AIdataNewsGlobal Explainer: Inside the Brown University Map Tracing AI’s Hidden Human Supply Chain

When digital platforms recommend videos, filter harmful posts, translate text, or generate replies, users usually see the machine output and assume the system is operating independently. A map developed by Cunyan Ma at Brown University challenges that impression by foregrounding a second layer of the AI stack: the human work that makes everyday algorithmic services possible.

PROVIDENCE, R.I. — The project, created while Ma collaborated with the Center for Technological Responsibility, Reimagination and Redesign at Brown University, assembles existing databases, reporting, and records Ma collected into a freely available interactive map. The map is designed to make visible a structure in which technology companies, intermediary firms, and dispersed data workers are linked across borders, much of it within the Global South.

From an industry-observer perspective, the map matters because it shifts the question from “Is this AI accurate?” to “Who is being paid, where, and under what conditions, to make it usable?” Ma frames the exercise as a way to expose not merely isolated workplace grievances, but a broader arrangement of labor, extraction, and value.

Ma said:

My primary motivation is to visualize the idea of AI colonialism in an impactful and immediately accessible manner

The gap I’m trying to fill is that most of what we know about data workers comes from anecdotal reports from authors and journalists. Those are great because they give us a very detailed look at working conditions and pay, etc. But I’m hoping this might start to fill in a larger picture of what AI colonialism actually looks like.

I’m hoping this might start to fill in a larger picture of what AI colonialism actually looks like. — Cunyan Ma

The routine tasks behind “automatic” intelligence

For an AI model to perform many functions, people must first prepare the material it learns from. The tasks can appear small on their own: marking what appears in an image, identifying grammatical roles in a sentence, converting speech into text, rating the usefulness of a chatbot response, or outlining regions in medical scans. Individually, these are forms of human judgment and categorization. Collectively, they are the scaffolding that supports many AI services users encounter online.

• Tagging objects in photos
• Tagging parts of speech in sentences
• Transcribing lengthy speech
• Judging whether chatbot responses are good or bad
• Marking tumor boundaries in medical scans

Ma emphasizes that the work is not only tedious. Systems must learn to identify harmful content in order to filter it. That requirement means data workers may be exposed to horrifying images or passages so the technology can recognize them and reduce their presence for others.

Outsourcing, intermediaries, and the geography of AI labor

Ma points to a chain of delegation. Corporations that operate online-facing products need labeling and evaluation work. They often send that work to specialist companies, and those firms then look for inexpensive labor. The result, in Ma’s description, is a geography in which difficult tasks are relocated across borders while profits are retained elsewhere.

Ma said:

How that manifests in reality is that companies outsource that work to the Global South — Kenya, Indonesia, Colombia and elsewhere

So you have folks around the world who have to read paragraph after paragraph of AI-produced child pornography, bestiality, violent content or whatever else, just to try to train AI to be safe.

She describes that arrangement as the essence of AI colonialism: hard, low-paid labor is pulled from the Global South so that profit can accrue in the West.

What the map records — and what it does not

To trace the chain, Ma built a network view rather than a single-country account. The map draws connections among data labeling companies, the corporations that use them, and the locations where the work appears to be performed. Ma treats delivery centers as a proxy for worker locations, acknowledging that the map is not a full census of AI labor.

According to the data cited in the project, the current form shows links among 13 data labeling companies, 141 corporate customers, and 31 countries containing data worker sites. Ma states that the map is not a comprehensive view, but is intended to be illustrative.

Map element	Cited value	Function in the visualization
Data labeling companies	13	Intermediary firms connecting corporate clients with distributed labor
Corporate customers	141	Organizations whose products rely on outsourced training and moderation work
Countries containing data worker sites	31	Locations linked to delivery centers that serve as proxies for worker presence
Examples named by Ma	Kenya, Indonesia, Colombia	Illustrative destinations cited in her description of Global South outsourcing

For readers trying to interpret the numbers, the map’s value is structural: it does not claim to measure every worker or every contract. Instead, it asks whether users and analysts can begin to see the pattern formed by firms, intermediaries, and dispersed workplaces.

Ma said:

One purpose for this is as a public education tool

When people want to learn about AI and its social impact, the map puts it in the context of this larger economic and political structure.

The argument for “power mapping”

Suresh Venkatasubramanian, a Brown computer science professor who leads the Center for Technological Responsibility, Reimagination and Redesign and advises Ma’s project, views the map as a corrective to a familiar narrative problem. He argues that much public discussion treats AI systems as if they possess agency of their own, while ignoring the collective human work embedded in them.

In his framing, the project helps restore the social substance underneath automated output: AI is not a self-generating oracle, but an aggregation of knowledge, understanding, and labor assembled by people.

Venkatasubramanian said:

It’s all humans underneath

That’s the real story behind the generation of AI. The problem with a lot of current discourse around AI is that it seeks to ignore that. The result is that we start constructing stories around AI as its own sentient being, as something that’s alive. But it’s not. It’s something that’s built — something that’s constructed by us, by people, all across the world.

From visibility to potential worker connection

Ma describes the labor chain as deliberately opaque, which makes accountability difficult. That opacity also fragments workers. Because data work is distributed across firms and countries, participants may not realize they share the same economic position or the same set of intermediaries.

Ma said:

One of the things about this is that it’s distributed

People don't even know about their fellow co-workers. That makes unionizing or anything like that just kind of impossible. Maybe they can use a future iteration of the map to identify each other.

Her longer-term plan is to expand the visualization beyond the current scope. Ma says she hopes to chart other aspects of the material foundations beneath algorithms, including labor implicated in the semiconductor industry. As a social analysis and research concentrator, she views power mapping as a method for understanding how global power structures shape the distribution of labor worldwide.

Frequently asked questions

What is the map meant to explain?

It is meant to make visible the human labor structure behind AI services, especially work performed in the Global South and mediated by data labeling firms and their corporate customers.

Why did Ma create it?

She wanted to visualize AI colonialism in an accessible way and to complement anecdotal reports with a broader picture of how the labor chain appears to be organized.

What data points are explicitly cited in the current version?

The map shows links among 13 data labeling companies, 141 corporate customers, and 31 countries containing data worker sites, while Ma notes it is not comprehensive.

Why are countries named in the article not presented as a full list?

Kenya, Indonesia, Colombia and elsewhere appear as examples in Ma’s description of outsourcing. The cited map scope is broader but still limited by available databases, Ma’s own collected data, and news reports.

Who are the main stakeholders shown in the map?

• Corporations that need labeled or moderated data for products and services
• Specialist companies that receive outsourced work
• Data workers, especially in the Global South
• Researchers, educators, and the public trying to understand AI’s social impact

What role does Suresh Venkatasubramanian play?

He is the Brown computer science professor who leads the Center for Technological Responsibility, Reimagination and Redesign and advises Ma’s project.

What does the map suggest about accountability?

Ma argues the chain is kept deliberately opaque, making accountability difficult. The map is therefore not only a data visualization, but a tool for public education and, potentially, future worker identification.

Key takeaways

1. AI systems depend on extensive human labor, including tagging, transcribing, rating outputs, and medical annotation.
2. Much of that labor can be outsourced to low-cost locations, with the Global South named as a major destination in Ma’s account.
3. The current map links 13 data labeling companies, 141 corporate customers, and 31 countries, but is explicitly described as illustrative rather than comprehensive.
4. The project frames this pattern as AI colonialism: difficult labor is extracted globally while profit concentrates in the West.
5. Beyond reporting, Ma and Venkatasubramanian emphasize the need to counter narratives that treat AI as detached from the people who build it.

Why this story fits a general AI news audience

The map sits at the intersection of infrastructure, enterprise applications, and data governance. It does not describe a single model release or a regulatory measure, but rather the operating conditions beneath automated services. That makes it useful for readers who want a compact way to connect technical adoption with labor geography and accountability questions.

Its relevance also lies in what it leaves unfinished. The visualization is a starting point, not a complete account, and future expansions may bring additional material layers into view. For now, the central contribution is simple: it gives readers a way to follow the trail from algorithmic convenience to the people whose work makes it possible.

Editorial contact: editorial@aidatanews.com — David Nakamura, AI & Data News