Liberalism’s Token Limit
A Marxist critique of the fantasy that artificial intelligence can be regulated, decentralized, and redistributed under capitalism
A growing section of the technology world has come to recognize that artificial intelligence, as presently developed, is not a neutral marvel. It is built on exploited labor, concentrated infrastructure, surveillance, environmental damage, and corporate power. The more serious versions of this critique no longer sound like Silicon Valley boosterism. They are humane, democratic, and attentive to real suffering. That is precisely why they deserve scrutiny. The problem with liberal-progressive AI politics is not that it ignores capitalism altogether. It is that it sees capitalism as a set of abuses to be regulated, decentralized, and redistributed around, rather than as the social relation that gives AI its present form. Its horizon is a better-managed AI capitalism. A socialist politics has to begin from a harder premise. Under capitalism, AI is not simply a tool being misused. It is machinery organized for the domination of labor.
One of the more serious examples of this liberal-progressive position is a recent essay by Alice I. Cecile, a technologist and transhumanist who argues that AI is powerful, cannot be uninvented, and must therefore be shaped rather than rejected. The transhumanist commitment matters here not as a personal defect, but as a symptom of the broader politics under examination. It is a tendency to imagine technical systems as routes of escape from material limits, rather than as forms of social labor embedded in energy, extraction, infrastructure, and class power. Cecile’s essay correctly identifies many of the harms. It notes the surveillance, labor displacement, environmental costs, and the concentration of power in a handful of corporations. The proposed remedies are regulation to stop avoidable harms, decentralization of control through open-source models, and redistribution of the productivity dividend. These are not frivolous demands. They reflect a genuine revulsion at the way AI under the productive forces of capitalism operates. But the essay also reflects a political horizon that cannot escape the logic of the system it criticizes. The question is not whether AI can be made open, regulated, or redistributive under capitalism; because it cannot. The question is who controls the machinery, whose labor does it embody, and for what social purpose it is deployed.
To understand why liberal AI politics reaches its limit, it is necessary to grasp what AI actually is in Marxist terms. Under capitalism, AI should be understood as machinery. It is past human labor embodied in systems that capital uses to discipline, cheapen, and reorganize living labor. It is what Marx called dead labor. A type of constant capital in perhaps its most developed form. The training data, the labeled datasets, the scraped content, and the hardware supply chains are all products of global proletarian labor. They come from miners extracting the minerals that feed the electronics supply chain and from the low-paid data workers and content moderators in the Philippines, Kenya, India, and elsewhere who prepare training data for poverty wages. When an AI model generates text or makes a prediction, it is not performing a magical act of machine intelligence. It is realizing the value that was already inscribed in it by the living labor that created it. The apparent autonomy of AI, its capacity to displace human workers and appear to function independently, is precisely the form in which dead labor dominates living labor under capitalism.
Cecile names the exploitation behind AI but does not follow that observation to its political conclusion. If AI is dead labor, then its capacity to displace living labor is not an accidental abuse. It is central to its deployment under capitalism. Capital introduces AI not to liberate humanity from toil but to cheapen production, weaken worker bargaining power, and increase the rate of surplus value extracted from those who remain employed. AI can raise productivity, generate outputs, and help firms capture profit. But in Marxist terms, machinery does not create surplus value by itself. It transfers value while reorganizing production around the exploitation of living labor. The benefits of AI cannot be durably democratized while the infrastructure, data, labor process, and surplus it organizes remain privately controlled.
The first major weakness of the liberal program is its faith in regulation. Cecile calls for legislation to stop environmental harms, restrict deepfakes, require renewable energy for data centers, and impose various other controls on AI companies. The underlying assumption is that the state is a neutral arbiter that can be pressured into acting against the interests of capital. But the capitalist state is not an instrument that can simply be turned against the class power on which it depends. It is, as Marx and Engels argued, but a committee for managing the common affairs of the whole bourgeoisie. Its legitimacy, its financing, its personnel, and its very function are embedded in and dependent upon the reproduction of capitalist social relations. The examples of regulation cited by Cecile inadvertently confirm this analysis. The EU AI Act, characterized as a bold proposal, regulates only high-risk applications while leaving the fundamental dynamics of capital accumulation through AI entirely untouched. California’s transparency legislation does nothing to halt the extraction of surplus value. These are not failures of political will. They are the logical limits of what bourgeois regulation can accomplish within a system dependent on private accumulation.
Even ambitious risk-based regulation can leave the ownership structure intact. It can restrict some uses, require documentation, and impose compliance obligations, while leaving untouched the deeper question of why a handful of firms control the infrastructure, data, compute, and distribution channels on which AI depends. The state can moderate some visible externalities of AI capitalism, and working-class movements can force real concessions from it. But regulation alone cannot overcome the ownership structure that gives capital command over AI in the first place. Regulatory capture by AI firms, compliance costs that favor incumbents, national-security competition around AI, and weak enforcement against labor surveillance all ensure that regulation manages the system rather than transforms it.
The second major weakness is the call to redistribute the productivity dividend. Cecile imagines redistribution dramatic enough to make the wealthy “flinch and gasp.” But the essay does not explain how this redistribution is to be achieved, who will enforce it, or what structural transformation of the economy would make it possible. The demand essentially amounts to a call for a more generous welfare state or a universal basic income subsidized by taxation of the technology sector. Without a theory of the organized power needed to impose it, redistribution remains a moral demand rather than a political strategy. Historically, the major redistributive settlements of modern capitalism have generally followed periods of intense class struggle, war, revolutionary pressure, or systemic crisis. They were not granted because moral arguments were convincing. They were won when capital faced organized pressure strong enough to threaten profits, property, or social stability. The weakness of Cecile’s proposal is not that redistribution is undesirable, but that it floats free of the class power required to impose it.
The third and most revealing weakness is the open-source illusion. Cecile calls to open-source everything, to decentralize control over AI technology so that the AI barons cannot monopolize its benefits. This proposal has an undeniable appeal to technologists accustomed to the culture of open-source software, and it gestures toward a genuine problem. It highlights the extreme concentration of power in a handful of technology corporations. But examined critically, it reveals itself as a fantasy of individualized liberation that ignores the structural imperatives of capital accumulation. The fundamental error of the open-source proposal is its assumption that the technology, once liberated from corporate control, can be made to serve human needs. Open source can democratize code, but it does not democratize compute. It can publish model weights, but it cannot by itself socialize energy grids, chip supply chains, cloud platforms, training data, or distribution networks.
AI is not a static tool that can be separated from the conditions of its production. Every major AI model requires enormous capital investment in computational infrastructure, energy, data acquisition, and skilled labor. Under capitalism, whoever controls these inputs controls the technology, and that control is exercised through enclosure - the private appropriation of socially produced means of production. Making model weights publicly available does not eliminate the barriers to running, maintaining, and improving those models precisely because the inputs remain enclosed. As Cecile acknowledges, training a frontier model costs billions of dollars. No amount of open-source ideology changes this material reality. What would actually happen if major AI models were released as open-source is precisely what has already occurred in the history of open-source capitalism: large corporations would enclose the free technology within their proprietary platforms, extract value from it, and compete even more ruthlessly with smaller developers who cannot afford the complementary investments needed to make open-source AI usable at scale. The problem is not that open-source code is insufficiently proprietary. The problem is that the means of production remain enclosed.
The history of open-source software shows how easily a commons can be absorbed into capitalist production through enclosure. Linux, the paradigmatic open-source project, now depends heavily on contributions from corporate employees, including workers at IBM, Google, Intel, and other firms. The open-source AI movement, far from undermining the AI barons, would likely provide them with free research and development while they continue to enclose the value through their control of distribution, infrastructure, and complementary services. The vision of decentralized AI imagines a world of small technical proprietors using advanced machinery outside the domination of monopoly capital. But the material basis of AI points in the opposite direction. The productive forces of AI have developed to a point where they are inherently social, requiring enormous collective investment and infrastructure, and the relations of production remain stubbornly private. This contradiction cannot be resolved by making the source code public. It can only be resolved by abolishing the enclosure of the means of production and bringing them under democratic control.
Cecile’s essay is useful not because it is idiosyncratic, but because it gives unusually clear expression to a broader politics among technical workers whose experience of capitalism is mediated through autonomy, expertise, credentials, and intellectual property. Within this milieu the tendency for political remedies focuses on the preservation of those forms while trying to soften the domination of capital. Their program does not require abolishing wage labor or private property but to make the system fairer, sustainable, and inclusive. They want the benefits of AI without its harms, the productivity of automation without the displacement of workers, and the creative potential without the exploitation of labor. All while ignoring the contradiction that an AI whose productive basis is collective its ownership and command would remain private.
The impasse of radicalized professional politics is that it can see the irrationality of capitalism while preserving the social position from which that irrationality is viewed. Its preferred remedies often protect autonomy, expertise, credentials, and intellectual property, while asking capital to behave less destructively. The result is not simple hypocrisy. It is a class contradiction. Technical professionals are increasingly proletarianized by the same systems they help build, yet many still experience their relation to capital through professional status rather than collective working-class power. The call to unite around reformist demands can therefore become, in practice, a call for workers to accept a slightly less brutal capitalism rather than fight for control over production itself. Reform struggles matter, but they cannot substitute for the transformation of the social relations that make AI a weapon against labor.
Decentralized protocols and local autonomy matter where they build working-class capacity and reduce dependence on brittle capitalist systems. But AI poses a different problem. Its material basis is already centralized, capital-intensive, and socially produced. The answer to such concentration cannot be the fantasy of isolated technical self-sufficiency. It must be collective control over the infrastructure that already organizes collective life. The question is not whether AI can be made humane under capitalism, but whether the class that owns it will ever voluntarily surrender its command over the labor process, the data, the energy, and the hardware that constitute the new means of production.
A socialist politics of AI cannot stop at access, transparency, or compensation. It has to raise the question of ownership and control. In capitalist production, AI is organized to extract surplus value from living labor, concentrate wealth and power, and discipline workers through surveillance, algorithmic management, and the threat of displacement. Regulation can restrain particular abuses. Open source can widen access to code. Redistribution can soften some effects of automation. But none of these, by themselves, changes the class power that determines what AI is built for.
A serious program would connect immediate struggles to socialist transformation. It would demand public ownership of cloud and compute infrastructure, democratic control over data-center siting and energy use, union veto rights over workplace automation, bans on algorithmic productivity surveillance, and reductions in working time wherever automation raises productivity. It would fight for public AI systems built for social need rather than profit, for expropriation or public-utility control of dominant cloud platforms, and for enforceable international labor standards for data annotation, moderation, and supply-chain work. These demands would not abolish capitalism on their own. Their value is that they shift the terrain from consumer access and technical transparency to class power over production.
None of this can be won by policy design alone. It requires unions, tenant organizations, public-sector workers, data workers, artists, engineers, and logistics workers acting not as injured stakeholders but as a class with the power to halt production. That is the limit of liberal AI politics. It wants to humanize machinery while leaving capital in command of the humans. A socialist politics begins from the opposite premise. The accumulated labor embodied in AI must be brought under the democratic command of living labor. The question is not how to make AI capitalism kinder. The question is how to take the machines out of capital’s hands.
Cecile’s essay invokes Marx’s view on the Luddites in support of embracing automation. The footnote cites seizing the means of production, not smashing it as the correct interpretation. But this reading gets the historical materialist position exactly backward. Marx did not celebrate the automatic loom as a neutral force that capital simply misused. He described how machinery under capitalism becomes a weapon against the worker, how the very development of the productive forces under capitalist social relations produces the domination of dead labor over living labor. The Luddites were not primitive for smashing machines; they were resisting the form in which capital deployed those machines, a form that made the worker an appendage of the machine. Marx’s critique was not that the Luddites misunderstood the neutrality of technology, but that their resistance could not succeed so long as the relations of production remained intact. Writing in Capital, Marx emphasized that the real target should be the specific social form in which machinery is used under capitalism, as constant capital embodying the power of dead labor over living labor. The working class must seize and transform those relations, not smash the material basis of socialized production.
Cecile’s misreading is not incidental to the argument. It is the hinge on which the entire liberal frame turns. By treating AI as a neutral tool whose harms flow from its capitalist deployment rather than from the logic of value accumulation itself, the essay can sustain the fantasy that regulation, open-source, and redistribution are sufficient remedies. The Marxist point is not that we should all become Luddites and destroy data centers. The point is that the contradiction is not between humans and machines, but between the socialized character of production and the private form of appropriation. The legacy Cecile invokes is not a craftsperson’s revolt against tools that can be re-appropriated as property. It is a class struggle over the entire social form of the productive forces.
Once this misreading is corrected, the political conclusions invert. The answer is not to update Luddism for a digital age, hoping we can smash our way back to artisanal autonomy. Nor is it to adapt, as Cecile counsels, to the inevitable rule of machinery. The only coherent position is to demand the full socialization of all productive forces. We must fight, and yes likely violently, for complete democratic command over the accumulated dead labor embedded in AI and more broadly all means of production. Our task is not to smash the machines but to abolish the class relation that turns them into instruments of domination.