AI Systems Inherit Colonial Power Structures

The Servers Are Not Neutral

The first thing you need to understand about artificial intelligence is that it is not built in a vacuum. It is built in a room full of people. Some of those people are in San Francisco, writing code and collecting salaries that exceed the GDP of small nations. Many more of those people are in Nairobi, Manila, or Bogotá, labeling data for pennies a day, their labor invisible to the final product.

This is not a bug in the system. It is the system.

In a 2023 paper published in Philosophy & Technology, political theorists James Muldoon and Boxi Wú argue that the global infrastructure of AI production mirrors the colonial structures that have organized the world for centuries. They call it the “colonial matrix of power,” a term borrowed from the Peruvian sociologist Aníbal Quijano, and they use it to show how AI systems inherit the hierarchies, labor regimes, and knowledge systems of colonialism itself (Muldoon & Wú, 2023).

The paper is not an accusation of bad actors. It is a structural diagnosis. And it is deeply uncomfortable to read.

What Is the Colonial Matrix of Power?

Quijano’s idea is deceptively simple. The colonial matrix of power is a framework for understanding how colonialism did not just end with independence. It restructured itself. The old hierarchies of race, labor, and knowledge persisted, just without the formal colonial administration. The global economy still runs on a logic where the Global North extracts value from the Global South, and the Global South supplies raw materials, cheap labor, and data.

Muldoon and Wú take this framework and apply it to AI. They argue that AI production today is not a break from colonial history. It is a continuation. The data centers, the supply chains, the annotation workers, the intellectual property laws, the algorithms themselves all sit inside this matrix (Muldoon & Wú, 2023).

The authors do not use the word “colonial” loosely. They trace specific mechanisms.

The Colonial Supply Chain of AI

The first mechanism is the supply chain. An AI model like GPT-4 or Stable Diffusion does not appear out of thin air. It requires massive amounts of data, computing power, and human labor. That data does not come from nowhere. It comes from the internet, which is dominated by English-language content produced in the Global North. The computing power comes from rare earth minerals mined in the Democratic Republic of Congo, often under conditions that human rights organizations have called exploitative. The human labor comes from a global workforce of data annotators, content moderators, and testers, many of whom are in the Philippines, Kenya, India, and Venezuela.

Muldoon and Wú argue that this supply chain is not just unequal. It is structurally colonial. The value flows upward. The costs and harms flow downward (Muldoon & Wú, 2023).

Consider the data annotators. These are the people who label images, transcribe audio, and flag toxic content so that AI models learn correctly. They are often paid below a living wage. They work on precarious contracts. They are exposed to traumatic content with minimal psychological support. And their labor is rendered invisible. The AI’s final output does not credit them. The user does not see them. The company does not protect them.

This is not an accident. It is a design feature of the global digital economy.

The International Division of Digital Labor

The second mechanism is what Muldoon and Wú call the “international division of digital labor.” This is a fancy way of saying that the world’s work is divided along lines that look a lot like the old colonial map. High-skill, high-pay work happens in the West. Low-skill, low-pay work happens everywhere else.

The authors point out that this division is not natural. It is enforced. Western technology companies hold the patents, the copyrights, and the data. They control the platforms. They set the terms. Majority world countries supply the raw material (data, labor, minerals) and then buy back the finished product (AI services) at a markup (Muldoon & Wú, 2023).

This is the same pattern that defined colonialism: extract raw materials, process them in the metropole, sell the finished goods back to the colony at a profit.

The authors find that this division is not just economic. It is epistemic. It determines what counts as knowledge.

Whose Knowledge Gets Built Into AI?

This brings us to the third argument of the paper, and perhaps the most unsettling one. AI systems do not just reflect the world. They actively shape what the world looks like. And the world they are shaping is a Western one.

Muldoon and Wú argue that AI production perpetuates “hegemonic knowledge production through Western values and knowledge that marginalises non-Western alternatives” (Muldoon & Wú, 2023). This is not a conspiracy. It is a structural outcome. If your training data is mostly English, mostly from the United States and Europe, and mostly from platforms that privilege certain voices over others, then your AI will learn those patterns. It will treat Western norms as universal. It will treat non-Western knowledge as noise or error.

The authors give the example of language models. A model trained on English text will not just be bad at Swahili or Tamil. It will encode assumptions about the world that come from English-speaking cultures. It will treat Western legal concepts, Western family structures, Western notions of privacy as default. Everything else becomes a deviation.

This is not a bug. It is a feature of how the training data is collected and curated.

The Problem of Universality

One of the most seductive claims about AI is that it is objective. It is math. It is neutral. Muldoon and Wú push back hard on this. They argue that the discourse of universality and objectivity is itself a colonial move. It erases the conditions under which the AI was produced. It makes the labor, the extraction, and the hierarchy invisible.

When a CEO says “the algorithm is fair,” they are not describing reality. They are performing an act of erasure. The algorithm is fair only if you accept the premises built into it. Those premises were chosen by people in a specific place, with specific interests, working within a specific power structure (Muldoon & Wú, 2023).

The authors do not claim that AI is inherently bad. They claim that the current mode of AI production is colonial. And that is a different argument entirely.

How the Study Was Done

Muldoon and Wú are political theorists, not computer scientists. Their method is not an experiment with control groups and p-values. It is a theoretical synthesis. They read the political economy literature on AI production. They read the decolonial AI literature. They read Quijano and the modernity/coloniality research program. Then they wove these threads together into a single analytical framework.

This is a common method in critical theory. You do not test a hypothesis with a dataset. You test it with logic, evidence from multiple fields, and historical comparison. The strength of the paper is not in a single data point. It is in the pattern recognition. The authors show that the same structures that organized colonial extraction in the 16th century are organizing AI extraction in the 21st.

The paper has been cited 98 times as of the writing of this article. That is a strong signal that it is being taken seriously in academic circles.

What the Research Does Not Prove

It is important to be clear about what this paper does not claim.

It does not claim that every AI system is colonial. It does not claim that all Western AI researchers are colonialists. It does not claim that AI cannot be used for liberation. It does not claim that decolonizing AI is impossible.

What it does claim is that the default mode of AI production, as it currently exists, is embedded in a colonial structure. The burden of proof is on those who want to build a different AI to show that they have actually broken the pattern, not just reformed it.

The paper also does not provide a detailed roadmap for decolonization. That is not its goal. It is a diagnosis, not a prescription. The authors are clear that they are opening a conversation, not closing one.

Why This Matters Now

The timing of this paper is not accidental. We are at a moment when AI is being deployed into every domain of life: hiring, policing, medicine, education, credit scoring, content moderation. If the systems are built on colonial logic, then they will reproduce colonial outcomes. They will extract value from the poor to enrich the rich. They will treat Western knowledge as universal. They will make the labor of the majority world invisible.

This is not a prediction. It is already happening.

The paper gives us a language to name it. That is its most powerful contribution. Once you see the colonial matrix of power, you cannot unsee it. You notice the data annotators in the Philippines. You notice the cobalt mines in the Congo. You notice that every AI assistant speaks English with an American accent. You notice that the people writing the ethics guidelines are almost all from the Global North.

The paper does not solve these problems. But it makes them impossible to ignore.

What This Actually Means

• ▸If you work in AI, ask who built your training data. Not just where it came from, but who labeled it, under what conditions, and for what pay. If you do not know the answer, that is a problem.
• ▸If you fund AI research, stop treating ethics as a box to check. Ethics is not a separate module. It is the structure of the system. Fund research that builds AI from the ground up with decolonial principles, not just Western ones with a multicultural veneer.
• ▸If you regulate AI, look at the supply chain. The harms of AI are not just in the output. They are in the production. Regulate the labor conditions of data annotators. Regulate the extraction of minerals. Regulate the terms of data ownership.
• ▸If you use AI, ask whose knowledge it is encoding. If the system cannot handle your language, your culture, your way of thinking, that is not a technical limitation. It is a political choice.
• ▸If you study AI, take the colonial matrix seriously. It is not a metaphor. It is a structural analysis. Treat it with the same rigor you would treat an economic model or an algorithm. It is a lens, and it shows things that other lenses miss.

The paper by Muldoon and Wú is not the last word on this subject. It is the first word in a conversation that has been overdue. The AI industry likes to talk about the future. This paper asks us to look at the present, and at the past that made it. That is harder. That is necessary.