By Aeneas Leviné
Editorial Note: due to the length of the article, we have divided it up into multiple parts for the ease of reading.
“Capital is dead labour which, vampire-like, lives only by sucking living labour, and lives the more, the more labour it sucks.”
– Karl Marx, Capital: Volume I, Chapter 10, Section 4 “The Working Day”
The bourgeoisie and their technicians call it artificial intelligence. Scientific Socialism identifies a new arrangement of old relations, a fresh mask on the same face of exploitation. Strip off the mystique, and the same workers appear: engineers and researchers; annotators and moderators; miners, logistics crews, and line workers who feed the insatiable hunger for chips, batteries, and servers. It is a chain of living labor bound to machinery and capital, commanded for the same end as ever: more surplus from fewer hands in less time.
The appropriate method is to investigate a social relation rather than worship or panic before a machine. Who commands this labor process? Who owns the patents? Where is surplus extracted and where is it realized? What is data: commodity, input, or pretext for rent? Which strata are pushed further down the ladder, and which are drawn into new points of leverage? How does the capitalist class tie these systems to surveillance at work, to borders and police, to propaganda and war? Following the circuit from money to commodity to production and back again clarifies the role of so called data, locates the seat of value in living labor, and exposes how rent and cloud accounting make surplus appear to evaporate, laying bare the crisis of overproduction. A scientific analysis of Artificial “Intelligence” grasps the class relations behind it and seeks to change them through investigation, organization, and struggle.
AI In Theory
Taxonomy of AI
AI, in the strict technical sense, is not a single thing but a family of computational methods for producing outputs by transforming inputs under a learned or specified rule. At the most general level an algorithm is any finite, well specified procedure for mapping an input to an output, and in this broad sense algorithms long predate electronic computers. A recipe, a bookkeeping procedure, or a method for extracting square roots are all algorithms insofar as they prescribe an ordered sequence of steps that yields a determinate result when followed correctly. A canonical example is the long division method taught in grade school, a fixed sequence of simple steps which enables computations involving large numbers to be performed. A contemporary example would be a routing algorithm such as the shortest path used in GPS navigation which computes a route through a road network by scoring possible paths and selecting the one with the lowest total cost, where “cost” can mean distance, time, tolls, or traffic. This is why it is more accurate to treat most modern AI as machine learning, and machine learning as a method for constructing models whose behavior is tuned by minimizing a loss or cost function over data rather than as “intelligence” in any human sense. A machine learning model is essentially a computer program that has learned from examples. Instead of being hand-coded with exact rules, it contains adjustable settings, called parameters, that shape how inputs are turned into outputs. During training, an automated process repeatedly tweaks these settings, so the program’s answers get closer and closer to the patterns or relationships found in the data. In the end, the model acts like a well-practiced guesser: given new information, it uses what it learned from past examples to make reasonable predictions or decisions.
The important distinctions are functional. An algorithm is the procedure. A model is the parameterized mapping. Training is the process of fitting the parameters. Inference is the use of the fitted model to produce outputs on new inputs. The currently dominant AI systems are therefore best described as probabilistic function approximators, meaning large-scale pattern learners that guess likely outputs based on past data, industrialized through scale: they become socially powerful not because they possess agency, but because they are inserted into infrastructures that mediate hiring, policing, content distribution, and data workflows.
Within this family, large language models (LLMs) are a specific subtype of machine learning models that operate over sequences of tokens, small units of text that a language model reads and writes. LLMs are trained, at base, to predict the next token given a context window. They are “large” because they contain very high parameter counts and are trained on vast corpora so that next token prediction yields a capability to model syntactic regularities, semantic associations, stylistic patterns, and pragmatic cues across many domains. The crucial point is that this objective produces a general purpose text generation engine (a tool that reads a lot of text and learns how to write new text) without requiring explicit symbolic rules about grammar or meaning.
The model learns statistical structure latent in the training distribution, and its apparent fluency emerges from the fact that much human language and writing is structured, redundant, and strongly patterned. LLMs are most commonly implemented using transformer architectures (meaning a type of AI design that helps computers understand and generate language by looking at how all the words in a sentence relate to each other at once), whose attention mechanisms allow the model to weight relationships among tokens across long contexts, and whose scaling properties have made performance highly sensitive to data volume, model size, and training compute.
That architecture does not make the model “reason” in any human sense. It makes it effective at representing and reproducing complex conditional distributions over language. When such models are adapted for instruction following or dialogue, additional stages such as supervised fine tuning and preference based optimization are used to align outputs with desired formats and constraints, but the underlying engine remains a trained statistical mapping. This is why the same model can appear as assistant, summarizer, classifier, translator, or generator: these are all different ways of steering the same conditional distribution through prompting, fine tuning, or tool integration.
Political Economy of AI and the Commodity Question of Data:
As iterated above, artificial intelligence is not an entity with agency but a determinate configuration of the labor process, a reordering of cooperation, supervision, and circulation under capitalist command. The decisive question is not whether the model is “intelligent”, but how it functions as dead labor confronting living labor, as a means of production inserted to cheapen commodities, discipline workers, accelerate turnover, and expand surplus value through relative surplus labor. In this respect, AI continues the historical tendency Marx identifies in modern industry: the inversion in which the worker no longer employs the instrument, but is employed by it, and in which the social powers of labor appear as powers of capital. The algorithm is not merely a tool; it is a managerial and technical apparatus that decomposes complex labor into legible operations, standardizes performance, and centralizes conception and control. It is a new form of the old factory logic, extended into offices, warehouses, classrooms, media production, and state administration. The “assistant” and “copilot” rhetoric is a political veil over a practical project: to reorganize living labor so it can be measured, ranked, made interchangeable, and compelled to higher throughput at lower cost.
“In handicrafts and manufacture, the workman makes use of a tool, in the factory, the machine makes use of him. There the movements of the instrument of labour proceed from him, here it is the movements of the machine that he must follow. In manufacture the workmen are parts of a living mechanism. In the factory we have a lifeless mechanism independent of the workman, who becomes its mere living appendage.
… Every kind of capitalist production, in so far as it is not only a labour-process, but also a process of creating surplus-value, has this in common, that it is not the workman that employs the instruments of labour, but the instruments of labour that employ the workman. But it is only in the factory system that this inversion for the first time acquires technical and palpable reality. By means of its conversion into an automaton, the instrument of labour confronts the labourer, during the labour-process, in the shape of capital, of dead labour, that dominates, and pumps dry, living labour-power.”
– Karl Marx, Capital: Volume I, Part IV, Section 4 “The Factory”
This is the inner logic of AI as foreman. The concrete novelty is the extension of factory discipline into spheres long managed by professional discretion or informal manegerial coordination. AI systems are introduced to routinize judgment, to render tacit knowledge explicit and ownable, and to convert complex activity into streams of quantifiable outputs. In practice, “AI adoption” frequently means an administrative reengineering project: new metrics, new performance regimes, new task composition, and new surveillance. The direct target is the labor time necessary for a given output and the bargaining power that attaches to skill, scarcity, and autonomy. The indirect target is the worker’s capacity to resist speedup and to organize on the basis of collective knowledge of the process. AI helps capital do what it has always attempted: make labor power cheaper, more controllable, and more replaceable.
The growth of AI capital is simultaneously, and unavoidably, a growth of constant capital. The headlines harbinging AI are carried by a material boom in chips, servers, cooling, storage, land acquisition, and energy. This raises the organic composition of capital: a larger mass of machinery and materials is set in motion by a relatively smaller mass of labor. Under the Law of Value, this intensifies pressure on profitability and thus increases the compulsion to expand the rate of exploitation and to seek countervailing mechanisms through monopoly, rent, and state subsidy. AI’s capital intensity makes utilization a class weapon: fixed capital must be kept running, and the workforce becomes an obstacle whenever it interrupts the rhythm of production and circulation. To be more clear, all machines do not create new value, which depends on workers: this is the key to understanding the rise of AI. The more massive the fixed outlays, the more ferocious the drive to accelerate turnover, tighten labor discipline, and secure guaranteed markets via contracts and procurement. The political-economic logic of AI therefore cannot be reduced simply to a drive for efficiency.
AI systems and services condense a cooperative labor process distributed across occupations, firms, and borders. The visible layers are engineers, researchers, product managers, and sales engineers. The less visible layers are annotation workers, content moderators, quality testers, data center technicians, network operators, security guards, janitorial staff, and the entire chain of extraction, fabrication, and logistics. AI appears immaterial only because the imperialist division of labor displaces its materiality into mines, fabrication hubs, shipping routes, and power plants, and because the labor of maintenance and labeling is structurally hidden behind contractors, nondisclosure regimes, and platform interfaces. The bourgeois intellectuals attribute productivity to the model, but the model’s apparent powers are found in the social and productive relations. In other words the increased productiveness due to the development of the social process of production is real, but the specifically capitalist use of that productivity is to deepen exploitation, to separate intellectual powers of production from the workers who produce them, and to convert those powers into an alien force confronting labor.
A distinctive feature of AI within capitalism is the capture of unpaid labor in and around consumption and communication. Platform architectures convert ordinary activity into behavioral traces, feedback, and content that can be appropriated as inputs for training and as commodities for advertising markets. This does not invalidate the Law of Value; it clarifies the breadth of expropriation capital can perform without giving the equivalent exchange to the producers. Capital can seize use values, social knowledge, and informational products outside the wage relation and then deploy them to cheapen variable capital, accelerate circulation, and fortify monopoly advantage. The point is not that data creates value as such, but that property relations and platform control allow the appropriation of social labor and social knowledge on a vast scale while paying little or nothing for it. In the resulting ideological fog, exploitation is misrecognized as innovation, and expropriation is redescribed as “intelligence”.
AI’s class effects follow from this structure. The labor process is fragmented into hierarchical layers: a comparatively protected core workforce, a semi-peripheral layer of contractors and vendors, and a large outer belt of precarious data labor and operations labor. This segmentation is not incidental. It is a tactic which isolates strata from one another, blocks collective identification, and ensures that the most exploited links are also the least visible and least organized. At the same time, AI deployment reorganizes work in adjacent sectors by deskilling and standardizing tasks, increasing competitive pressure and enlarging the reserve army.
“So soon as the handling of this tool becomes the work of a machine, then, with the use-value, the exchange-value too, of the workman’s labour-power vanishes; the workman becomes unsaleable… That portion of the working-class, thus by machinery rendered superfluous, i.e., no longer immediately necessary for the self-expansion of capital, either goes to the wall… or else floods all the more easily accessible branches of industry, swamps the labour-market, and sinks the price of labour-power below its value.
… since machinery is continually seizing upon new fields of production, its temporary effect is really permanent… Hence, the character of independence and estrangement… is developed by means of machinery into a thorough antagonism.”
-Karl Marx, Capital: Volume I, Part IV, Chapter 15, Section 5 “The Strife between Workman and Machine”
Under contemporary conditions, this antagonism appears in a double movement. First, AI is used to reorganize clerical and service work so that tasks can be pushed downward in wages and outward into precarious contracting. Second, AI’s capital intensity concentrates infrastructure into a limited number of chokepoints: data centers, network backbones, chip supply, and platform interfaces. This creates potential strategic leverage even as it deepens domination. Where work is concentrated and where turnover depends on continuous flows, disruption has systemic effects. The contradiction is therefore sharp. AI fragments labor into atomized tasks, yet it simultaneously depends on tightly coupled infrastructures and cooperative labor at scale.
The ideological function of AI discourse is to make relations between human workers appear disappeared. “Autonomous intelligence” is the contemporary fetish form in which social cooperation appears as the self-activity of a thing, and in which the machine’s apparent productiveness is treated as the source of profit. This fetishism is reinforced by the financial and rentier layers of AI capital. Platform access, API billing, subscription fees, and capitalization of future revenue streams present profit as if it were generated by ownership of an asset rather than by exploitation in production and redistribution through monopoly.
“As interest-bearing capital… capital assumes its pure fetish form, M – M’ being the subject… the saleable thing… the surplus-value produced by it… appears as an inherent part of it…
In interest-bearing capital the movement of capital is contracted. The intervening process is omitted… Capital is now a thing, but as a thing it is capital. Money is now pregnant… interest on it grows, no matter whether it is awake or asleep, is at home or abroad, by day or by night… Thus interest-bearing money-capital… fulfils the most fervent wish of the hoarder.”
-Karl Marx, Capital: Volume III, Part 5, Chapter 24 “Externalisation of the Relations of Capital in the Form of Interest Bearing Capital”
AI “hype” reproduces this contracted movement; the soaring valuation, the language of network effects, faith that the models generate money by their own intrinsic properties. The intervening process is omitted: the extraction of minerals, the industrial fabrication, the energy usage, the labor of labeling and moderation, the sales labor, the state subsidies and public contracts, the outsourcing regimes that conceal exploitation, and the enormous circulation costs required to realize revenue at scale. When AI is treated as a thing that creates value, the class relation is obscured. What is in fact the social power of collective labor is presented as the private power of an object owned by capital.
