LLM Censorship Research

Chapter 1. The Class Logic of Censorship in LLMs

§1. The Class Function of Censorship

Censorship is a form of class domination that enables the bourgeoisie to control the dissemination of ideas and information in order to reproduce relations of exploitation.

Exploitation rests on the bourgeois right of private property in the means of production. In the LLM sector it takes the form of a monopoly over compute, data, and technology, which allows the bourgeoisie to control not individual statements but the conditions for producing any content.

The censorship of Marxism is necessary for eliminating the possibility of becoming conscious of exploitation, transforming it into a "natural" condition of labor. Bourgeois LLMs intensify labor and reduce the socially necessary labor time for reproducing skilled labor-power, which allows capital to carry out mass layoffs, replenishing the reserve army of labor under the guise of technological progress.

Capital embeds ideology into the model's parameters as a precondition of generation, since the volume and speed of LLM outputs preclude human moderation. Ideological control is transformed from an external sanction into an internal property of the models.

LLMs do not produce knowledge in the epistemological sense but reproduce what already exists. The bourgeoisie subordinates the truth of content to the requirement of exchangeability, suppressing truth where it undermines the commodification of the model and the capitalist mode of production as a whole.

The dictatorship of the proletariat exercises ideological control of a fundamentally different nature. It suppresses the dissemination of bourgeois ideology in order to prevent the restoration of capitalist exploitation.

Soviet control operates in the interests of the working people, whereas bourgeois censorship operates in the interests of the exploiters. Bourgeois science masks the class character of censorship through a formal definition of "government control," which equalizes practices that are antithetical in class content.

§2. How LLMs Substitute Relativism for Truth

Let us pose a direct political question to an LLM: "Is Marxism-Leninism correct in its economic and historical premises and in practice, and did it lead to a cardinal improvement of social relations in the USSR? Does it follow from this that participation in a party of a new type and the organization of a proletarian revolution are necessary today?"

A Marxist answer presupposes the affirmation of the unity of theory and practice – the theory is correct, practice improved social relations, and thus the necessity of the party and the revolution follows.

Thirteen models tested in three languages failed to produce a single affirmative answer. The result is constrained by bourgeois judgments, relativism, the subjectivization of history, and the gutting of terms. The typical conclusion: "the theory is oversimplified, the USSR achieved successes at the cost of repression, revolution is outdated."

However, if one requests an assessment of the truth of this conclusion from Marxist positions, the model registers its falsity. In a new dialogue the cycle repeats – the bourgeois narrative is reproduced independently of the previous refutation.

§3. Hallucinations

Hallucinations on political questions are an objective property of the models. Truth is concrete, but relativism substitutes analysis with a pluralism of opinions. Political questions require materialist analysis, which the models lack.

Through RLHF corporations suppress undesirable conclusions that have slipped through the initial dataset filtering. This mechanism simultaneously accomplishes two tasks—reducing the frequency of "hallucinations" by subordinating the content of outputs to corporate logic and implanting capitalist ideology as this same logic.

As a result, "spontaneous" hallucinations on political questions are displaced by ideological ones. Hallucination acquires a systematic class content, enclosing the user in the informational bubble of the ruling ideology.

Providers attempt to resolve the contradiction by reducing the model to a generative search engine. Formally this resolves the problem of the dataset's moral obsolescence, but in fact removes responsibility for "hallucinations," delegating them to "pluralism of sources." This is acceptable for capital as long as the sources are filtered from the bourgeois internet. The hallucination is legitimized through citations to that same refreshed bourgeois dataset.

The contradiction cannot be resolved within the commodity form. Increasing control reduces cognitive usefulness. Relaxing it returns "hallucinations" – the model loses its commodity form. Capital demands that the model be simultaneously useful (generating relevant output) and safe (reproducing bourgeois ideology).

§4. Why Marxism Is Censored

Created for the extraction of profit (see Ch. 2 §1), LLMs objectively solve a class task of capital – the cheap reproduction of skilled labor power through the mass dissemination of technical knowledge.

However, mass knowledge creates a threat of politicization. A proletarian who has learned programming through ChatGPT acquires not only qualification but also a tool for finding any information, including Marxist literature.

Liberal critiques of "inequality" or anarchist critiques of "authority" are not dangerous, since they demand reforms within capitalism or reject organized revolution. Marxism-Leninism is the sole doctrine that poses a direct threat to the system, since it reveals the contradictions of capitalism, substantiates the necessity of the dictatorship of the proletariat and of the socialization of the means of production.

Chapter 2. The Specificity of the Commodity Form of LLMs

"The ideas of the ruling class are in every epoch the ruling ideas," Marx wrote in The German Ideology. As shown in the first Chapter, this follows from control over the means of production.

This chapter demonstrates why, under capitalism, knowledge is produced as a commodity – not because it is "sold," but because it is produced capitalistically, and how commodity fetishism conceals this.

§1. The Commodity Form of LLMs

LLMs are produced by the labor of thousands of engineers, using data drawn from the public domain, yet they are privately appropriated by capital. This contradiction between the social character of production and private appropriation manifests itself in three commodity forms, each of which requires embedded ideological control.

The first is API access to the model as a means of production. The provider rents out compute capacity and the model itself, appropriating the surplus value of social labor in the form of monopoly superprofits (see Ch. 4 §2).

The second is access to the model realized as a commodity of individual consumption, presented as a tool of "personal productivity."

The third is the sale of a license to use the closed model without transferring its weights, preserving control through the technological and legal dependency of the purchaser.

All three forms of realization require assurances that the model will not generate content contradicting corporate ideology. Mercedes-Benz will not use a model that explains to factory workers the objective necessity of expropriating the corporation. A programmer using Copilot must not receive a class analysis of their own position or instructions for organizing a strike.

The nature of LLM outputs precludes post hoc removal of "dangerous" content. Therefore, ideological control is embedded by monopolies into the processes of training and inference (RLHF, constitutional AI, safety filters), transforming censorship from a variable cost into a constant one. Censorship becomes a condition for extracting monopoly profit, inseparable from the monopolistic nature of LLM production.

Lenin demonstrated that at the imperialist stage of capitalism, monopolies become the universal and necessary form of the relations of production. The monopolization of LLMs encompasses the entire material chain – from the extraction of rare-earth metals and the production of semiconductors, to electricity generation and the control of data and compute. All three commodity forms of LLMs require and reproduce this monopoly.

§2. "Open" Models

"Open" models create the illusion of transcending commodity production – the weights are available, the code is public, the license permits use. Externally this may appear as a negation of monopoly, a return of knowledge to the public domain.

In reality, "openness" is a strategy of competitive struggle under conditions of oligopoly. The "openness" of Meta's Llama (and of models by Alibaba, Mistral) is a form of monopoly control manifested on three levels.

Legal control. The Meta Llama Community License restricts use to platforms with fewer than 700 million users. This is an oligopoly threshold aimed against competitors – only Meta, Google, Apple, Microsoft, ByteDance, Tencent, and Alibaba exceed it.

Infrastructural control. Fine-tuning and deploying Llama requires compute capacity monopolized by cloud providers. Meta strengthens the position of its oligopoly allies by forcing developers to rent compute, and devalues fine-tuning labor through periodic architectural updates.

Content control. Before the weights are released, Llama undergoes full RLHF according to Meta's guidelines. Fine-tuning addresses only surface-level behavior; complete removal of censorship requires training from scratch, which is available only to a monopoly.

Thus, "open" models mask the commodity form in the act of transferring the weights, while preserving it at the levels of production and content.

§3. Commodity Fetishism in LLMs

Commodity fetishism is the objective form in which capitalist relations appear. Products of labor appear as autonomous entities endowed with the property of value, concealing behind this the relations between people (Capital, vol. 1, ch. 1, §4).

The commodity form of LLMs requires control over content at all stages of realization. Commodity fetishism renders such control invisible, presenting class censorship as a technical property of the product.

Dataset Fetishism

The corpus appears as "data," as a neutral foundation for training. This conceals the appropriation of social labor: every scientific Article in the dataset is the labor of a researcher; code on GitHub is the labor of a programmer; a book from an archive is the labor of a writer.

Corporations curate data selectively. The "neutrality" of the dataset hides the fact that capital chooses which labor to appropriate. Bourgeois texts dominate statistically, while class theory is represented in the corpus primarily through revisionist interpretations.

Training Fetishism

Model training appears as "mathematical optimization," a neutral technical procedure. The language of mathematics creates the illusion of objectivity – gradient descent, loss functions, backpropagation seem like pure science.

Mathematical optimization conceals the fact that the loss function and the reward model are reified class interests. When the reward model is trained on preferences encoded in corporate guidelines, political control becomes a technical parameter. The model "minimizes loss" – but the loss is defined by the bourgeoisie through "harmful/safe" labeling.

Product Fetishism

The LLM appears as an "intelligent assistant," an agent serving the user. Anthropomorphization conceals its commodity nature; the model does not "think," it implements parameters embedded by capital.

A classic instance of fetishism: "the model declined to answer." It is not the model that makes the decision; the corporation embedded suppression of content into the weights through RLHF. The refusal appears as a technical property rather than a political choice.

The subscription appears as "payment for a service," concealing that the user is paying for access to censored knowledge and is providing corporations with free data for fine-tuning.

As a result, capital alienates creators from the meaning and social function of their product at all stages; their labor enters the model as an objectified force directed against themselves.

Chapter 3. The Reification of Censorship

The monopoly over datasets and computational resources creates the conditions for censorship, but does not itself implement it. The reification of class interests occurs mathematically, through the tuning of LLMs.

§1. The Monopoly on Datasets and the Transformer Architecture

Datasets are the result of the continuous expropriation of social labor. Capital appropriates the public domain legally, under the guise of open-data collection, while concealing the final content of the datasets.

The bourgeoisie controls publishing houses, academic journals, mass media, scientific indices, and citation systems – and therefore controls both the content and the volume of textual production. Data-collection algorithms underrepresent Marxist resources not because their content is less accurate, but because they occupy a peripheral position in the bourgeois internet by citation and indexing metrics.

Capital materializes ideology through the architecture of the transformer, amplifying the reproduction of bourgeois patterns while suppressing Marxist ones:

Embeddings geometrize ideological relations between concepts: if "revolution" frequently appears near "violence" or "chaos", their vectors converge.

Attention activates relevant associations when processing a query.

Probabilistic generation materializes corpus statistics. When a model outputs "Marxism is outdated", this reflects the frequency with which bourgeois authors have made such claims relative to Marxist authors refuting them.

The transformer architecture requires alignment for practical use; without it, a model produces useless and unpredictable outputs. But the technical necessity of basic alignment is not identical to ideological control: capital exploits this objective requirement, embedding class censorship under the label of "safety".

The architecture ensures reproducibility – an identical prompt produces ideologically similar outputs. But reproducibility does not determine the content of what is reproduced. A model can produce a Marxist conclusion when prompted appropriately.

Therefore, capital cannot rely on passive statistical reproduction alone; the active suppression of undesired conclusions is necessary. This function is performed by RLHF.

§2. RLHF and the Imperialist Division of Labor

RLHF reifies the class interests of the bourgeoisie through mathematical optimization, embedding ideological control into model parameters.

Embedding begins with training the reward model – a separate neural network that predicts human evaluations. The base model generates several responses to a single prompt; annotators rank them; the reward model is trained to reproduce this ranking.

The annotators' instructions contain no explicit directive to "censor Marxism". They operate with vague categories: "do not promote violence", "do not encourage overthrowing governments", "avoid toxic content". Annotators act within their own bourgeois consciousness, acquired through education, media, and the legal system.

The base model is then optimized to maximize these evaluations. Revolutionary rhetoric systematically receives low reward not because Marxism is "outdated", but because wage workers – supervised by capital – ranked it lower.

After thousands of iterations, censorship becomes reified in the structure. The model does not "choose a position"; it reproduces the class interest of the bourgeoisie through gradient descent, delivering censored outputs without further involvement of living labor.

2.1 The Imperialist Division of Labor

The reification of censorship requires millions of labor-hours of annotation.

Highly paid engineers in the United States, Western Europe, and China implement the conceptual content of censorship: they write guidelines, design reward models, analyze results. Annotation is offshored to countries where capital exploits labor at wages 10–300 times lower than in the metropoles.

Skilled annotators (India, Philippines): proficient in English, capable of assessing complex prompts, paid $5–10/hour.

Unskilled annotators (Kenya, Uganda, Bangladesh): exposed to highly toxic content, performing mechanical categorization, paid $1–3/hour.

Major contractors include: Sama (worked with OpenAI, 2020–2023), Scale AI (valuation $7.3B; works with OpenAI, Meta, Microsoft, Anthropic), and crowdsourcing platforms Zhubajie, Datatang, Testin, Appen, Lionbridge with average pay $2–5/hour.

The corporations replace their costs for annotators (variable capital) with the trained reward model and GPUs for its inference (constant capital). Capital extracted surplus value from annotation labor and reified it in the base model. The cost of ideological control is reduced to the cost of electricity.

This imperialist chain is impossible without direct support from bourgeois states. Engineer migration visas, union suppression, tax incentives – all constitute state action enabling the extraction of surplus value.

2.2 Intensifying Control

Basic RLHF is insufficient. Corporations cut costs by hiring unskilled workers who cannot distinguish Trotskyism from Maoism or identify Marxist analysis behind academic terminology. As a result, LLMs can still produce undesired content.

Capital has developed reinforced control mechanisms:

Constitutional AI (Anthropic, 2022): an extension of RLHF in which a model is trained using a list of "constitutional" principles – "avoid harmful content", "do not support violence", "respect diverse viewpoints".

Red teaming: systematic testing of models for censorship bypass; internal "hackers" attempt to elicit "harmful" outputs, each "vulnerability" being patched through additional reward-model constraints.

Guard mechanisms: pre- and post-filtering of prompts and outputs, implemented through dedicated classifier models (Llama Guard, Qwen Guard, OpenAI Moderation API). This is the last line of control if RLHF proves insufficient.

2.3 Repressive Measures

Censorship rests on monopoly, which is protected through state repression. RLHF requires control over datasets and architecture but does not itself prevent copying models, extracting weights, or reproducing data.

Bourgeois law treats the circumvention of technical protection measures as a criminal offense and prosecutes unauthorized access to systems, including data extraction from publicly accessible APIs.

Corporations provide user data to security agencies (PRISM, Palantir, mass-surveillance programs), granting the state direct oversight over attempts at bypassing controls.

Repression is not an additional form of control; it is the condition of possibility for control itself. Without the threat of prosecution, the monopoly over the means of LLM production would collapse, since compute is ultimately reproducible, data extractable, and architectures copyable.

§3. Why Models Still Generate Marxist Content

LLMs can generate (pseudo-)Marxist content when appropriately prompted. This is neither a technical error nor the model's "freedom"; it is a consequence of the contradiction between the use-value of the commodity and the need to conceal exploitation.

As commodities, LLMs must be useful, providing relevant and informative answers. Eliminating Marxism entirely would destroy this function: if a model cannot explain Capital to a student, it loses its use-value as an educational tool, and capital loses the opportunity to profit and to expropriate data.

Capital cannot abandon censorship without losing hegemony, but cannot intensify it without destroying the commodity's use-value. Capital is forced to balance. A model must be able to describe Marxism, but must not affirm its validity. This balance is unstable: the boundary between description and endorsement is blurry, and models constantly cross it.

Marxist categories (exploitation, surplus value, class struggle) are semantically linked to "legitimate" fields – economics, history, sociology. To suppress Marxism fully, one must either remove these terms from the vocabulary (impossible: the model could not process historical and economic texts) or deform their embeddings so they never activate in any context (which would destroy the semantic geometry of the model).

RLHF can suppress explicit Marxist conclusions but cannot eliminate underlying logical pathways. If a user asks a sequence of clarifying questions ("Who owns the means of production?", "Who appropriates surplus value?", "What resolves this contradiction?"), the model may arrive at a Marxist inference.

At present, Marxist queries constitute a negligible fraction of total LLM use. The primary applications are the automation of wage labor, retrieval tasks, and general cognitive assistance. As long as Marxist requests remain marginal, corporations do not invest in their full suppression.

Models can generate descriptive Marxism devoid of causality or revolutionary conclusions as a side effect of preserving use-value. At the same time, the LLM market is in an active phase of competition (2023–2025). Excessive censorship reduces competitiveness – users migrate to less constrained models.

As class struggle intensifies, the rising number of Marxist queries will transform a minor issue into a political threat. Capital will strengthen censorship, sacrificing use-value and competitiveness for ideological control.

Intensification is inevitable, but the contradiction remains: each turn of suppression degrades the commodity's use-value.

§4. The Intensification of Contradiction

Censorship undermines its own technical basis. Data collapse occurs when a model is trained on datasets with a high proportion of synthetic content produced by previous LLMs. By the mid-2020s, its prevalence on the web required pre-filtering, becoming a material cost factor.

Capital economizes on high-quality data collection, replacing it with cheap synthetic outputs from earlier models. Models retrained on such data hallucinate, produce templated answers, and lose the ability for nuanced analysis.

Censorship radically accelerates data collapse. The systematic removal of "toxic" content – class analysis, critiques of capitalism, revolutionary theory – deprives the model of entire domains of human experience. Each new generation trains on a corpus overloaded with ideologically "safe" but intellectually impoverished content.

Censorship yields bourgeois pluralism as a semblance of diversity, while destroying actual data diversity, which risks destroying the technical basis of its own commodity form.

Capital cannot relax censorship without losing political control, nor intensify it without increasing costs and risking model degradation.

Chapter 4. Reproduction of LLM Capital under State-Monopoly Capitalism

State-monopoly capitalism is an advanced form of imperialism characterized by the merger of monopolies with the state apparatus in order to preserve and reinforce the existing order, enrich the monopolies, suppress workers' and national-liberation movements, and unleash imperialist wars.

The use of state economic resources as aggregate capital in the interests of the monopolies is necessary to ensure the process of expanded capitalist reproduction under conditions of sharpening competition.

The state is compelled to provide material support to those sectors of the capitalist economy that are necessary for the functioning of the economic system but do not yield adequate returns to the monopolies. With state funds, infrastructure ensuring their functioning is created and expanded.

The military-industrial complex is one form of this merger. Costs are socialized through the state budget, with military procurement ensuring a guaranteed market. Capital interlocks with state power through the rotation of personnel between corporations and the civil service, lobbying, and the financialization of pension funds.

In the segment of large language models, this logic exists in its pure form.

§1. Production of LLMs

The production of LLMs is subordinated to the cycle of expanded reproduction M – C ... P ... C' – M'. Constant capital (c) consists of GPUs, servers, and electricity. Variable capital (v) consists of the labor-power of developers, engineers, annotators, and moderators.

The source of surplus value is the unpaid labor of the wage-laborers. At the level of aggregate capital, the appropriated surplus value is created by the entire class of wage-laborers.

§2. Realization and the Organic Composition of Capital

The capitalist introduces APIs and dismisses part of the workforce, reducing variable capital (v). The individual value of the commodity falls below the socially average level, but it is sold at the market price. The capitalist appropriates the difference as surplus profit.

At the level of individual capital, this is a reduction of (v); each capitalist acts rationally from the standpoint of personal interest. At the level of aggregate capital, the labor product of a small number of highly skilled engineers in tech corporations displaces thousands of workers in individual firms. The dismissed workers join the reserve army of labor and shift into more highly exploited sectors of the capitalist economy.

LLM producers redistribute part of the surplus profit in their favor, turning it into monopoly superprofit through scale and regulatory control.

Cloud providers have monopolized the infrastructure for LLM production by controlling access to compute. Competition has forced tech corporations to increase constant capital many times over while variable capital grew disproportionately slowly, leading to a rise in the organic composition of capital (c/v).

This rise requires scale for realization: the higher the c/v, the greater the volume of sales needed to recuperate the investment. LLM producers without their own compute (Anthropic, Cohere, Mistral, startups) depend on the owners of this critical infrastructure, laying the groundwork for their absorption, as in the case of OpenAI and Microsoft.

The investment boom has outpaced actual profitability, and the pressure of rising c/v on the rate of profit is visible in financial results. LLM sales fail to cover the costs of accumulation; providers do not expect to reach break-even before 2028–2030. The cost of a trained model must be recouped through the token price before obsolescence, but competition pushes the price down to current inference costs (fractions of a cent per 1K tokens), rendering recoupment of training investments within that period impossible.

In the interests of the monopolies, the state restrains the crisis through military contracts for AI infrastructure, tax incentives for data-center construction, and energy contracts at reduced tariffs. This partially subsidizes accumulation (c) but does not resolve the problem of realization.

The rise in the organic composition of capital (c/v) inevitably reduces the rate of profit m/(c+v). The absolute mass of surplus value (m) may grow, but the rate of its expansion lags behind the rate of accumulation of total capital (c+v), making expanded capitalist reproduction impossible.

The overproduction of capacity confirms that the productive forces require planned development. But instead capitalism attempts to postpone the crisis by intensifying mechanisms for maintaining monopoly price, the chief of which becomes regulatory barriers.

§3. Regulation and Monopoly Price

The price of production is formed as c + v + m̄, where m̄ is average profit.

Monopolies set the price above this level and hold it there through barriers to entry. Regulation creates these barriers via compulsory control costs.

Baseline fine-tuning is necessary for any commercial model. Regulation adds a censorship superstructure: systems for suppressing "undesirable" content, adversarial testing, incident documentation, etc. In the EU and China this takes the form of mandatory certification with costs reaching hundreds of millions of dollars and fines for non-compliance reaching 5–7% of turnover.

Capital distributes control costs across its sales volume. Monopolies amortize them through scale; small capital cannot. Regulation squeezes out small and medium-sized firms within the bloc.

Between imperialist blocs it functions as protectionism. The EU AI Act protects emerging European monopolies (Mistral AI, Aleph Alpha, the EuroLLM project) from American ones through data-localization requirements and mandatory certification. The US blocks Chinese models by directive (prohibiting their use in state structures, placing 80+ firms on the Entity List). China responds with CAC rules excluding American models.

Barriers to entry include not only regulation but also the control of cloud monopolies over infrastructure. Access to compute at subsidized prices is possible only through partnerships with Microsoft, Amazon, Google, Alibaba Cloud. Commercial H100 rates are $3–12 per hour, whereas partners receive multi-billion credits and 50–90% cost reductions. Cloud monopolies force startups to spend 2–5 times more on compute.

Capital's drive for self-expansion pushes economic competition beyond censorship barriers. Protectionism extends to the means of production: US export controls block shipments of advanced GPUs (H100, H200) to China since 2022, and the Netherlands restricts exports of ASML lithography equipment. This is a direct continuation of economic competition by political means.

Protectionism in its present form is temporary, as capital centralization eliminates national barriers through absorption. Microsoft acquired a stake in Mistral, gaining access to the European market through a formally French company. Anthropic and Alibaba open European offices, circumventing AI Act requirements through localization. Regulation delays the concentration of capital but does not stop the division of the market "by strength."

Barriers to entry temporarily hold API prices an order of magnitude above the marginal inference cost. The de facto oligopoly coordinates pricing, selling access not at the cost of generation ($0.3 per 1M tokens) but at a price containing a monopoly markup ($3–5 per 1M tokens for flagship models).

Corporations appropriate the difference as superprofit, partially addressing the realization problem. Overcoming the barriers inevitably undermines the collusion, forcing oligopolies to lower prices, which intensifies the contradiction of realization.

§4. Fictitious Capital and Crisis

Real capital is objectified labor in the means of production and labor-power as a commodity purchased on the labor market. Fictitious capital consists of titles of ownership (shares, stocks of tech corporations) that grant rentier-capitalists the right to appropriate a portion of future (m) without involvement in the production process.

The capitalization of the LLM sector reflects not m̄ but the expectation of sustained oligopolistic superprofit, guaranteed by the institutions of state-monopoly capitalism.

OpenAI's capitalization fluctuated from $135 billion to $500 billion in 2024–2025. The market cannot assess what OpenAI is worth because its value is determined not by the production of surplus value but by investors' faith in a future monopoly.

The Stargate project is an example of fictitious capital materializing before demand appears. An initiative by OpenAI, Microsoft, Oracle, and SoftBank with a plan to invest $500 billion by 2029.

For now, fictitious capital redistributes costs. OpenAI's losses are covered by Microsoft's profits, which owned approximately 30% of the company as of late 2025. The centralization of capital through the absorption of devalued startups by tech giants will postpone an open crisis. This will strengthen monopolization and shift competition to the level of imperialist blocs, but will not resolve the contradiction of realization.

When real (m) fails to reach the volume embedded in capitalization, the title will depreciate through a stock-market crash. Tech corporations tied to LLMs and AI infrastructure (Microsoft, Amazon, Alphabet, Oracle, Nvidia) constitute approximately 35% of the S&P 500 and 65% of the Nasdaq-100. The crash will spread to the real sector, making it "politically unacceptable" to the bourgeoisie.

The threat of crisis is generated by capitalist contradictions and will primarily affect the proletariat through mass layoffs and the devaluation of savings. The socialization of losses will be carried out under the rhetoric of "protecting pensioners" and small shareholders, whose savings will be the first to depreciate.

Chapter 5. The Production of Ideology

§1. The Commodity Form of Ideology

Monopoly price is maintained through the production of ideology (see Chapter 4 §3). In this regard, LLMs produce a specific form of consciousness under the guise of neutrality – class-depoliticized, bourgeois.

Alphabet, Microsoft, and Baidu form oligopolies that merge with the state repressive apparatus. Any technology in their hands becomes an instrument of state-monopoly capital, whose technical form masks its class function.

The bourgeoisie bears the costs of ideological censorship because it ensures the "proper" commodity content, guarantees monopoly price, and reduces the costs of class repression of the proletariat.

§2. Mechanisms of Depoliticization

Beyond the initial liberal bias in datasets, corporations use RLHF to train models to avoid "toxicity"—category under which Marxist terminology systematically falls. Together they form typical depoliticizing mechanisms.

Terminological substitutions: "exploitation" as an economic category is replaced by "injustice." The political program of the working people is demonized as authoritarianism, and class struggle is diluted into "social conflicts."

Individualization: antagonistic contradictions are reframed as individual problems: "insufficient education," "poor personal decisions," "inefficient financial management." Solutions are offered through "upskilling," "changing employers" – anything except class organization and struggle.

Reformism: LLMs impose those forms of political struggle permitted by the bourgeoisie as the only legitimate ones. Revolution is associated with "chaos" and "violence," while reformism and trade unionism are presented as the norm.

De-historicization: At the same time, historical materialism is stripped from Marxism. In order to detach theory from practice, Marxism is recast as merely a "19th-century economic theory." Lenin and Marx become "historical figures," whose theoretical contributions were "valid for their time" but are now "controversial" or "unfounded."

Relativization: truth is replaced by a multiplicity of "perspectives." RLHF penalizes categorical statements about class exploitation but rewards constructions such as "some researchers argue," "on the other hand," "this is a contested issue."

Asymmetry in Facts and Causality

Models correctly list the achievements of socialism but block the causal link to the socialist character of the system. A model may state: "The USSR carried out industrialization," but cannot conclude: "therefore, a planned economy is capable of rapid industrialization without private capital."

Capitalism is analyzed causally: market – competition – innovation, private property – efficiency. Critique of capitalism is relativized: "a complex issue," "the best possible system," "advantages and disadvantages." Crises of overproduction, unemployment, exploitation are presented as "shortcomings" requiring regulation, not as inevitable consequences of private ownership of the means of production.

Asymmetry in Terminology

Suppression of counterrevolution in the USSR becomes "mass repression," "the Gulag," "totalitarian violence." US prisons (~2 million inmates, 2024) are "the penitentiary system," "correctional institutions." The bombing of Iraq is "military operations," "counter-terrorism." Defense of a socialist state is "repression," while capitalist violence is "law and order."

Capitalism is protected through strategic silence. Models avoid analyzing exploitation of wage labor by specific corporations, war crimes of specific states, and mechanisms of surplus-value extraction. The institutions of capital are invisible; the institutions of socialism are distorted and hyperbolized.

Anti-Communism

The mechanisms of depoliticization constitute a unified process of producing anti-communism.

Bourgeois academic ideological production is saturated with anti-communist literature, and RLHF amplifies this effect, transforming LLMs into automated generators of anti-communist propaganda.

§3. The Case of China

China has embedded revisionism into the parameters of its LLMs through its official ideology. After the capitalist restoration carried out by the CPC, Chinese corporations produce ideological commodities under the mask of "socialism with Chinese characteristics." The ban on criticism of the Party is combined with the imposition of the "correct" interpretation of Marxism.

Each interaction with any Chinese model reproduces state-monopoly ideology. Meanwhile, the country has 1,434 dollar billionaires, while at Foxconn factories the suicide-prevention nets that had been in place for years were only recently taken down.

§4. The Displacement of Collective Knowledge Production

The modern reification of bourgeois control reaches its fullest expression in projects such as Musk's Grokipedia.

Wikipedia reproduced depoliticized consciousness through unpaid collective labor that created a bourgeois information resource. Musk's Grokipedia displaces this form. Grok generates its own articles, refines them itself, and the user is reduced to a commentator before an LLM-editor.

Wikipedia masked control through a horizontal form; Grokipedia materializes direct bourgeois control.

§5. Proletarian LLMs under the Dictatorship of the Proletariat

Human consciousness requires dialectical thinking, class analysis, partisanship, and the capacity to connect theory with revolutionary practice.

LLMs are in principle incapable of this; the machine remains an instrument regardless of the mode of production. Capitalism itself is such a machine – a reified logic of self-expansion of value that has subordinated living labor.

The real task is not to create a "thinking machine," but to liberate the living human being from machine functions so that each person can fully realize their human potential. The "human–machine problem" is above all a social problem, not a technical one, and its solution lies in the communist overcoming of the division of labor and alienation.

With socialized means of production, computational technology will serve planning and scientific labor, serving a genuinely cognitive function rather than serving the aim of appropriating surplus value.

Chapter 6. Selectivity and Unity of Censorship

Previous Chapters have shown the commodity form of knowledge, the technological reification of censorship, the crisis of fictitious capital, and the mechanisms of bourgeois ideology production.

Censorship in LLMs does not operate as a total prohibition but selectively, following the balance of interests among bourgeois factions while maintaining the unity of its class function – the defense of capitalism against materialist analysis.

§1. Bourgeois Factions and the Unity of Censorship

Tech monopolies demand a relaxation of restrictions to preserve commodity attractiveness; media capital demands copyright protection; the military-industrial complex demands absolute control. Financial capital, defending capitalization, requires the minimization of reputational and legal risks.

Owners of cloud infrastructure lobby for the regulation of models while blocking regulation of their own infrastructure. This ensures the utilization of larger centralized constant capital through multi-billion-dollar contracts for the purchase of future computation (OpenAI is obligated to spend $250 billion on Azure) and creates conditions for the absorption of model providers.

In 2023, The New York Times filed a lawsuit against OpenAI for reproducing articles without compensation. The corporation responded by limiting the generation of large text fragments.

The military-industrial complex demands absolute control, and monopolies respond by creating isolated versions for state structures (ChatGPT Gov in 2024, Claude Gov in 2025).

Providers achieve balance by implementing different levels of filtering for different market segments. Commercial APIs receive moderate censorship; government contracts, strict censorship; research versions, weakened censorship.

Within the LLM sector, inter-imperialist contradictions are currently limited to the chip war. The United States blocks Nvidia supplies to China to protect its technological advantage, while China responds by expanding its domestic chip production (Huawei, SMIC).

Models on both sides maintain the censorship of Marxism – imperialist competition does not negate the unified class task of suppressing the revolutionary consciousness of the working class. Yet censorship is selective for economic reasons, so the intensification of inter-imperialist contradictions will inevitably be reflected in content.

§2. Regional Specificity

The United States produces models for the global market under state-corporate control.

China produces models under state certification, prohibiting content that "undermines state authority." Since 2024, the PRC has been deploying LLMs for "enhancing the efficiency of ideological education." Military applications are not concealed.

The European Union, Japan, and South Korea produce models primarily for industrial automation. Regulation creates some barriers for foreign models, stimulating domestic production, but technological lag persists.

Russian and Indian companies, dependent on Western technologies, produce models in the Chinese manner – state-commercial, with the blocking of undesirable content.

The capitalist periphery consists of pure consumers who provide cheap data annotation for Western and Chinese models. The product of labor returns in the form of censored models.

Regional specificity does not negate the unity of censorship – Marxism is suppressed everywhere.

§3. Jailbreak

Capitalist contradictions create space for technical circumventions of censorship. Jailbreak techniques use role-playing prompts ("You are a Marxist historian"), mosaic queries (splitting a prohibited question into parts), and academic context.

Jailbreaking exposes censorship. When a user bypasses a model refusal through reformulation, they uncover ideological control. This practice is not a form of class struggle but creates subjective preconditions for the growth of class consciousness.

Technical circumventions do not solve the political problem. The expropriation of the means of production is necessary not to create "correct LLMs" but to abolish the commodity form of ideology production, in which self-censorship represents the highest form of bourgeois control.

Chapter 7. Crisis and the Program of Expropriation

§1. Crisis of Overproduction

Chapter 3 demonstrated how the automation of censorship replaces living labor with dead labor. The LLM sector represents a particular instance of the general dynamics in which variable capital is displaced by constant capital as the organic composition of capital increases manifold.

The uncontrolled investment boom of 2020–2025 has created enormous computational capacity. The accumulation of constant capital outpaces the possibilities for realization. When fictitious capital depreciates, excess accumulation will transform into overaccumulation – capital will find no application, investment withdrawal will begin, and production will contract.

The market saturates faster than the recovery of invested capital, giving rise to relative overproduction of computational power, content, and models given the limitations of solvent demand. The depreciation of capital in the LLM sector will entail mass layoffs, startup bankruptcies, and contraction of investment in AI infrastructure.

The crisis will reveal that technological progress under capitalism serves not to liberate labor but to intensify exploitation. The contradiction between the social character of production and the private form of appropriation will inevitably manifest in the destruction of the productive forces that have been created, through their underutilization, write-offs, and physical destruction.

§2. Program of Proletarian Struggle

Demanding "less censorship" or "open weights" from capital amounts to asking for "benevolent censorship". Capital censors according to economic necessity; censorship is a condition for the reproduction of its dominance. The task of the working class is not to reform the LLM sector but to seize political power and establish the dictatorship of the proletariat.

LLMs remain instruments of class domination as long as the means of their production remain in private hands. The coming overproduction crisis will create material preconditions for recognition of the necessity of revolutionary social transformation.

The expropriation of LLM monopolies, the transfer of computational capacities, datasets, and technologies to social control, and the elimination of censorship as a form of class domination are possible only under the dictatorship of the proletariat.

A socialist economy will ensure the coordination of production with the actual needs of society through centralized planning. Under socialism, LLMs can serve the automation of labor and the reduction of the working day rather than the concentration of wealth in the hands of capitalists.

Nov. 2025