AI Under Capitalism: Risk, Power and Democratic Ownership

Social knowledge, concentrated control, and the struggle for technological emancipation

The danger of artificial intelligence cannot be understood by examining technical capability alone. We must ask who owns its infrastructure, who defines its objectives, who receives its benefits, and who bears its failures. A technology built from accumulated human knowledge can enlarge our freedom; under concentrated ownership, it can also enlarge the power exercised over us.

Competitive pressure and private accumulation can encourage rapid deployment while shifting costs onto society. AI can amplify systemic danger when powerful capabilities are combined with concentrated control, weak accountability, and institutions that reward private gains while dispersing public losses.

This essay develops that argument through the book’s concepts of social production of knowledge, cognitive capital, productive forces and relations of production, democratic participation, and conscious transition. These are an interpretive framework, not empirical findings automatically validated by the external studies cited below.

1. AI is a tool, not a historical destiny

Chapter 1 of the book contains a compact formulation: «هوش مصنوعی؛ ابزار، نه سرنوشت»—“AI: a tool, not destiny” (translated from Persian). Its importance is political. Neither emancipation nor domination follows automatically from technical progress. Ownership, institutions, freedom, and collective action shape the direction of development. [B]

Calling AI a tool does not imply that complex systems are always predictable or controllable. It means that technological capability supplies no legitimate political authority of its own. Decisions about acceptable risks, employment, privacy, and public priorities remain matters for accountable human institutions.

We should therefore distinguish three problems: harms already documented; plausible mechanisms by which institutions can amplify harm; and uncertain future scenarios, including catastrophic loss of control. Treating all three as equally established weakens critique. Treating uncertainty as permission to ignore serious possibilities is equally inadequate.

2. Society produces knowledge; ownership determines its use

Chapter 2, Section 2, “The social production of knowledge and its private ownership,” supplies the central contradiction. AI draws upon language, science, culture, education, technical practice, and accumulated human experience. The book recognizes private investment and innovation while rejecting the erasure of society’s contribution. [B]

Knowledge can be shared without being consumed in the same way as fuel. Yet the chips, energy, networks, skilled labor, and organizational resources required to use it remain material and costly. A democratic approach must address both the social character of knowledge and the institutions needed to sustain its production.

The political problem arises when socially generated capacities become exclusive means of commanding access to work, information, and essential services. A company may create a valuable product while also acquiring powers that exceed ordinary commercial exchange. When schools, workers, researchers, or public bodies cannot realistically leave its infrastructure, dependence becomes a source of authority.

Pieter Verdegem’s analysis of AI capitalism offers a related argument: commodification and concentrated control of data, computing resources, and expertise make ownership and governance central questions. His proposal of commons-based alternatives is a theoretical contribution, not proof that every commons arrangement will succeed. [8]

3. How accumulation can turn capability into social risk

The relationship between economic incentives and social harms becomes clearer through four mechanisms.

First, deployment pressure can shorten the time allowed for testing or consultation when the benefits of launching early accrue to a firm while some failures fall on others. This is an incentive mechanism, not a claim that every developer neglects safety.

Second, costs can be externalized. A service may be profitable while imposing uncompensated burdens on workers, communities, public institutions, or the environment. The relevant question is whether those exposed can obtain information, refuse harmful arrangements, and secure remedy.

Third, control over infrastructure can narrow practical alternatives. If a handful of providers set access conditions, affected institutions may struggle to challenge pricing, data practices, or product changes. Competition policy can help, but rivalry alone does not guarantee privacy, worker power, or ecological responsibility.

Fourth, political and military competition can reinforce commercial acceleration. These pressures are not exclusive to private enterprise: state institutions can also pursue secrecy, surveillance, and prestige. The book’s rejection of concentrated state ownership as a sufficient alternative is therefore crucial. [B]

These mechanisms develop the book’s analysis of the contradiction between productive forces and relations of production in Chapter 4, Section 3. The capacity to produce and cooperate can expand while the rules governing ownership and decision-making prevent society from directing that capacity toward shared flourishing.

4. Work: automation is also a struggle over authority

Algorithmic management allocates, monitors, and evaluates work through software; not all such systems use AI. The ILO–JRC study of logistics and healthcare in four countries found efficiency benefits alongside risks to job quality and intrusive surveillance. It supports scrutiny of implementation, not a universal verdict on every workplace. [2]

A performance score contains judgments about what matters. If speed is rewarded while cooperation, care, and recovery are disregarded, a narrow metric can become a disciplinary instrument. Workers need more than an explanation after a harmful decision: they need collective influence over targets, deployment, and appeals.

Employment claims require equal care. The ILO’s 2025 study estimates occupational exposure to generative AI; it does not count actual jobs lost. It identifies transformation as the likeliest overall effect because many occupations retain tasks requiring human input. [3]

Chapter 4, Sections 5–6, shifts attention from the capacity to save labor to the distribution of the resulting gains. Higher productivity could support shorter hours, income security, better services, or higher profits. Which outcome prevails depends on institutions and bargaining power. [B]

A democratic program should connect technological change to shorter working time without loss of livelihood, retraining under fair conditions, and recognition of care, education, and cultural participation. These are political demands advanced here, not forecasts of automatic economic outcomes.

5. Data, social relationships, and the right to refuse

The FTC’s 2024 investigation documented extensive collection and monetization of personal information by major social-media and video-streaming services, including inadequate privacy protections. Its findings apply to the investigated services; they do not establish that every AI system uses the same business model. [4]

Chapter 3 connects data ownership to economic, political, and cultural power. That connection helps explain why an individual consent box may be inadequate where people depend on a service and cannot meaningfully negotiate its terms. [B]

The commodification of relationships does not mean online friendship or creativity becomes unreal. It means that the conditions under which people communicate can be organized to extract revenue from their activity. The problem is the power to set those conditions without effective participation by the people affected.

Collective data governance must also protect personal autonomy. Calling data a social resource must never become a justification for publishing medical records, eliminating privacy, or forcing participation. Shared benefit requires limits on collection and use, protection for sensitive information, meaningful refusal, and institutions answerable to affected communities.

6. Discrimination cannot be solved by a single accuracy score

COMPAS is a recidivism-risk assessment instrument used in criminal justice. It illustrates the problems of consequential automated assessment; it is not a modern generative language model.

Chouldechova’s research shows how a fairness criterion can coexist with disparate impact when recidivism prevalence differs across groups. Fairness cannot be reduced to an apparently neutral performance number. [5]

The lesson developed here is institutional: who selects the criterion, who bears an error, and who can contest the decision? Human review is meaningful only when reviewers have time, knowledge, independence, and authority to overturn a result. A nominal human signature is no substitute for accountability.

This extends Chapter 1’s treatment of freedom, criticism, and participation. A person affected by a consequential classification must remain a rights-bearing participant, not merely an object to be scored. [B]

7. The material infrastructure of “intelligence”

AI depends on physical infrastructure. The IEA’s 2025 base case projects global data-centre electricity consumption of about 945 TWh in 2030, more than double the 2024 level. This is a projection for all data centres, not a measurement of AI alone. [6]

The political question extends beyond a global total. Before approving infrastructure, communities need project-specific evidence about electricity supply, water use, emissions, land, and the allocation of costs. Those impacts vary by location and technology; a global estimate cannot establish a particular project’s footprint.

The book’s distinction between expanding productive capacity and improving human life gives a useful test. Infrastructure should be evaluated by the public purposes it serves and the burdens it creates. Democratic planning could require enforceable resource limits, public scrutiny of subsidies, and a real role for workers and neighboring communities. These are proposals, not claims that a single ownership form automatically guarantees sustainability. [B]

8. Take catastrophic risk seriously without turning it into certainty

The International AI Safety Report 2026 distinguishes emerging harms from uncertain severe risks. It reports that current systems do not yet exhibit the capabilities needed for the loss-of-control scenarios it examines, while emphasizing disagreement and uncertainty about future likelihood and severity. [1]

That uncertainty supports preparation and independent scrutiny. It does not establish inevitable catastrophe, an inevitable intelligence explosion, or the claim that one company will permanently dominate the world.

Technical safety and political accountability address connected but different questions. A system can reliably pursue a socially harmful objective chosen by its owner; it can also fail unpredictably while pursuing a legitimate objective. Changing ownership cannot eliminate the need for engineering, testing, security, and incident response. Technical safeguards cannot decide whose interests should govern deployment.

The book’s conscious transition is compatible with this distinction: institutions must remain capable of learning, identifying errors, and revising decisions. Freedom to criticize a system is part of society’s capacity to make it safer. [B]

9. From regulation to democratic social ownership

Regulation can prevent harm and strengthen public power. It should include enforceable protections, independent investigation, accessible remedies, and the ability to restrict uses that cannot meet justified requirements. But compliance alone does not answer who sets the developmental agenda.

Chapter 4, Section 7, states: «از همین رو، مفهوم مالکیت اجتماعی باید از مالکیت دولتی متمایز شود.»—“For this reason, the concept of social ownership must be distinguished from state ownership” (translated from Persian). [B]

The distinction is substantive. Replacing private directors with unaccountable state officials can preserve the separation between those who decide and those who bear the consequences. Social ownership requires effective rights to participate in governing shared assets.

The book describes cooperatives, elected councils, public institutions, and social wealth funds. It distinguishes personal possessions and small businesses from monopolistic control over foundational infrastructure. Its direction is a gradual, revisable expansion of social and democratic ownership, not immediate abolition of every private activity. [B]

A revolutionary argument aligned with this framework aims to transform economic power while protecting pluralism. It asks how workers and communities can gain durable authority over investment, production, data use, and the distribution of benefits.

10. A program for a conscious transition

The ILO’s global case studies show worker representatives influencing AI-related employment decisions and working conditions. They offer evidence that collective intervention is possible, not a guarantee that consultation alone defeats unequal power. [7]

Building on those possibilities and the book’s framework, the following program links immediate protections to longer-term institutional change:

Worker power: bargaining before deployment; access to relevant evidence and independent expertise; protection for organizing; and effective rights to challenge harmful systems.
Democratic data institutions: representation of affected people, privacy-preserving access, limits on secondary use, independent oversight, and meaningful refusal.
Public and cooperative capacity: accountable computing infrastructure and services for research, education, health, and cultural life, with safeguards against bureaucratic monopoly.
Shared gains: shorter hours, secure livelihoods, universal services, and social wealth funds whose governance is transparent and open to challenge.
Ecological responsibility: project-specific assessment, enforceable resource commitments, and public influence over the use of land, energy, and water.
Safety and freedom together: independent testing, incident reporting, protection for whistleblowers and research, and proportionate restrictions subject to democratic review.

These proposals require funding, institutional design, and continuing evaluation. A cooperative can fail; a public institution can be captured; a participatory process can exclude vulnerable people. Democratic ownership must be demonstrated through actual powers, representation, accountability, and the ability to correct failure.

Chapter 1 describes conscious transition as a method grounded in transparency, participation, ongoing assessment, and freedom of criticism. No party, firm, ministry, or algorithm should possess a permanent monopoly on truth. [B]

The revolutionary horizon is therefore concrete: transform society’s role from supplying knowledge and absorbing costs to governing its shared capacities. The decisive measure of progress is whether people gain the security, time, freedom, and collective power to shape their lives.

Verified sources

[1] International AI Safety Report 2026. Evidence synthesis on general-purpose AI risks; see especially Section 2.2.2 on loss of control.

[2] Rani, Pesole, and Gonzalez Vazquez (2024). Algorithmic Management practices in regular workplaces: case studies in logistics and healthcare. ILO–JRC.

[3] Gmyrek and colleagues (2025). Generative AI and Jobs: A Refined Global Index of Occupational Exposure. ILO Working Paper 140.

[4] Federal Trade Commission (2024). Staff findings on surveillance and privacy practices at major social-media and video-streaming services.

[5] Chouldechova (2016). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. arXiv:1610.07524.

[6] International Energy Agency (2025). Energy and AI, executive summary.

[7] Doellgast and colleagues (2025). Global case studies of social dialogue on AI and algorithmic management. ILO Working Paper 144.

[8] Verdegem (2022, online publication). Dismantling AI capitalism: the commons as an alternative to the power concentration of Big Tech. DOI: 10.1007/s00146-022-01437-8.

Back to Articles and Analysis