers handling disturbing content, have answers that are rarely published. The same opacity runs through AI and jobs data more broadly: measuring AI's effect on work is hard when the work itself isn't counted. And the IMF's widely cited estimate of AI's job exposure, explained in our IMF explainer, says nothing about the people whose labor trains the models.

What's confirmed, reported and uncertain

• Confirmed (Brown University, and the map's data files): the project, its creator and adviser, the 13/141/31 counts and the names of the 13 firms, the data sources and Ma's own statement of its limits.
• Confirmed (ILO): Convention No. 193's adoption in June 2026, its online-work definition and entry-into-force rule; the February 2026 brief's conclusion that platform workers can't yet be counted accurately.
• Reported (TIME, 2023): Sama's pay and contract figures, with Sama's competing figures.
• Uncertain: how current each link is, since the data files were last updated on October 27, 2026 and may keep changing, and why 31 of the 172 customers named in the files aren't drawn on the map.