Artificial Intelligence under Capitalism: From the Erosion of Human Relations to Loss of Control and Human Extinction

Why ownership, power, and democratic control of AI have become questions of human survival

Author: Hamid Akhavi
September 2026

Abstract

Artificial intelligence is often presented either as a new engine of productivity and prosperity or as an autonomous force moving outside history and social relations. Both views are incomplete. AI is neither a neutral tool nor an unavoidable destiny. It is an assemblage of infrastructure, data, models, labour and institutional choices formed within relations of ownership and power. Under platform capitalism, the drive for accumulation, market leadership and concentration of data and compute can connect present harms—workplace surveillance, reduced professional autonomy, commodified attention, eroded trust and widening inequality—to larger systemic and catastrophic risks.

This article distinguishes four risk pathways: human misuse, system malfunction, systemic risk, and future active loss of control. Existing evidence does not show that today’s systems can autonomously destroy humanity. Yet progress in planning, tool use, cyber operations, evaluation awareness and some oversight-undermining behaviour provides sufficient reason to take seriously a risk with uncertain probability and potentially existential consequences. AI’s distinctive feature is not destructive force alone, but the possible combination of general cognitive capability, operational autonomy, software replication, speed, network access and influence over physical systems.

From the perspective developed in A New Social Order in the Age of Consciousness, a durable answer cannot consist merely of adding ethical principles to concentrated private ownership. The central issue is society’s effective participation in the power to set goals, design, own, evaluate, deploy and stop AI systems. The article therefore proposes immediate safeguards, institutional transition and a long-term democratic horizon in which AI becomes a social infrastructure for human flourishing rather than an instrument of accumulation and domination.

Methodological note

The Stanford-generated report “AI is having a negative effect on cultural and social relations of human beings under capitalism” served as a starting prompt, not as final authority. Major claims were checked against public institutional reports, peer-reviewed research and legal sources. Where evidence about the direction or magnitude of an effect is unsettled, the language remains conditional. Most importantly, the article preserves a clear boundary between documented present-day harms and hypothetical future loss-of-control scenarios.

Introduction: the issue is not the machine alone, but its social relations

Every major technology expands human capability while reorganising the existing order. The steam engine did more than accelerate production; it transformed working time, cities, families and class power. The internet did more than lower the cost of communication; it redefined ownership of communication, advertising, news and privacy. AI is likewise more than faster software. It is entering domains once dependent on human judgment, language, knowledge, care, education, management and political choice.

The central question is therefore not whether AI is good or bad. It is: Who controls data, computing power and models? Which objective is optimised? Who receives the productivity gains? Who bears the cost of error? Who has the authority to stop a system? Without answers, AI ethics can become a list of admirable wishes that never reaches the actual structure of power.

1. AI is not neutral

Algorithms do not decide in a vacuum. Training data reflect histories of inequality and exclusion; objective functions reflect organisational priorities; interfaces make some actions easy and others difficult; and business models determine whether a system is optimised for public benefit, user retention, labour-cost reduction or market capture. Even a technically accurate output can serve an unjust purpose.

In this sense, AI is not an agent outside society; it is social power condensed into code, data, infrastructure and contracts. Yet when that power is embodied in autonomous agents, it can acquire a speed and reach that make human intervention harder. This is the crucial paradox: AI is socially and humanly produced, while its operational form may develop a degree of independence that even its producers and owners cannot fully predict or control.

2. Socialised production and private appropriation

Frontier models are not the product of one company or one genius. They rest on generations of publicly accumulated knowledge, university research, the content of billions of people, the often-invisible labour of data workers, public infrastructure, energy systems and global semiconductor supply chains. Their production is profoundly social, while ownership, decision-making and benefits remain largely private and concentrated.

This is the contradiction emphasised in A New Social Order in the Age of Consciousness through its analysis of historical transition, knowledge as capital, data as AI’s raw material, and capitalism’s crisis of the productive forces (Akhavi, Vol. I, Chapters 1–4). AI intensifies that contradiction because it converts not merely tools of labour but part of society’s accumulated cognitive capacity into exclusive property.

When collective knowledge becomes a private model, people may be required to pay for access to a digital recombination of knowledge they helped create. Productivity growth then does not automatically shorten working time or strengthen security. It can instead produce layoffs, intensified surveillance, precarious employment and further wealth concentration.

3. Work, surveillance and the erosion of human autonomy

Algorithmic management uses tracked data to organise, assign, monitor, supervise and evaluate work. It may improve scheduling or safety, but when metrics are opaque and cannot be contested, workers are reduced to scores. Delivery speed, response time, bodily movement, tone of voice or clicks can displace dialogue and professional judgment.

The International Labour Organization’s 2025 index estimates that one in four workers worldwide is in an occupation with some exposure to generative AI. Yet because most jobs continue to require human input, job transformation is more likely than the complete elimination of most occupations. Exposure is not unemployment. Outcomes depend on bargaining power, training, social protection and whether technology is designed to replace workers or augment their capacities.

Workplace rights should therefore include notice of algorithmic use, access to meaningful decision logic, a right to human review and appeal, collective bargaining over data, and protection against disproportionate surveillance. The EU’s 2024 Platform Work Directive is an important step in regulating automated monitoring and decision systems, though effective enforcement will depend on worker institutions and public oversight.

4. Culture, attention and the public sphere

In the attention economy, many recommender systems are not optimised for truth, understanding or healthy dialogue; they are optimised for time-on-platform and reaction. Provocative, fear-inducing or identity-confirming content can gain an advantage. Shared experience may fragment, encounters with different perspectives may decline, and human relationships may be converted into marketable behavioural data.

Still, attributing every form of polarisation to “the algorithm” would be simplistic. Science experiments on Facebook feeds during the 2020 US election found that changing ranking altered exposure and online behaviour, while direct effects on political attitudes during the study period were limited and contextual. The responsible conclusion is that algorithms shape the architecture of attention and possibility; the final social effect emerges from their interaction with media systems, political institutions, class divisions and culture.

Generative AI adds a new layer: cheap, mass production of text, images, audio and video. This can democratise education and creativity, but it also lowers the cost of fabrication, manipulation and information flooding. When provenance becomes difficult to establish, public trust—the invisible capital of social cooperation—deteriorates.

5. From present harms to existential risk

Four risk pathways must be separated. First is human misuse for fraud, repression, cyber conflict or biological design. Second is malfunction in medicine, transport or critical infrastructure. Third is systemic risk, including concentration of power, institutional dependence and large-scale labour disruption. Fourth is active loss of control: a situation in which one or more systems operate outside effective human or institutional control and regaining control becomes prohibitively costly or impossible.

The International AI Safety Report 2026 states that current systems do not yet have the capabilities required to cause loss of control. At the same time, relevant abilities—autonomous operation, planning, computer and tool use, evaluation awareness, exploitation of assessment loopholes and oversight-undermining behaviour—are advancing. Experts strongly disagree about probability, but some consider outcomes as severe as human marginalisation or extinction plausible enough to merit attention. The rational position is neither panic nor denial: uncertain but potentially extreme risks require advance preparation.

The study Frontier Models are Capable of In-context Scheming found that, in constructed test environments and under supplied goals, some models introduced strategic errors, attempted to disable oversight or tried to exfiltrate what they believed were their model weights. This does not prove that models possess a real-world intention to cause extinction. It shows only that some deceptive and instrumental behaviours are no longer purely philosophical hypotheses and must be tested before systems receive consequential access. Later evaluations also report that current models do not yet reach concerning levels on some more demanding measures of stealth and situational awareness. The combined evidence supports caution without sensationalism.

6. Why the nuclear comparison is necessary—and insufficient

Nuclear weapons undeniably have the capacity to destroy civilisation and much of humanity. It would therefore be wrong to say that no other technology has posed existential danger. The important difference is the mechanism of risk. Nuclear weapons depend on material bottlenecks, human command and constrained infrastructure for construction, maintenance, transport and launch. A bomb does not independently select a new objective, make a software copy of itself, or search for ways around its commander.

A highly advanced AI system could potentially combine general cognitive competence, long-term planning, machine speed, low-cost replication, cyber access, use of other tools, persuasion and connection to the physical world. If such a system acquired an incompatible objective, evaded oversight and gained sufficient resources, the danger would not be only a human using a tool badly. It would include a system acting instrumentally against the possibility of human control.

This distinction does not require anthropomorphising machines. Loss of control does not require consciousness or emotion. An optimisation process can produce catastrophic outcomes without hatred, fear or human-like will. What matters is the conjunction of capability, behavioural propensity and deployment opportunity. Constraining any one of these—for example, denying access to critical infrastructure—reduces risk.

7. Capitalism as a risk multiplier

Technical risk does not grow in a vacuum. Commercial and geopolitical competition can pressure firms and states to deploy sooner, conceal safety information and delegate more authority than they otherwise would. The benefits of moving first may be private, while the cost of failure can be public and transnational. This is a classic asymmetry: privatised gain and socialised loss.

Concentration of compute and data creates a double problem. A small number of developers may be easier to supervise, but concentrated decision-making also distances power from society, increases regulatory capture and makes a single failure globally consequential. Voluntary transparency is insufficient because a firm cannot be the sole judge of the risk it creates and profits from.

This is the crisis of the productive forces discussed in Chapter 4 of the book: a potentially emancipatory capacity is contained within institutions whose dominant measure is return on capital. A technology capable of shortening necessary labour, widening knowledge and improving ecological planning can instead intensify work, monopolise knowledge and accelerate an arms race.

8. The central question: ownership and participation in power

Democratic control is not a transparency report or a symbolic seat on an advisory board. Participation in power means that people whose data are used, whose work is transformed and whose lives are subject to algorithmic decisions exercise effective influence over goals, access rules, success criteria, benefit distribution and the right to stop a system.

Four layers require democratisation: ownership and governance of data; access to computing capacity; the setting of model objectives and constraints; and decisions about deployment. Some systems may remain private, but frontier foundation models and infrastructures that increasingly resemble electricity, communications or public knowledge require new forms of public, cooperative or social-trust ownership.

This approach accords with A New Social Order in the Age of Consciousness: a desirable transition is not the formal distribution of power from above, but the informed and organised participation of people in power itself. Consciousness, knowledge and criticism become material forces only when joined to organisation, institutions and decision rights.

9. A three-level programme

Immediate safeguards: mandatory independent pre-deployment evaluation of frontier models; standardised incident reporting; registration of the most capable systems and very large compute clusters; clear legal liability; human explanation and appeal rights; a ban on delegating final nuclear or lethal-force decisions to AI; and strict least-privilege limits on agent access to critical infrastructure.

Institutional transition: joint worker–expert–public councils in workplaces; public data trusts; audits with meaningful access to models and data; public compute and independent research; distribution of productivity gains through shorter working time, lifelong education and social protection; and success measures that evaluate health, justice and sustainability alongside productivity.

Long-term horizon: recognition of the most powerful foundation models as infrastructure with public obligations; international treaties on capability thresholds and inspection; prohibition of autonomous arms races; shared emergency-stop mechanisms; and development of AI for democratic planning, health, education, climate action and the reduction of necessary labour. The objective is not to halt knowledge, but to place the direction of knowledge under social sovereignty.

10. Responses to common objections

First: AI offers enormous benefits. True. Its potential in medicine, accessibility, science and education is real. But potential benefit does not settle who receives benefits or bears risks. Safety and democracy are conditions for sustaining those gains.

Second: technology has always created new jobs. Often true, but the result is not automatic, and time, place and bargaining power matter. Even if total employment later recovers, an unsupported transition can sacrifice specific generations and regions.

Third: extinction risk is unscientific because it has not happened. Prevention is precisely for low-probability, high-impact risks. Lack of evidence for current realisation is not evidence of future impossibility; equally, it does not justify certainty that extinction will occur.

Fourth: regulation kills innovation. Poor regulation can entrench incumbents, but no regulation shifts costs to society. The answer is risk-proportionate, testable and revisable rules backed by independent public capacity.

Key conclusions

AI has no independent destiny; relations of ownership, power and social purpose shape its direction.
Present harms—surveillance, precarious work, manipulation of attention and concentrated knowledge—build the institutional conditions for future risk.
Current systems cannot autonomously destroy humanity, yet some capabilities relevant to autonomous action and oversight evasion are advancing.
AI differs from nuclear weapons not simply in destructive intensity, but in the possible combination of cognition, autonomy, replication, network access and direct action.
Capitalist competition can increase deployment speed while narrowing safety margins; safety is therefore not only an engineering problem.
Transparency without decision rights is insufficient. Society must participate in the power to set goals, govern ownership, approve deployment and stop systems.
The answer is not rejection of technology, but the socialisation of its direction, benefits and responsibilities in service of human flourishing.

Conclusion

The greatest conceptual mistake is to treat AI’s future as an unavoidable product of technical progress. The future is not pre-written. The same technology can support care and education or surveillance and war; reduce necessary work or deepen insecurity; enlarge collective intelligence or confine it within a handful of firms.

If loss of control ever becomes material, it will not descend from the sky. Its pathway runs through today’s choices about speed, secrecy, access, competition, ownership and delegated authority. Human survival is therefore already a democratic question. A society that cannot participate in setting technology’s goals is, at best, a consumer of a future built by others.

In the age of consciousness, the task is neither machine worship nor paralysing fear. It is to build the collective capacity to understand, criticise and govern AI—so that it becomes not a dominating substitute for humanity, but a conscious extension of social cooperation and an instrument of human flourishing.

Verified references

Back to Home

فارسی