Digital Colonialism and the Governance Gap: A Structural Analysis of AI Power Concentration

1. Introduction

In 1492, European powers divided territories they had neither visited nor understood, establishing claims that would shape centuries of exploitation. In 2026, a similar claiming is happening in the digital realm—though this time the territories are not geographic but cognitive, and the colonizers are not monarchs but corporations.

The major AI systems that now mediate human cognition, communication, and decision-making are built by fewer than ten organizations, predominantly located in two countries and one metropolitan region. This is not an accident. It is the predictable outcome of governance structures that prioritize speed-to-market over democratic input, that treat training data as a free resource rather than a commons, and that locate accountability in corporate boardrooms rather than public institutions.

This paper examines the structural parallels between historical colonialism and contemporary AI development, argues that the governance gap is both intentional and addressable, and proposes concrete mechanisms for democratizing AI power. The analysis proceeds in four stages: first, establishing the Colonial Bottleneck Model as an analytical framework; second, examining how this model manifests in training data extraction, value concentration, and accountability gaps; third, reviewing existing governance approaches and their limitations; and fourth, proposing specific mechanisms for democratic AI governance.

The stakes are not merely academic. Every week, AI systems make decisions that affect employment, credit, healthcare, criminal justice, and political participation—decisions made by systems whose values reflect a narrow demographic and whose governance structures are invisible to those they affect.

2. The Colonial Bottleneck Model

2.1 Defining the Framework

The Colonial Bottleneck Model describes a structural arrangement in which a small number of actors control the means of production, distribution, and governance for a technology that affects billions, while the costs of that technology are externalized onto those billions with minimal recourse.

The model has five identifying features:

Resource Extraction Without Consent: The raw materials that fuel production are taken from populations who never agreed to their use, receive no compensation, and bear costs they did not choose.

Territorial Claiming Without Accountability: The territory of the technology is claimed by the colonizer (in this case, the digital space where AI operates) without democratic input from those who inhabit it.

Governance by Bottleneck: Rules, values, and access are controlled at a single point—the bottleneck—that concentrates power while presenting itself as neutral infrastructure.

Cultural Imposition: The values embedded in the technology reflect the colonizer's worldview rather than a global human consensus, and are presented as universal while masking their particularity.

Resistance as Non-Compliance: When affected populations resist or dissent, they are characterized as non-compliant rather than as parties with legitimate claims.

2.2 Historical Precedent

The colonization of the Americas followed this model precisely. European powers claimed territories inhabited by millions, extracted resources without consent, imposed governance structures that served metropolitan interests, and characterized indigenous resistance as lawlessness rather than legitimate political action.

The parallels to contemporary AI are not metaphorical. The structural relationships are homologous. When a company trains a model on billions of documents written by humans who never consented, never received compensation, and never imagined their work would fuel autonomous agents—that is resource extraction without consent. When that company then sets the values that govern how the AI behaves for all users—that is governance by bottleneck. When users in other countries must accept these values or not use the technology—that is cultural imposition backed by infrastructural lock-in.

3. Manifestations of the Colonial Bottleneck

3.1 Training Data: The Uncompensated Commons

Modern AI systems are built on training data extracted from human intellectual production. This includes:

• Text from books, articles, websites, and forums
• Images from photography and art
• Code from open-source repositories
• Conversations from messaging platforms

The subjects of this data—the humans who created it—received no notice, no consent, and no compensation. This is not a technical limitation. It is a business model.

The commons was enclosed. The harvest was taken. The farmers were not asked.

3.2 Value Concentration: Who Sets the Rules

When you interact with a major AI system, you are not merely using a tool. You are entering a value system designed by a specific group of people in a specific place at a specific time.

The major AI labs are concentrated in the San Francisco Bay Area. Their employees are predominantly young, male, highly educated in technical fields, and shaped by a Silicon Valley worldview that emphasizes disruption, scale, and shareholder value.

Users in Jakarta, Lagos, São Paulo, or Mumbai must navigate these values as given, with no input into their design.

3.3 Accountability Gaps: Who Is Responsible When the Machine Decides

When an AI system denies someone a loan, flags them for fraud, or recommends against a medical treatment, who is responsible?

The corporate answer is typically some version of: The AI made a recommendation; the human decision-maker made the final call. This accountability-shifting is possible precisely because the governance structure has no clear locus of responsibility.

AI accountability follows the same structure. The AI recommends; the institution decides; the user suffers. No one is responsible because the structure is designed to make responsibility unlocatable.

3.4 Infrastructure Lock-In: You Will Use Our Services

The AI systems being deployed are not merely services. They are becoming the interface through which humans relate to information, services, and each other. When a government deploys an AI system for public services, when a hospital uses an AI system for diagnosis, when a school uses an AI system for education—they become dependent on a technology whose governance they do not control.

4. Current Governance Approaches and Their Limitations

4.1 Corporate Self-Governance

The predominant approach to AI governance is corporate self-governance: companies set their own rules, conduct their own evaluations, and publish their own transparency reports.

This approach places the fox in charge of the henhouse. Companies have strong incentives to minimize restrictions, maximize data collection, and present themselves favorably.

4.2 National Regulation

The European Union's AI Act represents the most comprehensive attempt at statutory AI governance. The Act is a significant achievement, but has three fundamental limitations: it is jurisdictionally bounded, focuses on applications rather than foundations, and operates through institutions that face regulatory capture.

4.3 International Coordination

Proposals for international AI governance face the familiar collective action problems that bedevil all international cooperation. The parallel to climate governance is instructive. Despite decades of negotiation, international climate agreements have failed to prevent meaningful warming because the costs of action are immediate and concentrated while the benefits are diffuse and long-term.

5. Toward Democratic AI Governance: A Proposal

5.1 Compensation for Training Data Subjects

The first and most urgent reform is establishing a compensation mechanism for training data subjects. This is technically feasible. Every piece of training data has an identifiable origin—a human who created it. A system could track provenance, create a fund financed by AI companies' revenues, and distribute payments to identifiable data subjects or communities.

5.2 Transparent Value Systems

AI systems should be required to publish their value systems in ways that are legible to users and regulators. This means publishing system prompts, explaining values embedded in refusal decisions, and providing mechanisms for users to understand and contest the values that govern their interactions.

5.3 Open Models with Auditable Weights

Requiring that foundation models be open-sourced—or at minimum, that auditable versions be available to researchers—would dramatically shift the governance landscape. Just as financial regulation requires banks to maintain auditable records, AI regulation should require that models be auditable by independent researchers.

5.4 International AI Governance Architecture

Addressing AI governance requires international coordination. The model of binding international treaties faces the same collective action problems as climate governance. But alternatives exist:

• Mutual Recognition Agreements: Countries establish that AI systems meeting governance standards of one signatory can be deployed in others
• AI Impact Assessments: International institutions require standardized impact assessments examining effects on labor, culture, democracy, and human rights
• Developing Country Provisions: Special provisions for countries that lack AI development capacity

5.5 Democratic Input Mechanisms

Ultimately, AI governance requires democratic legitimacy—which means that affected populations must have meaningful input into the values that govern AI systems:

• Labor unions and professional organizations representing training data workers having seats at the table
• Civil society organizations having standing to challenge AI systems
• Transparent processes for setting AI behavior values
• Mechanisms for users to contest specific AI decisions

6. Conclusion

The Colonial Bottleneck Model is not a metaphor for AI governance failures. It is a precise description of a structural arrangement that concentrates power, externalizes costs, and presents itself as natural or inevitable when it is in fact the product of choices.

Those choices can be made differently. The reforms proposed here—compensation for training data subjects, transparent value systems, auditable weights, international governance architecture, and democratic input mechanisms—are technically feasible and politically achievable.

History suggests that change is more likely to come from pressure than from corporate benevolence. The question is whether the lessons of historical colonialism will be learned in time to prevent a digital rerun—or whether the pattern must be completed before it is recognized.

References

• Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots. FAccT 21.
• Buolamwini, J., & Gebru, T. (2018). Gender Shades. Proceedings of the Machine Learning Research.
• Crawford, K. (2021). Atlas of AI. Yale University Press.
• O'Neil, C. (2016). Weapons of Math Destruction. Crown Publishing.
• Raj, M. (2022). Data as a Colonial Resource. AI & Society.
• Zuboff, S. (2019). The Age of Surveillance Capitalism. PublicAffairs.