Large language models (LLMs) come wrapped in an old promise: the machine will take the toil away. The awkward fact is what the machine is made of. It is built from other people’s writing, code, and commentary, then rented out as a product. Debates about sentience, cyber threats and jobs all matter. But there is a question about ownership running through them: how does the accumulated work of millions become a machine controlled by a few? Marx gives us a way to think about that transformation.
The oldest story in capitalism
Marx’s central claim in Capital is that capitalists profit through surplus value: the gap between the value workers create and the wages they receive (Marx 1867). Machinery raises productivity. Ownership gives capitalists a claim on the profits it helps generate, and power over how the gains are shared. He called capital dead labour: past work put to use in the production process by living workers. Machinery gives that past work a physical form. When someone else owns it, it stands over the people using it.
Now look at a large language model. It is trained on billions of pages of text and code, the products of countless writers, programmers, artists and forum commenters. In Marx’s terms, it’s dead labour on a scale he could scarcely have imagined: the cognitive output of a civilisation, compressed into weights and sold back as a subscription.
Enclosing the commons of the mind
Marx described the enclosure of common land as part of primitive accumulation: the process that separated people from the means of making a living and forced them into wage work. Land they had used in common became someone else’s property.
Much of the internet grew through people sharing knowledge, writing code and answering one another’s questions. AI firms have harvested that shared work, along with copyrighted books that were never freely offered to the public (U.S. Copyright Office 2025). They have used it to build proprietary models, then charged for access.
The analogy with enclosure has a limit. Training a model does not, by itself, fence readers out of the original texts. The fence surrounds the productive capacity built from them: the model that can turn this accumulated knowledge into answers, drafts and code. Its owners decide who can use it, on what terms, and at what price. Customers pay the company; after wages and other costs, any profit belongs to its owners. The people whose shared work helped build the model have no automatic claim on that income.
Marx had a phrase for society’s accumulated knowledge once it becomes a force of production: the general intellect. LLMs give that idea a striking contemporary form. The knowledge is social, while control over the machine is private.
That ownership can reach back into the lives of the people whose work made the model possible. A writer or programmer may find their accumulated craft turned into a tool an employer can use to bargain down the price of their next piece of work. Dead labour returns as leverage over the living.
The counterarguments
Developing, training and serving a model takes real research, engineering and investment. The companies have added something valuable. But that does not settle the terms on which everyone else’s work was taken, or who should control what was built from it. The engineering can be real and the appropriation real too.
Open-weight models complicate the picture. Releasing the trained weights lets others run and adapt a model, subject to its licence, and falling costs broaden access (Stanford Institute for Human-Centered Artificial Intelligence 2025). Some smaller models can run on phones and laptops. The largest demand costly hardware, and releasing weights does not distribute the infrastructure needed to train their successors. These are real openings in the fence, but access to a model and control over its future are different things.
Who should own the machine?
Marx distinguished machinery from its capitalist use. He saw technology that reduces necessary labour as a precondition for a better society: it could shrink the working day and free time for human development. Whether it does so depends on the social relations around it, including who owns it and whose interests govern its use.
A Marxian response today could mean public or cooperative control of models, compute and data, a democratic say over deployment, and gains coming back as shorter hours and higher living standards. LLMs could make us more productive and still leave those questions unresolved. A machine built from our collective work gives us reason to ask for a collective say in its use and a collective share in the wealth it produces.