When a new machine appears, debate almost inevitably begins with the machine itself. Some see it as a threat to the worker; others, as the liberation of human beings from arduous labour. Marx poses the question differently. For him, it is not enough to establish that a new technology is capable of producing faster, more cheaply and with a smaller expenditure of human labour. What must be examined is the set of relations of production into which the machine enters, who owns the means of production, and how the increased productive force is employed by capital. In the first volume of Capital, Marx draws a fundamental distinction between machinery as a productive force and its capitalist application. He states explicitly that machinery in itself is not responsible for depriving the worker of the means of subsistence. Moreover, the machine increases the quantity of the product and reduces the quantity of labour required for its production. The contradiction arises not from the technical construction of the machine, but from the social form of its application. This distinction is of decisive importance to any discussion of artificial intelligence as well, and it would be a mistake to begin the analysis by asserting that artificial intelligence is in itself the enemy of the worker, for such a conclusion would contradict Marxism. The productive force must first be separated from the social relations within which it is employed. Marx demonstrates this with particular clarity in his account of machine spinning. The Jenny, Throstle and Mule dramatically increased the production of yarn and made it considerably cheaper, yet the initial result was paradoxical: hand-loom weavers obtained an enormous quantity of cheap yarn, were able to work full time, their earnings rose for a certain period, and the industry itself began to attract new workers. Marx writes of 800,000 weavers whom the development of these spinning machines had effectively ‘called into existence’. But the next technical revolution, accompanied by the spread of the power loom, struck precisely this enormous mass of workers.
The history of machinery, therefore, cannot be reduced to the simple formula ‘the machine appears — the worker disappears’. An increase in productivity at one stage can bring about an expansion of production at another, increase the demand for raw materials, equipment and transport, give rise to new occupations and temporarily increase the number of people employed. Marx demonstrates directly that the spread of machinery increased the production of means of production, expanded the extractive industries and created a new category of workers — the producers of the machines themselves. But it is precisely here that the essential question begins. The increased productive force does not develop in a social vacuum; it is incorporated into the process of the production of capital. What must therefore be examined is not merely how much labour the machine saves, but what capital does with this saving of labour.
Herein lies the apparent contradiction that Marx exposes with particular force. The machine in itself reduces the labour-time required for production, yet its capitalist application can lead to a lengthening of the working day; the machine in itself lightens labour, whereas its capitalist application, on the contrary, intensifies it; the machine in itself increases the productive force of labour and the mass of use-values produced, yet under capitalist production the result of this increase assumes the form of the result of a production process belonging to capital. Consequently, the question of ownership cannot be understood here in merely juridical terms — as the question of who formally owns a particular machine, server or piece of software — but must be examined as a relation: Who has command over the means of production? Who combines them with labour-power? Who determines the purpose of production? And to whom, ultimately, does the surplus value produced belong? It is precisely for this reason that the capitalist application of machinery is capable of transforming a technical means for the liberation of labour into a means for strengthening the power of capital over labour. Marx demonstrates this directly: when machinery made muscular strength less necessary, capital used this technical possibility not simply to lighten existing labour, but to enlarge the mass of labour-power available to it by drawing women and children under the direct domination of capital. Marx therefore calls the labour of women and children ‘the first word of the capitalist application of machinery’.
Here lies the starting point for the analysis of artificial intelligence.
AI likewise represents a development of society’s productive force: it is capable of reducing the time required to process information, taking over some of the operations previously performed by human intellectual labour, automating repetitive operations and sharply increasing the volume of work that can be performed within a given period of time. But none of this, in itself, tells us anything about its social outcome. To understand that outcome, we must pose the same question that Marx posed in relation to the machine production of the nineteenth century: in whose hands does this new productive force lie, and within what relations of production is it employed? If an artificial intelligence system is under the control of private capital, then a reduction in the quantity of labour technically required to perform a particular task does not in itself mean a reduction in the worker’s working time. An increase in productivity does not in itself mean that the additional wealth produced as a result will pass to the worker, while the disappearance of some existing functions does not mean the simple and simultaneous disappearance of labour as such. As Marx’s account of the Jenny demonstrates, a new productive force may initially expand entire fields of activity and create new categories of workers, only for a subsequent stage of the same technical revolution to alter their position once again. It would therefore be just as mistaken to declare artificial intelligence the cause of humanity’s future liberation as it would be to declare it the cause of future unemployment.
The machine does not establish relations of ownership.
The machine does not appropriate surplus value.
The machine does not determine the length of the working day.
The machine does not decide who receives the fruits of increased productivity.
All of this is determined by the social relations within which it operates.
The modern computing system did not arise alongside the earlier system of machinery as something fundamentally separate from it; it grew out of that system through a succession of technical revolutions, in the course of which mechanical control gave way to electrical and electronic control, the individual machine was incorporated into increasingly large automated complexes, production became integrated with computation and the transmission of information, and information processing itself became a direct element of the production process.
Artificial intelligence represents a further stage in this development, since the machine acquires the capacity to perform not only a predetermined sequence of mechanical operations, but also some of the operations that previously required the direct intellectual labour of human beings: analysing large volumes of information, recognising patterns, generating text and images, managing processes, comparing alternatives and producing results within a period of time incomparable with the expenditure of human labour required to perform the same mass of operations.
The growth of productive force, however, simultaneously entails a growth in the material conditions required for its application. The more operations are transferred to the computing system, the more complex the models become and the wider the field of their application, the greater the requirement for computing power, specialised processors, data centres, electricity, networks and supporting infrastructure; consequently, the technical development of the machine is accompanied not by the disappearance of its material foundation, but by its expansion and concentration.
Here arises the contradiction between the increasingly social character of production and the private ownership of the material means by which it is carried out. An algorithm may be developed by a small group of researchers, software code may be distributed freely, and a scientific discovery may become the common possession of millions of people; yet the ability to transform accumulated social knowledge into a continuously functioning productive force is determined by access to computing power, specialised processors, energy and network infrastructure, whose creation on a large scale already requires a considerably greater mass of capital than the emergence of an individual idea itself.
The openness of a technology, therefore, does not yet imply equality of economic opportunity. An enterprise may possess its own model and its own product while remaining dependent upon the owner of the computing infrastructure, since, in order to train, refine and maintain its system, it is compelled to purchase computing power as a condition of its own production. The result is a structure in which numerous capitals retain their formal independence and compete with one another, while the material foundation upon which that competition takes place belongs to a considerably smaller number of owners. The more the downstream market develops, the more computing power it consumes; the more computing power it consumes, the more resources the owner of the infrastructure obtains for its expansion; the larger the infrastructure becomes, the greater the capital required to create a comparable system, and consequently the development of competition at one level may simultaneously intensify concentration at another.
Small capital, therefore, is not necessarily destroyed directly by large capital. It may continue to exist, grow and even compete successfully while remaining dependent upon the owners of the infrastructure to which access constitutes a necessary condition of its own production, since a portion of the capital it advances continually assumes the form of payments for computing power and other necessary means of production, while the conditions of access to those means are determined by their owners. In such circumstances, the existence of thousands of independent enterprises does not disprove the concentration of capital: economic multiplicity may persist on the surface of the market simultaneously with the concentration of the material conditions of production in the hands of a considerably smaller number of capitals.
It is precisely for this reason that the concentration of computing resources has significance extending far beyond the technology industry itself. As computing systems penetrate industry, transport, finance, communications, medicine, scientific research and public administration, control over computing infrastructure gradually becomes control over one of the general conditions governing the operation of numerous other industries, with the result that the owner of such infrastructure occupies not merely the position of one producer among others in the market, but that of a supplier of a means of production to other capitals.
At this stage, the question of artificial intelligence finally ceases to be a question of a particular program or an individual machine: the technical revolution leads to an increase in productive force; the growth of productive force requires an expansion of its material foundation; the expansion of that foundation increases the mass of capital required for its large-scale application and, under certain conditions, accelerates the concentration and centralisation of capital, with the result that control over the means of computation is capable of becoming economic power over the conditions of production of other capitals.
By the middle of 2025, the development of artificial intelligence in the United States had already ceased to appear exclusively as the affair of individual technology companies, since the scale attained by computational production had linked the further expansion of this industry not only to the accumulation of individual capitals, the acquisition of equipment and the hiring of labour-power, but also to such general conditions of production as the energy system, semiconductor production, transport and network infrastructure, the training of labour-power, public procurement and the foreign-trade regime, none of which lies directly at the disposal of the individual capitalist.
In July 2025, this relationship received its fully developed political expression in America’s AI Action Plan, in which the development of artificial intelligence was linked to the economic competitiveness, national security and international position of the United States, while the country’s technological superiority was associated with economic advantages, military power and the ability to influence the formation of global standards; Eastern Post has already examined the substance of this programme in detail in a separate article — America’s AI Action Plan 2025 — and what matters here, therefore, is not to repeat the measures it provides for, but to uncover the social relation that stands behind them.
An individual capital commands the means of production belonging to it, purchases labour-power, obtains credit, expands the enterprise and enters into competition with other capitals; yet even the largest private owner does not possess state power as his private property: he cannot, by his own decision, alter legislation, reconstruct the national energy system, determine the regime governing semiconductor exports, allocate federal funds, establish the rules of public procurement or, through administrative power, alter the conditions under which entire territories operate. But it by no means follows from this that the economic power of capital exists separately from political power and only at some later stage of development turns to it for assistance; on the contrary, capitalist production exists from the outset within a definite social and political order that safeguards the relations of property corresponding to it, while the development and concentration of capital increase the scale of those conditions of its reproduction that are organised no longer by the individual enterprise, but by the state.
It is precisely for this reason that the relationship between large technology capital and the state can be reduced neither to the transformation of the corporation into the state nor to the transformation of the state into the executive committee of any single corporation. Individual capitals retain their independence, compete with one another, demand different rules and seek advantages over one another; state power, in turn, is capable of restricting one capital, supporting another, altering the conditions of competition and pursuing objectives that do not coincide with the immediate interest of any individual enterprise, yet all of this takes place within a social order in which the principal means of production remain capital and production is carried on as the production and accumulation of capital.
America’s AI Action Plan reveals this connection with particular clarity, since private companies remain the direct owners of a substantial part of the means of production and the commercial agents of the technology’s development, while the state organises a significant part of the general conditions necessary for the industry’s further development: energy and computing infrastructure, the permitting regime for the construction of data centres, support for the production of technological components, the training of labour-power, public procurement, scientific research, the military application of the technology and the foreign-economic conditions governing the expansion of the American technological system.
State intervention here, therefore, does not stand in any simple contradiction with the proclaimed freedom of private initiative. The freedom of capital has never meant the absence of social and state conditions for its existence; on the contrary, private control over the means of production presupposes an entire system of legal, administrative and material conditions without which it could not reproduce itself. When existing state restrictions begin to obstruct the expansion of a new industry, those restrictions themselves become the object of change: America’s AI Action Plan proposes that federal agencies identify and review regulations, administrative decisions and guidance documents regarded as unjustifiably impeding the development and deployment of artificial intelligence, so that state regulation appears here not only as a limit upon the movement of capital, but simultaneously as a means of altering that limit.
The same connection appears in the relationship between the federal government and the states, since, in the allocation of certain federal funds, the regulatory climate of a particular state and the possibility of restricting funding where existing rules are regarded as an obstacle to federal artificial-intelligence policy are to be taken into account; the formal legislative competence of the state does not thereby disappear, yet the material conditions under which that competence is exercised are brought into connection with the centralised financial power of the federal state.
Deregulation in this case, therefore, means not the withdrawal of the state from the economy, but a particular direction of its activity: in order to remove a state-imposed restriction, the state itself must decide to alter the rule, the permitting procedure, the conditions of financing or the administrative practice. What appears on the surface as the liberation of private initiative from the state is accomplished through state power; the boundary is moved not because that power disappears, but because the power itself moves the boundary.
Nor is it necessary to assume either a secret agreement between the government and the heads of technology companies or a complete coincidence of interests among different corporations. On the contrary, competition among capitals remains a necessary relation of capitalist production: each individual capital seeks to expand its own share of the market, secure more advantageous access to energy, computing power, public procurement and credit, displace its rival or subordinate that rival to itself, while competition itself, as large-scale production develops, is capable of accelerating the concentration and centralisation of capital. It is precisely for this reason that Marx distinguished accumulation, concentration and centralisation: the growth of capitals already functioning may take place simultaneously with the formation of new capitals, whereas competition and credit create a particular mechanism for drawing together already existing capitals and transforming many capitals into a smaller number of larger ones.
Competition among individual technology companies, therefore, does not alter the fact that they all operate within a common material structure whose reproduction requires energy, a scientific base, an education system, finance, infrastructure, legislation and foreign trade; they may struggle fiercely over the distribution of these conditions among themselves while simultaneously sharing an interest in the expansion of the conditions themselves, since without an expansion of the computational, energy and productive base, the accumulation of each of them ultimately encounters the limits of the material foundation upon which that accumulation takes place.
Here, then, we discover the real content of the apparent contradiction between the demand for the freedom of private capital and the demand for state support for its further development: the command over capital and the appropriation of the results of its movement remain private, while an ever greater part of the material and organisational prerequisites of that movement acquires a social scale. Energy capacity, the training of specialists, fundamental science, infrastructure, public procurement, budgetary financing and the administrative apparatus are drawn into the creation of the general conditions for the industry’s expansion, yet this socialisation of the conditions of production does not in itself abolish the private form of capital; on the contrary, a socially organised condition can continue to function as a condition of private accumulation.
It would therefore be a mistake to regard the expansion of state participation as an automatic negation of private capital. The increasing social scale of the organisation of production does not in itself signify a change in the relation of production: Marx shows that capitalist development itself concentrates the means of production and transforms production into an increasingly social process without thereby abolishing the capitalist form of appropriation; indeed, it is precisely the large scale of production that creates the technical and social prerequisites for enterprises whose realisation is bound up with enormous masses of capital and with the acceleration of its centralisation.
But the movement of capital does not end within national borders, since accumulated capital must continue its movement, while the expansion of production also expands the need for markets, raw materials, means of production and spheres for the application of capital. The integration of the development of artificial intelligence with foreign-trade and international policy therefore represents not an external addition to domestic industrial policy, but a continuation of the same movement on the world market: American models, computing infrastructure, standards and the technologies associated with them begin to appear not only as commodities produced by competing enterprises, but also as elements of a national technological system whose expansion beyond the country’s borders is supported by the trade, diplomatic and other instruments of the state.
It is precisely here that the interest of large capital most readily assumes the outward form of the general national interest: what, for individual enterprises, means the expansion of the market, the enlargement of the sphere in which their capital can be employed and the strengthening of their position in international competition appears, in political form, as the task of securing the country’s competitiveness, national security and technological superiority. This transformation of a private economic interest into a politically formulated ‘general’ interest does not abolish competition among capitals themselves; it merely demonstrates that their struggle takes place within a broader class and state relation. It was precisely in relation to such material that Lenin observed how the interests of large-scale industry begin to be identified with the interests of the ‘country’, after which the economic movement of capital finds its continuation in the foreign policy of the state.
Herein lies the essential content of the transition that has taken place: the development and concentration of technology capital do not abolish the distinction between the corporation and the state, but make their belonging to a single system of social relations increasingly apparent, within which private capitals continue to compete with one another while state power organises, safeguards and alters the general conditions of their reproduction; and the more the further development of artificial intelligence comes to depend upon the energy system, science, infrastructure, the public budget, state procurement and the world market, the more evident it becomes that this new productive force has long since passed beyond the confines of the individual enterprise, even though the means of its large-scale application continue to function as capital.
America’s AI Action Plan therefore represents considerably more than a programme for the technical development of a new industry: private ownership of concentrated computing resources is preserved, socially organised material conditions are drawn ever more deeply into securing their further expansion, while the state apparatus, through legislation, the public budget, procurement, infrastructure policy and foreign-trade policy, organises a substantial part of the legal, material and international conditions within which the further movement of technology capital takes place.
But from this arises another question, one that cannot be derived directly from the ownership of servers, semiconductors and data centres, although it develops within the same social relation: state power not only organises the material conditions for the application of the new productive force, but, through legislation, administrative decisions and public procurement, also establishes certain limits upon its permissible social use, with the result that the dispute shifts from the question of who commands the machine to the question of which functions of that machine are permitted or prohibited, what information may be provided to an individual, what restrictions must be built into the system, which forms of its application are deemed dangerous and, finally, who possesses the power to establish these boundaries.
A characteristic manifestation of this struggle can already be seen in Britain, where, as the BBC reported on 14 September 2026, the Joint Committee on Human Rights concluded that existing legislation was inadequate to address the risks arising from the development of artificial intelligence and proposed the creation of a single independent supervisory body, stricter requirements for the most powerful and ‘high-risk’ systems at different stages of their life cycle, and a complete prohibition of certain applications of the technology, including particular forms of ‘covert influence’, profiling and the use of biometric data.
The category of ‘high risk’, however, explains nothing in itself until it has been established what particular social relation lies behind this formula. The machine does not declare itself dangerous, nor does it draw the boundaries of its own permissible application; such boundaries are established by people and institutions through law and state power, and while the technical characteristics of a system may determine its power, speed, capabilities or the probability of a particular outcome, no political decision as to which capabilities should be permitted, to whom they should be available and under what conditions their use should be restricted can be derived from those technical characteristics themselves. Here the technical measure ends and the social relation begins.
It is precisely for this reason that the appeal to privacy reveals a broader contradiction in state policy, since a state capable of restricting certain civilian applications of artificial intelligence on the grounds of the collection, processing or use of personal data itself possesses apparatuses for the systematic acquisition and analysis of information, which arose long before the present debate over ‘high-risk’ models and which, with the development of computing technology, are acquiring ever more powerful technical means.
The American experience is particularly revealing in this respect: the PRISM programme, disclosed in 2013, formed part of a broader system of intelligence collection under Section 702, through which US intelligence agencies obtained certain categories of electronic communications from electronic communications service providers, while Section 702 itself remains in force; consequently, the collection and machine processing of large volumes of information do not appear to state power as something unconditionally impermissible in themselves, since the permissibility of a particular application is also determined by the actor involved, its purpose and the legal regime established for it.
The distinction becomes still clearer where the computing system enters directly into the military and intelligence apparatus, since in this case the state does not restrict the development of the machine merely because of the power of its computational capabilities but, on the contrary, is capable of financing their further expansion and combining state information resources with the technologies of private corporations. The United States Department of Defense has awarded Palantir major contracts connected with the Maven Smart System, within which algorithmic tools are applied to sensor and other data to build an operational picture, detect objects and support processes associated with the identification and confirmation of targets.
The simple opposition of ‘dangerous artificial intelligence’ to a state supposedly standing outside social relations and merely protecting society from a technical threat is therefore insufficient to explain what is taking place. Lenin criticised precisely this conception of the state as a force standing outside classes and capable, at its own discretion, of choosing between opposing social interests; political arrangements, he argued, are rooted in economic relations, express them and serve them. The same computational capability can therefore acquire a different legal and political significance depending upon the social relation into which it is incorporated, who commands it and the functions for which it is employed.
If, therefore, the expansion of the capabilities of artificial intelligence in one sphere becomes subject to restriction as ‘high risk’, while the development of the computational capabilities of the military or intelligence apparatus is simultaneously financed and expanded on a state scale, what confronts us is no longer a technical question of whether the machine is dangerous in general, but a question concerning the different social functions of the same productive force: why is a particular capability restricted within one relation and deemed necessary within another, who draws this boundary, and through what power does it acquire binding force?
The answer cannot be derived from the technical characteristics of the machine, since the same computational capability acquires a different social significance depending upon the relation within which it is employed; therefore, before accepting the classification of ‘high risk’ as a self-evident explanation, it is necessary to establish what specific harm this category denotes, for whom the risk arises, who determines its measure, and why a particular capability of the new productive force is restricted in one case while in another its development is supported by state resources.
It is particularly revealing that the question of the limits to the further development of artificial intelligence is now being raised not only by parliamentary and state institutions. According to the material cited by the BBC, Anthropic chief executive Dario Amodei has called for the pace of development of the most powerful models to be slowed, for independent oversight during their development, for industry-wide regulation and for international co-ordination, while OpenAI chief executive Sam Altman and Elon Musk have publicly supported a substantial part of the proposed approach; on the surface, a paradoxical situation thus arises in which representatives of capitals directly engaged in the development of the new productive force themselves demand that certain limits be imposed upon its further development.
But the contradiction arises only if capital is assumed to have the development of the productive forces as such for its objective. Capital has no such objective: for it, production is simultaneously a process of producing a product and a process of the self-expansion of value, and already in the first volume Marx demonstrates this dual character directly — the social labour process is subordinated to the movement of capital and to its own ends. In the third volume this connection appears still more clearly: the purpose of capitalist production is the expansion of the value of capital, the appropriation of surplus labour, and the production of surplus value and profit.
Capital is therefore interested in the development of a new productive force not unconditionally, but only insofar as that development can function as a movement of capital; new technology receives an extraordinarily powerful impetus precisely because it is capable of reducing individual costs, giving an individual capital a competitive advantage, opening new spheres for the application of capital and serving further accumulation, yet the same logic means that technically possible development need not be pursued immediately and without restriction if the concrete conditions of its application threaten profit, property, the possibility of realisation, the position of capital in competition or the stability of the conditions of its further reproduction. In Marx, the expansion of production is determined not by society’s abstract need for the maximum development of the productive forces, but by the conditions governing the production and realisation of profit; capitalist production comes to a halt not where technical possibilities have been exhausted, but where further movement ceases to satisfy the requirements of the self-expansion of capital.
The demand to regulate or even slow particular directions in the development of artificial intelligence does not, therefore, in itself constitute a negation of the interests of capital, still less does it transform the private owner of the new productive force into an opponent of capitalist development. The question is not whether capital demands the maximum possible speed of technical development under all circumstances, but under what conditions the new productive force can continue to function as capital — that is, preserve private command over the means of its application, secure the process of the self-expansion of value, and reproduce the social relation within which the technical power of the machine appears as the power of capital belonging to its owner.
Hence the need to distinguish between the danger that the application of a technology may pose to an individual and the consequences that the development of a new productive force may have for existing social relations, since these phenomena may intersect but are not identical. A violation of privacy, the unlawful use of biometric data, a discriminatory decision or the deprivation of procedural safeguards constitute concrete harms that may become the object of social protection; an entirely different matter is a change whereby a new productive force reduces the quantity of labour required to perform particular intellectual operations, eliminates the need for certain previously existing intermediary functions, or gives a substantially wider circle of people access to knowledge and means of activity that previously required considerable financial expenditure or membership of a specialised professional institution, since in such a case what changes is no longer merely the technical conditions under which particular operations are performed, but the existing division of labour, the position of particular professions, the distribution of income, and established relations of economic and political power.
The participation of the heads of the largest technology companies in determining the future rules of the industry therefore acquires fundamental significance, since, while continuing to compete with one another for markets, computing power, capital and technological superiority, they simultaneously act as representatives of capitals whose existence presupposes the preservation of private command over the means of production and the possibility of transforming the new productive force into a means of further accumulation. It does not follow from this that different corporations must have an identical interest in any particular form of regulation: one may demand acceleration, another restriction, and a third rules that give it an advantage over its competitors; yet the very intensity of this struggle does not abolish the common system of property relations within which it takes place, since the competing enterprises remain independent capitals and confront one another precisely as capitals.
It is precisely here that the high-sounding categories of privacy, safety and high risk cease to provide a sufficient explanation of what is taking place, since public safety cannot be determined by the label that a state, an expert or a corporation has attached to a particular measure, but requires answers to material questions: what exactly has been identified as a danger, who is threatened by the supposed harm, who determines its permissible measure, who retains access to the capability being restricted, and how the restriction introduced alters the actual distribution of economic and political power. The BBC material itself does not allow us to conclude that the British parliamentary committee or the heads of technology corporations are consciously seeking to slow artificial intelligence specifically in order to preserve their own dominance, since they state no such motive and the available facts do not directly establish one; political-economic analysis, however, has no need to attribute a secret intention to them, since the objective contradiction exists independently of the subjective intentions of its participants: the conditions governing the permissible application and further development of the new productive force are determined by state institutions and the largest capitals, which themselves operate within the existing relations of property and power.
The question must therefore be posed differently: not only what risks artificial intelligence creates for society, but also how existing relations of property, the division of labour and power affect the establishment of the boundaries of its development and application, who acquires the ability to present the crossing of a particular boundary as a danger to society, and who possesses the power to decide beyond what point the further development or a particular application of the machine must be slowed, restricted or prohibited.
The dispute over ‘high-risk artificial intelligence’ then ceases to be exclusively a dispute among engineers about safety, since behind the technical classification there emerges a political question of considerably greater magnitude — who possesses the right to determine the socially permissible extent of the development of the new productive force.
The right to determine the socially permissible extent of the development of artificial intelligence, however, exists not only in the form of legislation prohibiting a particular application of the technology, since public procurement proves to be a considerably more effective instrument: the state has no need to prohibit a corporation from producing a particular model if it is capable of establishing the conditions that the model must satisfy in order to gain access to the enormous government market.
It was precisely such a mechanism that was incorporated into America’s AI Action Plan in the summer of 2025. The document called for federal procurement rules to be amended so that the government would enter into contracts only with developers of frontier large language models that ensure the objectivity of their systems and the absence of top-down ‘ideological bias’; at the same time, the National Institute of Standards and Technology was instructed to revise the artificial-intelligence risk-management framework and remove references to misinformation, Diversity, Equity and Inclusion, and climate change.
In this case, however, the matter did not remain at the level of political declaration. On the same day, 23 July 2025, the President of the United States signed a separate Executive Order, Preventing Woke AI in the Federal Government, establishing two principles for large language models procured by the federal government — Truth-seeking and Ideological Neutrality. According to the order, a model must strive for factual truth, historical accuracy, scientific inquiry and ‘objectivity’, acknowledging uncertainty where reliable information is incomplete or contradictory, and must not manipulate its answers in favour of political or ideological dogmas. An excellent requirement; only one small matter remains — to establish who determines where scientific inquiry ends and dogma begins, what degree of historical accuracy is sufficiently accurate for a government contract, and at precisely what point a political judgement ceases to be political and acquires the considerably more respectable name of ‘objectivity’.
It is precisely here that the question acquires material content, since the state does not confine itself to expressing the wish that the machine should be truthful: the conception of objectivity formulated by state power is translated into procurement rules, then into the terms of a federal contract, and compliance with those terms becomes one of the requirements governing a private developer’s access to the government market. A small administrative miracle thus takes place: a philosophical category passes through the procurement process and emerges from it as a contractual obligation.
The value of a contract can be expressed in dollars, computational performance in technical metrics, and information security in standards and tests; ‘ideological bias’ is considerably less amenable to measurement, since it cannot be identified in a model until the norm against which an answer is to be judged biased has first been established. What therefore becomes decisive is no longer the word ‘neutrality’, but the material question of who possesses the power to determine its measure.
And here we encounter a distinction that the elegant word ‘objectivity’ tends to conceal rather than explain. Scientific inquiry proceeds not from whether the state likes or dislikes the result obtained, but from actual relations, their emergence, movement and contradictions; it is under no obligation to recognise the existing order as natural merely because that order exists. It was precisely for this reason that Marx treated the capitalist mode of production and the relations of production and exchange corresponding to it as a historically determined object of inquiry, requiring independent thought rather than the adaptation of conclusions to the prejudices of public opinion.
This gives rise to a considerably more serious problem than the dispute surrounding DEI, which the order itself cites as an example of ideological dogma. If an examination of property, wage labour, accumulation, the state or technology capital itself leads to a conclusion inconvenient to the existing order, does that make the inquiry incorrect — or does the question of its ‘ideological bias’ arise only from the moment when the conclusion reached proves inconvenient to whoever establishes the criterion of neutrality? Marx, in his time, formulated this metamorphosis without diplomatic embellishment: once the class struggle had openly declared itself, the question for bourgeois political economy was no longer whether a theory was correct or incorrect, but whether it was useful or harmful, convenient or inconvenient, and whether it accorded with ‘police considerations’.
It is therefore particularly striking that the same state power which requires artificial intelligence to be free from political and ideological dogmas officially declares communism a threat to the country, while the President contrasts loyalty to Karl Marx with loyalty to America. The point here is not to replace one state-approved dogma with another: Marxist analysis interests us not as an article of faith, nor as a set of conclusions to be accepted irrespective of the facts, but as an investigation of actual social relations, their historical emergence, class antagonisms and movement; its conclusions must therefore be tested against the social process under examination itself, rather than against how convenient those conclusions may be for the administration in power.
And at this point ‘ideological neutrality’ appears in a somewhat unexpected light. If the scientific method arrives at an approved conclusion, it remains objective; if the same method reveals the historical character of existing relations of property and power, the temptation arises to discover ‘ideological bias’ within it. Nothing supernatural is taking place here, of course: the political views of an opponent continue to be called political views, critical analysis of the existing order risks acquiring the name of ideology, while the state’s own premises undergo a considerably more fortunate transformation — they disappear as politics and return as the neutral measure of objectivity. But the significance of this mechanism becomes clear only when the scale of the government market upon which this principle is to operate is placed alongside the legal text.
By August 2025, the General Services Administration had already incorporated Claude, Gemini and ChatGPT into federal procurement mechanisms, after which special OneGov agreements were concluded: OpenAI offered ChatGPT Enterprise to participating federal agencies for a symbolic one dollar during the first year, Anthropic offered Claude to federal bodies across the executive, legislative and judicial branches for one dollar as well, while Google provided Gemini for Government at a price of 47 cents per agency for the duration of the initial offer.
Naturally, if one looked solely at the invoice, one might conclude that the largest technology companies had unexpectedly inaugurated an era of almost free artificial intelligence for the American state. Yet the sequence of prices — one dollar, another dollar and 47 cents — demonstrates that what confronts us is no longer an ordinary struggle for immediate revenue from the initial sale. The difference between one dollar and 47 cents is utterly negligible for the federal budget and just as obviously tells us nothing about the actual economic scale of the models, computing infrastructure and engineering organisation required for their operation; what becomes the object of competition is access to a considerably larger sphere for the application of the technology.
The state consolidates the previously fragmented demand of numerous agencies, confronts several competing suppliers with its own centralised purchasing power and thereby acquires the ability to reduce the initial price of access dramatically, while the supplier may agree virtually to forgo immediate revenue during the initial period because an exceptionally cheap point of entry facilitates the diffusion of its system among an enormous number of potential users. Forty-seven cents therefore cannot be taken as the actual economic measure of Gemini, just as one dollar is not the economic measure of the entire technological system of OpenAI or Anthropic: what we have before us is the price of one moment within a considerably longer-term relationship.
The economic result of the first stage is real for both sides, although it has a different content for each of them. By centralising demand, the state acquires the ability to set competing suppliers against one another and reduce its own expenditure; the corporation, by contrast, accepts an exceptionally low initial price in order to expand the sphere in which its own product is used. But the expansion of this sphere is not yet profit and does not in itself create new value: it merely creates broader conditions within which the subsequent use of the technology may acquire an economically significant scale.
It is precisely here that one dollar and 47 cents cease to be curious price tags and become the starting point of further movement. The cheaper the initial access, the easier mass adoption becomes; the broader the adoption, the greater the number of users, administrative processes and technical connections that develop around the system; and the more deeply the system enters into the everyday activity of the state apparatus, the less significant the initial licence price becomes and the greater the significance acquired by the volume of subsequent use. The state gains from the entry price, while the supplier acquires an expanded sphere for the potential realisation of its services; this possibility, however, must still be transformed into actual effective demand.
By September 2026, the scale of this sphere had already become visible: the General Services Administration reported that its artificial-intelligence agreements had expanded access to the relevant systems to approximately 3.5 million federal employees, while the savings calculated by the government under its artificial-intelligence agreements had reached approximately $1.4 billion; at the same time, a new agreement with OpenAI no longer provides for the symbolic one-dollar price of the initial period, but for a 27-month arrangement under which model usage is charged at a 50 per cent discount, with purchasing eligibility extending not only to federal bodies across all three branches of government, but also to state, local and tribal governments.
Before turning these 3.5 million people into the basis of any calculation, however, it is necessary to define precisely what the figure itself represents: it is not the number of artificial-intelligence users in the United States, not the American audience of OpenAI, Anthropic or Google, and not even the final size of the government market for these systems, but a considerably narrower category — the approximate number of federal employees to whom, according to GSA data, the relevant agreements have already opened access to artificial-intelligence systems; outside this figure remain state and local government bodies, tribal governments, private enterprises, educational and other institutions, individual users and the entire overseas market, while some of these spheres do not fall within the government contracts considered here at all, and to add them to a single calculation would therefore mean transforming the technical possibility of consumption into effective demand before the consumption itself had even taken place.
We shall therefore proceed in precisely the opposite direction: instead of first presenting the entire potential audience for artificial intelligence and then declaring it a market, we shall confine ourselves to the one already documented federal sphere and ask what monetary magnitude it might assume at a given level of actual usage; the 3.5 million thus become not an estimate of the American market, but the starting boundary of our hypothetical calculation.
If, for ease of calculation, we assume average consumption of artificial-intelligence services by one active user at $100 per month — understanding this figure not as the price of a single subscription, but as a hypothetical aggregate cost of consuming models, computing resources and additional functions — and further assume that all 3.5 million federal employees who have already been granted access became active users with this average level of consumption, the potential volume of paid services would amount to:
3,500,000 × $100 × 12 months = $4.2 billion per year;
if the same conditions were maintained for five years:
$4.2 billion × 5 = $21 billion.
If, however, only half of the identified federal audience — 1.75 million people — were active users, the corresponding figure would be:
1,750,000 × $100 × 12 months = $2.1 billion per year, or $10.5 billion over five years.
Thus, to arrive at a figure of approximately $4.2 billion annually, there is no need whatsoever to add the private American market, the entire population of the United States or an overseas audience to the calculation: a scale of this magnitude already arises within a single documented federal sphere under the assumptions we have adopted, while state, local and tribal governments, to which the new procurement mechanism also extends, remain outside the calculation; the $4.2 billion figure is therefore neither an estimate of the entire American artificial-intelligence market nor a forecast of government expenditure, but the arithmetical expression of the potential scale of paid services within one defined sphere of application.
But it is precisely here that we must stop at the point where contemporary financial journalism usually only begins to gather speed: neither $4.2 billion, and still less $21 billion, can be called the profit of the technology corporations, since these figures represent only a hypothetical volume of paid services under the assumptions we have adopted; from the revenue received, the costs of computing infrastructure, electricity, cooling, networks, data centres, maintenance, security, the development and updating of models, labour-power and other elements of production must be replaced, while the mere fact that 3.5 million employees have been granted access by no means implies that all of them will become regular consumers and generate effective demand on the scale we have assumed.
The significance of the calculation, therefore, lies not in awarding technology corporations in advance a profit they have not yet received, but in revealing the distinction between the price of initial entry and the possible scale of the subsequent relationship: the state obtains virtually free initial access and the savings it has calculated, while the supplier gains access to a multimillion-user sphere of potentially recurring consumption; one dollar does not turn into billions — it is not the dollar that changes, but the scale of the relationship in which it initially appeared.
The expansion of this relationship, however, still tells us nothing in itself about the origin of profit, since computing infrastructure does not begin to create new value merely because a greater number of requests have passed through it: servers, accelerators, buildings and other means of labour that function over prolonged periods transfer their value to the services produced with their aid gradually, in proportion to their use and wear; electricity and other consumed elements must continually be replaced as they are used up; labour-power, by contrast — that of engineers, programmers and other personnel — occupies a fundamentally different position in the process of creating new value. Fixed and circulating capital must therefore not be confused with constant and variable capital: the first distinction is determined by the manner in which advanced value turns over and is transferred, the second by the different roles played by the means of production and labour-power in the process of valorisation.
For artificial intelligence, this distinction is particularly important, since a server is not purchased anew for every response produced by a model, nor is an already constructed computing infrastructure reproduced from scratch for every new user: its value is recovered gradually through the mass of services produced by means of it, while the infrastructure itself continues to function throughout a multitude of successive operations; the economic advantage of scale therefore lies not in any mysterious capacity of the machine to produce value independently, but in the possibility of using already advanced means of labour to produce an enormous mass of services.
But here, too, capital has not been presented with an eternal server: the means of labour undergoes physical wear, while in a high-technology industry physical deterioration is accompanied by even more rapid moral depreciation, since a new generation of accelerators, a more powerful architecture or a more efficient method of computation may render existing equipment economically obsolete long before it physically ceases to function; the same science and technology that increase the productive capacity of capital already in operation are simultaneously capable of devaluing its existing material forms and compelling its owner once again to advance funds for their replacement.
The profitability of such a system is therefore determined neither by the value of an individual server nor by the price of an individual subscription, but by the movement of the totality of relations — the magnitude of the capital advanced, the mass of services produced and realised, the costs of replacing consumed elements, the depreciation of the means of labour, expenditure on labour-power, the rate of turnover, the actual selling price and the scale of effective demand; it is precisely here that it becomes clear why one government employee ceases to mean one subscription, since, as artificial intelligence is integrated into everyday activity, that employee’s work may be broken down, from the supplier’s standpoint, into a succession of separate computational services — document analysis, information retrieval, text generation, programming, access to a specialised model, processing of datasets, the operation of an agent, the storage and retrieval of information, and the use of additional computing power.
The state thus begins to purchase not merely a finished program, but the ability continually to pass individual operations of its employees’ intellectual labour through computing infrastructure owned by private capital; it is precisely for this reason that our hypothetical figure of $100 per employee represents not the price of a single licence, but an average monetary magnitude for the possible aggregate of services consumed, while an increase in the number of functions, the complexity of operations and the intensity of use may increase the mass of paid consumption without any corresponding increase in the number of employees.
The economically significant expansion of the market therefore begins not with the mere formal granting of access to ChatGPT, Claude or Gemini to a government official, but as documents, searches, analyses, program code, agents and data processing begin regularly to pass through a private computing system and the corresponding operations assume the form of services continuously purchased by the state; yet even the technical capacity to produce such a mass of services does not guarantee their realisation, since effective demand still stands between the productive capacity of the system and the revenue received — the state must actually purchase and pay for the corresponding volume of services.
Capital therefore does not obtain surplus value from circulation merely because it has managed to issue more invoices: through sale, the value produced assumes its money form and the surplus value contained in the commodity is realised, whereas competition itself, or the movement of money, does not create its source; the number of users, revenue, profit and the price of a financial title belong to different economic relations, although financial journalism possesses the remarkable ability to transform them into a single magnitude whenever doing so produces a larger number of billions.
But the state’s initial saving contains its own contradiction: while suppliers are competing for entry, the state occupies an exceptionally strong position as purchaser, since it commands a budget, centralised demand and the ability to set one system against another; if, however, a particular system becomes deeply embedded in administrative processes, with user skills, procedures and technical integrations forming around it, then subsequently changing supplier may potentially entail not simply replacing one licence with another, but expenditure on migrating processes, retraining and new integration.
It does not yet follow from this, however, that dependence has already been established: a low entry price merely creates the possibility of future switching costs, whose actual magnitude is determined by system compatibility, data portability, the nature of integrations, contractual terms and the preservation of genuine competition; moreover, the relationship may develop in both directions, since the state apparatus may become increasingly dependent upon private technological capacity that it does not itself possess on the required scale, while technology capital, having acquired access to a large and regularly recurring government market, itself becomes interested in preserving access to the public budget, procurement mechanisms and the rules of admission established by the state.
We are therefore dealing with two qualitatively different forms of social power, which can neither be identified with one another nor regarded as entirely external to one another: corporations command capital, models, computing infrastructure and organised labour-power, while the state centralises effective demand, controls the public budget, establishes the rules of public procurement and possesses administrative power that an individual capital, as a private owner, does not directly possess. The state can use competition among suppliers to reduce prices and alter the conditions of access to the government market, while individual capitals, through participation in that market, expand the sphere of application of the means of production belonging to them; their relationship therefore represents not a simple external confrontation between the state and capital, but the movement of different forces within a single system of economic and political relations.
And here competition begins to reveal its own contradiction: competition through the cheapening of commodities and services gives an advantage to capital capable of operating on a larger scale, sustaining a longer period of reduced initial prices and spreading costs across a considerably greater mass of production; competition therefore not only separates capitals but, under certain conditions, accelerates their concentration and centralisation, displacing weaker participants, transferring part of the market to stronger ones and, through credit and the financial system, combining with the gathering of dispersed monetary resources into the hands of the largest capitals. The resulting movement bears remarkably little resemblance to a static picture of the ‘free market’: the state uses competition to obtain a low price; the low price becomes one of the means of struggle for market expansion; the victors in this struggle may further increase their own scale; and further concentration may weaken precisely the competition that the state initially exploited. The outcome of the transaction thus begins to alter the conditions under which the transaction itself arose.
Here another transformation appears, one that must be strictly distinguished from the technology corporation’s actual profit: if a particular monetary income becomes regular and an expectation arises that it will recur in the future, the right to that future income, when it assumes the corresponding financial form, may acquire a capitalised valuation in the present; this valuation, however, is neither a second profit of the enterprise nor a second copy of its actual productive capital.
Actual capital functions in the process of production — in the means of production and purchased labour-power — whereas a financial title represents a claim to the corresponding income and acquires a market price of its own; that price may rise because the market expects enormous future revenues from artificial intelligence and money capital flows into the corresponding securities, yet the fact that such a title has appreciated by billions in no way means that the enterprise has at that same moment produced billions in new value or received a corresponding amount of net profit.
The multi-trillion-dollar stock-market valuations of technology companies therefore prove nothing in themselves about the profitability of the government contracts under consideration: actual capital advanced, revenue, realised profit, the financial title and the capitalised expectation of future income are distinct economic magnitudes; the price of the title may move under the influence of expected returns, the rate of interest, inflows of money capital and speculation, whereas the actual profit of the enterprise must still be produced and realised.
On the stock exchange, therefore, the expectation of future income may be capitalised, whereas what is sold to the state here is an actual computing service; confusing the two magnitudes would be particularly convenient precisely when one wishes to present wealth that is still merely expected as wealth already produced — the financial form, after all, possesses a remarkable ability to present the present with a bill for income that the future has yet to deliver.
The initial conjuring trick of one dollar and 47 cents now reveals itself completely: the state can obtain a perfectly real saving today — approximately $1.4 billion under its artificial-intelligence agreements, according to its own calculations — while suppliers simultaneously gain the possibility of a considerably larger sphere of subsequent realisation; under our hypothetical average consumption of $100 per month, the 3.5 million federal employees who have already been granted access would, if all of them became active users, correspond to a potential volume of paid services of approximately $4.2 billion annually and $21 billion over five years, while half of that audience would correspond to $2.1 billion per year and $10.5 billion over five years.
This is not a forecast of profit, not a promise of future government expenditure and still less an estimate of the entire American artificial-intelligence market; it is merely the arithmetical expression of the possible scale of the relationship within one documented federal sphere, with the state recording in its calculation a reduction in present expenditure, while capital assesses the possibility of subsequent recurring demand. There is no miraculous transformation of 47 cents into billions here: one side’s account simply ends where the other side’s calculation is only beginning.
The economic scale of the connection between the state and private artificial intelligence became still more apparent in the military sphere: in 2025, the Department of Defense’s Chief Digital and Artificial Intelligence Office concluded separate agreements worth up to $200 million each with OpenAI, Anthropic, Google and xAI for the development of prototypes of advanced artificial-intelligence capabilities to address critical national-security challenges in military and administrative fields; the aggregate maximum contractual ceiling of the four agreements therefore amounts to as much as $800 million — precisely the potential ceiling of the contracts, not the amount of money already actually spent by the state.
Here there emerges a rather peculiar picture of a technology industry ‘free from state intervention’: the state deregulates the conditions governing the development of artificial intelligence, facilitates the construction of its material infrastructure, uses federal funding as an instrument for influencing state regulatory policy, establishes the criteria governing the admission of models to public procurement, opens up a sphere of application for them among millions of government employees, and simultaneously concludes agreements worth hundreds of millions of dollars with the largest developers for the application of the same productive force in the sphere of national security; freedom from the state, it would seem, reaches its most perfect form precisely when the state determines the rules, creates the market and pays for the order.
It would, however, be equally mistaken to conclude from this that the state has simply subordinated itself to a handful of technology corporations: it possesses enormous purchasing power of its own and, through centralised procurement, is capable of altering the position of suppliers, while the General Services Administration reported that, in its first year alone, OneGov consolidated the previously fragmented procurement of federal agencies into twenty centralised agreements with major technology suppliers and generated estimated savings of approximately $1.1 billion.
It is precisely here that the question of ‘objectivity’ ceases to be abstract, since if the state were purchasing a single licence for a single official, a dispute over a procurement criterion would have very limited significance; when, however, the systems in question are made accessible to approximately 3.5 million federal employees, while the largest developers simultaneously receive agreements to develop military capabilities with maximum contractual ceilings running to hundreds of millions of dollars each, the criterion governing a model’s admission to the government market becomes one of the conditions governing the mass social application of the new productive force.
It is therefore necessary to distinguish between two requirements that political language can rather conveniently combine into one: the first concerns verifiable factual accuracy — not inventing facts, indicating uncertainty, accurately conveying the content of a document, and not substituting supposition for what is known — and here an answer can be checked against a source or an established fact; the second concerns ‘ideological neutrality’, yet in order to assess the latter it is first necessary to decide what exactly is to count as the neutral point of departure.
If the system incorrectly states the date of a law, the error is established by comparison with the text of the law; if it calculates a sum incorrectly, the result can be checked by arithmetic; but if it explains the origins of private property, the class structure of society, the role of the state, the causes of war, the relations between capital and labour or the historical nature of the existing social order, a considerably more entertaining measuring instrument appears — the person who first establishes the norm and then discovers a deviation from the norm he himself has established.
American regulation thus encounters its own internal contradiction: it requires artificial intelligence procured by the state to be free from ideologically biased influence imposed from above, yet the criterion determining what is to be recognised as such bias is itself established from above — by the executive branch, then translated into administrative guidance and, finally, made a condition of a government contract; the state, therefore, must intervene vigorously enough to guarantee the absence of state ideological intervention.
But the problem becomes considerably more serious than this sarcasm as soon as artificial intelligence ceases to be experimental software and enters the everyday activity of government institutions, because the question of a model’s criteria then becomes directly connected with the legal consequences of decisions taken by the state apparatus: if artificial intelligence is used to prepare a briefing, analyse a document, assess information or draft an administrative decision, it is necessary to determine whether the result it produces constitutes an official government document, supporting material or a recommendation; who is required to verify its content; whether a citizen can find out that a model was used; whether they have the right to obtain information about the material grounds for the decision; who is responsible for an error — the government official, the agency, the contractor or the developer; and, finally, how such a decision can be challenged if a substantial part of the intellectual operation was performed by a system whose workings the citizen does not control and may be entirely unable to examine.
And here we approach a considerably more dangerous boundary than a government chatbot on an official’s work computer, because a private artificial-intelligence system is now entering that part of the state apparatus which possesses not only a budget and administrative authority, but the direct power of coercion; the question of who produced the text, who verified its content and who bears responsibility for an error contained within it ceases to be academic precisely at the moment when machine-generated text can become part of an official document that directly alters the position of a particular individual.
Police
When Eastern Post first addressed this problem in 2024, artificial intelligence had already begun to perform a function that, until very recently, had seemed too closely bound up with the personal perception and responsibility of a public official to be entrusted to a machine: Axon Draft One took audio from a police officer’s body-worn camera and transformed the recorded incident into an initial draft of a police report, after which the officer merely had to review, correct and sign the narrative produced by the machine. At the time, we asked what might happen if the generative system made a mistake, confused speakers, lost context or produced a convincingly worded detail that had never in fact occurred; two years later, practice itself has already supplied part of the answer.
In September 2024, the King County Prosecuting Attorney’s Office in Washington State informed police agencies across the county that it would not accept reports prepared using artificial intelligence at all, and the basis for this decision was no longer the abstract possibility of a ‘hallucination’, but errors that had actually been identified: according to the prosecutor’s office, the system could incorrectly identify witnesses and officers by name, and in one of the cases examined it included in the report a police officer who had not been present at the scene at all. For an ordinary generative service, such an error might have ended with the correction of a line; in a police document, it meant that an officer would be required, by signing the report, to attest as his own recollection a statement about an event he had not witnessed, while the subsequent discovery of the error would cast doubt not on the machine — the machine, after all, does not give evidence under oath — but on the credibility of the police officer himself. The prosecutor’s office also pointed to a considerably more serious circumstance: the original design of Draft One preserved neither the initial machine-generated text nor the sequence of what the officer subsequently corrected or added, with the result that, once the report had been approved, it became impossible to reconstruct the origin of a particular error — whether the machine had written it and the officer had failed to notice it, or whether the officer himself had subsequently inserted the statement in question. For criminal proceedings, this distinction is by no means merely technical, since the disappearance of the original machine-generated text deprived the prosecution, the defence and the court of the ability to establish the origin of a particular false statement. If the machine had indeed produced the error, the officer might have signed it without noticing; but the reverse possibility also existed: the officer himself might have inserted a false statement, after which the absence of the original Draft One output would make it impossible to prove that the machine had never generated that statement. The machine thus acquired a remarkably convenient property which it could not, of course, exploit for itself: the origin of an error could be attributed to it precisely because the original product of its work no longer existed. This is why the problem lay not merely in the capacity of a generative system to make mistakes.
A human-written report can likewise contain an error or even a deliberate falsehood, but under the traditional procedure the author of the text is known and bears responsibility for his own statement; the use of Draft One introduced, between the event and the signed report, an additional machine operation of text generation that is not itself a subject capable of bearing responsibility, while the destruction of the original machine-generated version erased the boundary between what the system had initially produced and what a human subsequently altered or added. What emerged, therefore, was not simply a new possibility of error, but a new possibility of losing the ability to establish the origin of a particular statement and to determine responsibility for its inclusion in an official document.
And here there arose the first transformation that in 2024 we could only fear: the body-worn camera had originally been intended to create an additional material record of an incident, existing independently of the officer’s subsequent account, but now that recording became the source material for a computing system owned by a private corporation, which used it to produce an initial narrative; the police officer received this text, altered it and, by signing it, transformed it into his own official report, while the original machine-generated version was not preserved. As a result, between the directly recorded event and the final government document there emerged an intermediate product that participated in the formation of the latter but then disappeared, together with the possibility of reconstructing precisely the history of the changes made: the camera preserved what had occurred before its lens and microphone, the final report preserved what the police officer asserted, but it became impossible to establish what the machine had initially written between those two points and what the human subsequently added, deleted or altered. Artificial intelligence thus created not only an additional possibility of error, but a new break in the chain by which its origin could be established — a break that the machine itself could not exploit, whereas the person responsible for the content of the report potentially could.
By 2026, Axon had introduced the ability to preserve the original text generated by Draft One, and it was precisely this function that could make the machine’s participation in the preparation of a police report considerably more transparent: the original machine-generated text could subsequently be compared not only with the body-worn camera recording and the final report, but also with the changes made by the police officer himself. Yet here the same relation revealed itself that has accompanied the machine throughout the history of its application: the technical possibility of control does not yet amount to social control over technology. Preservation of the original Draft One output is disabled by default for American agencies and must be specifically enabled by the police themselves; consequently, the machine is capable of leaving a record of its own work, but the decision as to whether that record will be preserved belongs not to the citizen, the defence, the prosecutor or the court, who may subsequently need it in order to scrutinise the report, but to the police agency itself, whose officer prepares that report. If preservation is not enabled, the officer receives the machine-generated text, may delete some information from it, alter other information, add still more, and sign the revised result as his own official report, while the original text that would later make it possible to establish precisely what the system had generated before human intervention may no longer exist.
Let us imagine an entirely hypothetical case: a body-worn camera records an arrest; the person offers no resistance, he has no weapon, and one of the police officers is the first to use force; Draft One receives the recording and generates an initial text describing the sequence of events in precisely those terms. The officer then opens the machine-generated draft, deletes the statement that there was no resistance, adds an assertion that the detainee ‘made a sudden movement towards the officer’, and states that force was used only after this movement; the final text is signed and becomes the official police report. If the original Draft One output was not preserved, the defence will subsequently have the camera recording and the final report before it, but there will be no document capable of showing that the machine itself initially described the incident differently and that the disputed assertion appeared only after the police officer intervened. And if the question then arises as to where the false statement came from, an almost perfect circle of responsibility appears: the police officer can claim that the machine suggested it, the machine is incapable of confirming or denying anything, and its original text, which could have resolved the dispute, does not exist. The machine here concealed nothing, fabricated nothing and deceived no one; the entire trick consists in a considerably older art — a human being acquired the ability to alter a document, while the means capable of showing exactly what he had altered might not have been preserved.
But it is precisely here that another, considerably more unpleasant aspect of what has occurred reveals itself. If a citizen wishes to challenge an ordinary police report, there is a text before him whose author is known and can be asked why he wrote precisely what he did; if, however, the initial narrative was generated by a machine, subsequently edited by a police officer, and the original version was not preserved, the question of authorship of a particular statement may become technically impossible to resolve. The defence can see the final result, can see the bodycam footage and can attempt to compare one with the other, but the intermediate link — the very machine-generated text that formed the initial structure of the account and thereby provided the human author with the original sequence of events — may be absent.
But the application of artificial intelligence by the state apparatus represents only one side of what is taking place; on a considerably broader scale, the new machine is entering directly into the labour process itself. China’s automated factories demonstrate clearly what artificial intelligence is changing in modern production. Industrial robots are already capable of performing a substantial proportion of physical operations, while artificial intelligence can monitor equipment, detect defects, allocate tasks and manage individual sections of production. Work that previously required dozens of workers can increasingly be performed by a smaller number of people tending a system of machines; the quantity of human labour required to produce a given quantity of commodities is thereby reduced, while a considerably greater mass of products can be produced within a given working time.
For the owner of the enterprise, the meaning is obvious. If, thanks to new machinery, his factory produces a commodity faster and more cheaply than its competitors, he gains an advantage: he can sell more, reduce the price or, for a time, obtain an extra profit. But such an advantage cannot remain exclusive for long. Competitors also install robots and artificial-intelligence systems; yesterday one factory produced a million components with a thousand workers, today the same quantity can be produced by several hundred, and tomorrow this form of organisation of production becomes commonplace throughout the industry. As a result, society expends progressively less human labour-time on the production of each component, and the value of the individual commodity falls.
And here the most interesting part begins. A machine can produce an enormous quantity of things, but it does not in itself create new value. Human labour has already been embodied in it — in the manufacture of the equipment, electronics, software and other elements of the system — and this previously created value is gradually transferred to the commodities produced, whereas new value arises through living human labour; capital therefore finds itself in a peculiar contradiction: every individual capitalist has an interest in replacing as many workers as possible with machines, because this gives him an advantage over his competitors, but when all capitalists do the same, the proportion of living labour directly employed declines relative to the enormous mass of machinery.
This by no means implies that, as machinery spreads, the profits of enterprises must immediately fall, since a factory equipped with more productive technology can produce several times as many commodities, expand its market, displace less productive competitors and, as a result, obtain a considerably greater mass of profit; the mass of profit must, however, be distinguished from the rate of profit, that is, the ratio of the profit obtained to the total capital advanced, which makes possible a picture that is paradoxical only at first sight: the number of factories, machines, robots and commodities produced continues to increase, labour productivity reaches levels never seen before, and the total mass of profit may continue to grow, while the changing ratio between the growing mass of means of production employed and the relatively diminishing mass of living labour creates, other things being equal, a tendency for the rate of profit to fall. Herein lies one of the contradictions in the movement of capital: every individual capitalist, by introducing a more advanced machine, seeks to gain an advantage over his competitors and increase his own profit, but competition compels the other capitals to follow him, with the result that yesterday’s exceptional advantage becomes an ordinary condition of production, and capital once again finds itself compelled to seek a new machine, a new technology and a new means of reducing the expenditure of labour required to produce the commodity.
Artificial intelligence therefore does not abolish the laws that revealed themselves with the development of machine production in the nineteenth century but, on the contrary, makes them considerably more visible: whereas the earlier machine primarily took over the physical movements of the worker, the modern automated system is beginning also to take over the monitoring of the production process, calculation, the detection of deviations and part of the functions of management, enabling an ever greater mass of products to be produced with a relatively smaller direct expenditure of human labour. Technically, this represents an enormous saving of labour-time; yet saving the labour-time required to produce a commodity and freeing the worker’s own time are very far from being the same thing, since, so long as production remains subordinated to the accumulation of capital, the hours released by the machine may assume the form not of a shorter working day, but of fewer workers, a greater mass of commodities produced, an intensification of the labour of the workers who remain, and a renewed struggle to expand the market; the machine frees production from the necessity of employing the previous quantity of human labour, but in doing so it does not in the slightest free the worker from the necessity of selling his labour-power so long as the means of production confront him as the property of others.
But the same machine that, in an automated factory, takes over part of the operations of monitoring, calculation and management of the production process simultaneously enters an entirely different system of social relations — the apparatus of military coercion, and here the scale of state financing itself already demonstrates how far the process has advanced. Through direct Department of Defense contracts, research programmes, the CHIPS Act, the Defense Production Act, Small Business Innovation Research and the accelerated mechanisms of Other Transaction Authority, the state links the development of computing power, microelectronics, robotics, unmanned systems and artificial intelligence with military procurement; surrounding this system are the traditional defence corporations — Lockheed Martin, Northrop Grumman, RTX, General Dynamics and BAE Systems — while a new technological layer is formed by Palantir, Anduril, Microsoft, Amazon Web Services, Google and specialised manufacturers of autonomous systems. In other words, the state is no longer simply purchasing weapons: it is financing the creation of an entire computing environment within which intelligence, communications, data analysis, command and control, and the employment of weapons are gradually being integrated into a single technical system.
This can be seen particularly clearly in Palantir and Project Maven. Maven’s original task was the machine processing of enormous streams of imagery and intelligence data, but the subsequent development of such systems now brings together considerably more functions: TITAN receives and integrates data from satellites and other sensors; Army Vantage brings together information on forces, supplies and resources; Global Force Information Management is used for force management; DCGS-A and All-Source integrate different sources of intelligence information; Maven Smart System processes intelligence, surveillance and reconnaissance data; JADC2 is being built around the integration of the different branches of the armed forces and their information systems. The significance of Palantir therefore lies not in the existence of yet another program on a military computer, but in the fact that a private computing platform enters between the incoming signal and the subsequent human decision: the machine selects what is significant from the mass of data, classifies objects, correlates sources and presents the operator with an already organised picture of what is taking place.
It is precisely here that major government contracts are also being directed: the expansion of Maven is associated with a contract worth approximately $480 million, while the subsequent expansion of Palantir programmes has involved further hundreds of millions, and a considerably broader stream of orders for Anduril, AeroVironment, Lockheed Martin, Northrop Grumman, General Dynamics and other contractors is forming around autonomous systems and military analytics. But the economic significance of this expenditure lies not only in its magnitude. Government procurement creates for private capital a large sphere of effective demand and realisation, finances through contracts particular directions in the development of the technical base, and simultaneously turns the state apparatus into one of its largest consumers; the private corporation, therefore, no longer merely sells the state a finished machine, but participates in the creation and development of the infrastructure through which the state apparatus sees, correlates information and organises its actions.
The nature of military control changes accordingly. Whereas previously a mass of intelligence material had to pass successively through human analysts, the machine is now capable of continuously sorting incoming images, signals and reports, detecting objects, establishing connections between different sources and updating the overall picture as new data emerge; the human remains in the decision-making chain, but the material on the basis of which that decision is made is increasingly pre-selected and organised by a computing system. The AI Action Plan’s formula of ‘continuous adaptation based on AI net assessments’ therefore signifies more than merely an acceleration of traditional staff work: what is involved is a transition to a continuously operating protocol of observation, analysis and adjustment, in which the time between the emergence of information, its processing and subsequent action is reduced.
And here the same contradiction that we have just seen in the factory reveals itself, but in a considerably more dangerous form. There, artificial intelligence economises the social labour-time required to produce a commodity; here, it reduces the time required to process intelligence information and prepare a military decision. In both cases, the machine increases the productive power of human activity, but the social form and outcome of the application of that power are not determined by the machine: in the factory, the means of production function as capital and are under the command of their owner, whereas in the military system the same technical capability is incorporated into the activity of the state apparatus, which makes use, among other things, of computing infrastructure, models and other means belonging to private corporations. Competition for more productive technology, therefore, which within an individual industry appears as a struggle among capitals to reduce costs, expand markets and obtain profit, becomes connected at the international level with the struggle among states for computing power, semiconductors, models, energy infrastructure and autonomous systems, since possession of the corresponding technical base can be transformed not only into an economic advantage, but also into a means of strengthening state and military power.
When a new productive force extends beyond the boundaries of an individual enterprise and becomes one of the conditions determining the competitiveness of entire industries, the struggle over the conditions of its development likewise extends beyond the boundaries of individual capital; what for an individual corporation is a question of productivity, markets and profit becomes, in the movement of numerous capitals connected by a common national market, credit system, infrastructure and state, intertwined with the struggle for position on the world market. The advantage here no longer belongs simply to the enterprise possessing a more advanced artificial-intelligence model, since the large-scale application of this productive force presupposes computing power, microelectronics production, energy infrastructure, scientific personnel, financial resources and the ability to extend the technological system that has been created beyond its own market; the expansion and international competition of large capitals therefore become connected with the activity of the state, which, through trade, financial, industrial and foreign policy, is capable of facilitating access to markets, resources and the technical conditions of production, with the result that the struggle of private capitals on the world market becomes intertwined with the struggle of states for economic and technological position.
This also makes another contradiction intelligible: states may simultaneously facilitate the widest possible diffusion of technologies produced by capitals connected with their national economies and restrict their rivals’ access to those elements of production whose possession confers a substantial technological advantage; private capital seeks to expand the world market for its commodities and services, while state power can employ trade, financial and industrial policy to influence the access of competing economies to raw materials, equipment and other conditions of production. The same duality appears in relation to raw-material resources: foreign sources may be significant not only as a condition for the profitable employment of capital and the supply of domestic industry, since control over access to them can simultaneously affect the productive capabilities of rivals. International competition in artificial intelligence therefore cannot be reduced to the attractive picture of laboratories competing against one another, with one group of engineers attempting to create a more advanced model before another; behind the laboratory there gradually appears the semiconductor factory, behind the factory — the energy system, behind that — the banks and enormous masses of capital, behind the corporation — public procurement, subsidies and foreign policy, and behind the language of technological leadership — the entirely material question of who will own the means of the new productive force and who will be compelled to use them on the owner’s terms. This is precisely why competition, which begins among individual capitals as a struggle to reduce costs and increase profit, can, as production becomes concentrated and capital enters the world market, become connected with the struggle of considerably larger economic complexes for markets, resources and the conditions of further accumulation; throughout all this, the machine remains a machine — it knows neither national borders, nor competitors, nor world domination, whereas the social relations into which it is placed know all of these perfectly well. And here we come to the final question of the article: if the machine itself does not determine to whom it belongs, why it is used or against whom it is employed, then responsibility for the consequences of its use cannot be transferred to it either.
This also makes another contradiction intelligible: large capital has an interest in the widest possible diffusion of the technologies it produces and in the expansion of the world market, yet the same international competition compels it to seek to preserve advantages in access to raw materials, equipment, computing power and other conditions of production; at a certain stage in the concentration of capital, this struggle no longer remains exclusively the private affair of individual corporations, because the economic power of the largest capitals becomes increasingly closely connected with state power, which commands the instruments of trade, financial, industrial and foreign policy. Foreign sources of raw materials therefore acquire significance not only as a condition for supplying industry and for the profitable employment of capital: control over access to them can simultaneously strengthen the position of one group of capitals and restrict the productive capabilities of its rivals, with the result that the private interest of large capital acquires the possibility of appearing in the political form of state interest. International competition in artificial intelligence therefore cannot be reduced to the attractive picture of laboratories competing against one another, with one group of engineers attempting to create a more advanced model before another; behind the laboratory stands the semiconductor factory, behind the factory — the energy system, behind that — the banks and enormous masses of capital, behind the corporation — public procurement, subsidies and foreign policy, and behind the language of technological leadership — the entirely material question of who will own the means of the new productive force and who will be compelled to use them on the owner’s terms. Competition, beginning among individual capitals as a struggle to reduce costs, secure markets and obtain profit, thus becomes connected, as production is concentrated and capital enters the world market, with the struggle of states for markets, resources and the conditions of further accumulation; the state does not thereby become an individual corporation, nor does the corporation become the state, yet the economic power of large capital and the political power of the state no longer operate as two worlds entirely external to one another, but as interconnected aspects of the existing order, while competition among individual capitals within this connection by no means disappears. Throughout all this, the machine remains a machine — it knows neither national borders, nor competitors, nor world domination, whereas the social relations into which it is placed know all of these perfectly well. And here we come to the final question of the article: if the machine itself does not determine to whom it belongs, why it is used or against whom it is employed, then responsibility for the consequences of its use cannot be transferred to it either.
It is precisely here that the question of responsibility acquires decisive significance, because in all the cases we have considered the machine performs an increasingly substantial part of the operation, yet nowhere does it become the subject of the social relations within which that operation takes place: artificial intelligence may propose the text of a police report, select an object of interest from millions of images, allocate tasks in a factory or prepare a response for a government official, but it does not own the enterprise, conclude a government contract, establish the rules governing the preservation of the original report, determine the military adversary or obtain profit from its own application. The expression ‘artificial intelligence made the decision’, therefore, however convenient it may prove after the latest error, conceals far more than it explains: decisions about what system to create, what data to train it on, where to deploy it, what powers to entrust to it, what results to regard as permissible and who is to benefit from the resulting saving of time are in every case made by people and organisations possessing the corresponding property and power; the machine may make a mistake, but responsibility for the social conditions under which its mistake becomes a dismissal, a police document or an element of a military decision cannot be transferred to the machine.
From this follows the limit of purely technical control over artificial intelligence: it is possible to require the preservation of the original machine-generated text, the recording of changes, the disclosure of data sources, the auditing of the model and the retention of a human being in the decision-making chain, and all of this can indeed make the application of the machine more open to scrutiny, but none of these measures in itself answers the question of who determines the purpose for which it is used. A factory machine may be entirely transparent and yet be used to reduce the number of workers rather than shorten the working day; a police system may record every one of its actions in detail, yet the rules governing its use are still established by the police agency; a military algorithm may show impeccably the provenance of every signal it processes, but this says nothing about who determined the military objective. Technical transparency is therefore a necessary means of control, but it is not yet social control itself, for to control a machine means not merely to see what it has done, but to possess the power to decide why it should do it at all, where the limits of its authority lie and to whom the fruits of the productive force it has increased belong.
It is precisely for this reason that the question of artificial intelligence cannot be resolved by the empty formula of ‘social control’ without first answering a considerably simpler question: what society, what class and through what power is to exercise that control? In a society divided into classes, there is and can be no special force standing above the opposition between labour and capital and administering their interests with equal impartiality; whoever holds the principal means of production in their hands possesses not only economic power, since without corresponding political power they could not preserve their economic domination, and the hope that the existing state, itself having arisen on the basis of these same relations of production, will place itself above classes and, through rational regulation, subordinate capital to the interests of society as a whole means leaving the very foundation of the relation untouched and then attempting to repair its consequences. A corporation can be required to disclose how its system works, preserve machine-generated drafts, submit algorithms to scrutiny, prohibit particular forms of their application and establish ever more supervisory bodies; yet so long as computing power, data centres and the other material means through which the new productive force is created and applied function as private capital, such measures can alter the conditions under which the owner operates, but they do not abolish the property relation itself by virtue of which he remains the owner of the means of production; the existing relation may be patched and mended indefinitely, but that does not make it cease to be the very relation that has to be patched and mended.
The question is thereby returned to the point from which it is constantly being diverted: not to a struggle exclusively against individual abuses by corporations, nor to the search for particularly conscientious officials who are supposedly capable of compelling capital to serve an interest opposed to its own, but to the struggle between classes arising from the relations of production themselves. The capitalist is not an accidental ‘bad owner’ of the machine who need only be replaced by a good one, just as the bourgeoisie is not a collection of individuals who have accidentally appropriated economic power; it is a class whose position is continually reproduced by the very system of private ownership of the means of production, and therefore only another class, whose material position places it in an opposing relation to those same means of production, is capable of confronting a class as a social force.
The proletariat becomes such a class — not because any special moral virtue is attributed to it in advance, but because it is deprived of ownership of the means of production, compelled to sell its labour-power and, at the same time, through the very development of large-scale production, brought together, organised and transformed from a multitude of dispersed workers into a social force; its struggle, therefore, if carried through to its foundation, is directed no longer against individual capitalists but against the very order that transforms the social means of production into the private property of capital. Artificial intelligence makes this contradiction particularly visible, because the existence of this productive force presupposes the labour of workers, engineers, programmers, scientists and energy workers, enormous energy and computing systems, global production chains and human knowledge accumulated over generations, yet the result of this increasingly social process can confront the very people who created it as an alien power belonging to capital, commanding their labour and the products of that labour.
Capital thereby creates the material preconditions for its own negation: it concentrates the means of production, brings together previously dispersed labour and transforms production into an increasingly social process, while simultaneously increasing, uniting and organising the class that directly confronts capital; yet the social character of production continues to be opposed by the private form of appropriation of its results. But the economic domination of capital cannot be abolished separately from the political power through which the property relations corresponding to it are preserved, and therefore the struggle of the proletariat against the private appropriation of the social means of production, when carried through to its consistent development, inevitably becomes a political struggle, a struggle for power, without a change in whose class character it is impossible to change the social relation that transforms the means of production into capital.
The actual resolution of this contradiction therefore lies no longer in the supervision of the private owner by an abstract ‘society’, as though society existed somewhere above property relations and class power, but in the expropriation of the expropriators, in the abolition of the private ownership by capitalists of the principal means of production, and in the transfer of those means into the hands of the producers themselves. Only then does the very expression ‘social control’ acquire a definite content: its subject is no longer an imaginary society standing above classes, but the organised class of direct producers, commanding the means of production that previously confronted them as alien property.
This does not mean that the struggle for transparency in artificial intelligence, the preservation of machine-generated drafts, the accountability of public officials, the protection of workers or restrictions on particular forms of the machine’s application are devoid of significance before property relations are changed; such measures can indeed limit particular consequences of the application of the new productive force and alter the immediate conditions under which it is used, but they do not abolish the relation itself within which the means of production continue to confront the producers as alien property. It is precisely for this reason that state regulation of an individual corporation cannot be mistaken for the abolition of the power of capital, just as restricting an individual capitalist does not yet abolish the social relation that repeatedly produces the capitalist as capitalist and the worker as wage labourer.
The question of power over artificial intelligence therefore ultimately proves to be a question not of who will exercise better control over the owners of the machine, but of whether this productive force will have a private owner at all. The machine is capable of reducing the quantity of human labour required to produce a material product, process information, manage a process or perform a multitude of intellectual operations, but it cannot itself determine the social form assumed by the saving of time it creates; so long as this productive force functions as capital, the labour-time it releases can be transformed into a reduction in the number of workers, an increase in output, an expansion of the market and further accumulation, whereas the abolition of the relation of private appropriation itself opens the possibility of setting the machine an entirely different social task — using the saving of human labour that it creates not for the accumulation of capital belonging to others, but for the reduction of the necessary labour-time of the producers themselves.
($1 ≈ £0.7475). Indicative exchange rate as at 17 September 2026. Source: Investing.com.
Illustration / Иллюстрация: Eastern Post.
