The Robot Is Dead Labour

The detail that stayed with me was buried a long way down the article. A company had begun putting a humanoid robot into homes in San Francisco, the coverage was about the strangeness of a machine folding laundry in a domestic kitchen, and several paragraphs past the photographs came the operating caveat: when the robot cannot manage a task on its own, a remote employee puts on a headset and drives the arms. Somewhere in an office there is a person whose hands are inside a stranger’s house.

Most of the reaction I read was about privacy, which is reasonable to be alarmed by and also the smaller half of what is happening. The more interesting question is what that operator’s hands are doing economically. Every shirt they fold through the headset is a training example. Every recovery from a dropped cup, every awkward reach behind a cupboard door, every moment of improvisation that a policy model cannot yet produce on its own is being recorded, labelled by the act itself, and folded back into the next version of the system. The operator is being paid a wage to manufacture the asset whose entire purpose is to make the wage unnecessary. That is the robotics economy in a single image, and it is why I think the argument currently on offer about robotics, which is an argument about open weights against closed ones, runs past the thing that will actually decide how this lands.

The constraint is data that has to be produced

It helps to separate the frontier from the coverage. The photogenic part of robotics is the form factor, the walking, the human silhouette, and those are the parts that fill a launch video. The hard part is contact. Manipulating objects the system has never seen, under friction and deformation and partial occlusion, recovering when a grip slips rather than continuing confidently into a mistake, generalising from a kitchen in one house to a kitchen in another where the drawers are different and the light is worse: that is where the field actually sits, and progress there has been real and much slower than the demonstrations imply.

The reason it is slow is worth being precise about, because it is not a shortage of ideas. Language models arrived into an unusually lucky situation, which is that a corpus of human expression already existed, produced over decades for entirely unrelated reasons, sitting on public servers, and the work was to scrape it and to spend enough compute on it. Robotics has no equivalent lying around. There is no internet of hands. Every hour of manipulation data has to be produced deliberately, by teleoperation rigs, by instrumented fleets running in warehouses, by cheap open hardware in university labs, and increasingly by machines placed in real homes with a person in the loop to cover the gaps. The data has a cost of production, paid in wages and capital equipment, and its supply is controlled by whoever is running enough machines in the world to generate it.

In language models the raw material was a commons that got enclosed after the fact, and the arguments about it have been arguments about enclosure: what was scraped, from whom, with what consent, and whether releasing weights gives anything back. Robot data is private from the moment it exists. It has no prior public life to be taken out of, because it did not exist until somebody paid for a machine and a person to generate it. The fleet is the wellspring, the model is the byproduct, and a company that gives away its weights while keeping its fleet has given away the cheap half.

Dead labour, and the awkward arithmetic

Marx has a vocabulary for this that is older and sharper than anything the current discourse has produced, and it is worth using precisely rather than decoratively.

Machinery in his accounting is constant capital, and his phrase for it is dead labour: the congealed product of past human work, transferred into the product a bit at a time as the machine wears out, incapable of creating new value on its own. Labour power is variable capital, and it is the only input that can produce more value than it costs, which is where surplus comes from. Put those two together and a robot is not merely a tool that displaces a worker. It is previous human labour, embodied in metal, set against present human labour in a competition the present cannot win on cost.

The teleoperation loop makes that transfer literal in a way the nineteenth century could only describe by analogy. In a textile mill the weaver’s skill was decomposed into machine motions by engineers, over years, through observation and redesign. In a teleoperated fleet the demonstration is the mechanism. The skill leaves the body and enters the model directly, is copyable from then on at close to zero marginal cost, and returns to the labour market as a competitor. Marx’s line in the chapter on machinery, that the instrument of labour strikes down the labourer, was written about a process that took a generation. The version running now takes an afternoon and a data pipeline.

Then the arithmetic gets awkward for capital before it gets awkward for anybody else. If surplus value arises only from living labour, an economy that systematically replaces variable capital with constant capital is raising what Marx called the organic composition of capital, and the rate of profit tends downward as it does so, since the base that generates surplus is shrinking relative to the total capital advanced. A perfectly automated firm, in these categories, produces goods and no surplus value at all. I want to be honest that this tendency has been argued about for a hundred years, that the countervailing factors Marx himself listed are substantial, cheapening the machines, raising the intensity of the labour that remains, opening new markets, and that plenty of serious people regard the whole construction as unsound. It is still the case that the most thoroughgoing theory anyone has of what happens when machines do everything was written by someone who thought it would break the system’s own accounting rather than deliver abundance to its owners.

The passage that reads most strangely today is in the Grundrisse, in the section usually called the Fragment on Machines. Marx describes a stage where general social knowledge, the accumulated understanding of a society, becomes a direct force of production, embodied in fixed capital, at which point labour time stops functioning as a sensible measure of wealth even though the system keeps insisting on measuring with it. A foundation model is that description with the metaphor removed. It is collective knowledge, extracted and compressed, and a robot is that knowledge issued with hands. Marx treated the moment as the point where the value form comes under a strain it cannot indefinitely absorb, and he was clear that the machine itself decides none of it. The same machinery immiserates or emancipates depending on who owns it, which is the whole of his argument and the part that the current conversation keeps mislaying.

Why open weights are not a commons

The reflex answer to concentrated ownership in this industry is to open the models, and I understand the appeal, because it is concrete, it is available now, and it has an honourable history. It also socialises the layer that is cheapest to reproduce while leaving every expensive layer where it was.

Consider what open weights actually distribute. They distribute a blueprint. They do not distribute the fleet, the factory, the actuators, the precision reducers that come from a small number of Japanese suppliers and remain a genuine chokepoint, the electricity, the capital to keep several thousand machines running in the field, or the data pipeline those machines feed. A researcher who downloads open robot weights still cannot generate the demonstrations that would improve them at scale, and a company with a large deployed fleet can absorb the open model, fine-tune it on proprietary data nobody else has, and end up further ahead than before the release.

There is a precedent here. Free software socialised the infrastructure layer of computing comprehensively, and the largest single consequence was that the cost of building hyperscale cloud platforms fell dramatically for the handful of firms with the capital to build them. The commons was real, the labour donated to it was real, and the enterprises that captured the most value from it were the ones that owned the physical layer underneath. Robotics has a harsher version of the same asymmetry, because the marginal input is not compute purchased on a spot market, it is machines in the world, and machines in the world are owned.

None of that makes the vertical integration case wrong, and I would rather concede its strength than argue against a weak version of it. Owning the actuator, the model, the fleet and the deployment closes the loop between what the machines do and what the model learns, and loop speed compounds in a way that breadth does not. The open ecosystem, the pooled cross-institution datasets, the cheap hardware, the university rigs that made manipulation research affordable, wins on cost and on the number of people who can participate. Neither arrangement touches ownership of the fleet. That is why the open-against-closed argument, however it resolves, is a technical argument that has been asked to carry political weight it was never built for.

What would actually count

If the question is whether there is a socialist model of robotics in the strict sense, meaning social ownership of the productive apparatus and of the surplus it throws off, then the useful move is to run the candidates against a single test: does the mechanism change who owns the fleet and the data it generates, or does it only redirect some of the income the fleet produces.

Most of what gets proposed touches income. A robot tax redirects income. A universal basic income funded from corporate profit redirects income. A sovereign wealth fund of the Norwegian kind holds a socially owned stake in capital and distributes the return, which is meaningfully more than a transfer and still leaves control at the point of production untouched, and the Alaskan version is thinner again. These are serious instruments and they belong to social democracy, where private ownership continues and the state taxes and redistributes around it.

The proposal that went at ownership directly was Rudolf Meidner’s, drafted for the Swedish trade union confederation in 1976, which required firms above a certain size to issue new shares each year into union-controlled wage-earner funds, so that over a few decades the workforce would come to own the firms outright without anybody being expropriated in a single dramatic act. It was diluted, fought over for fifteen years, and wound up in the early nineties, and it remains the clearest worked example of a mechanism aimed at the asset rather than the cash flow. Translated into this industry it would mean a levy on robot-driven profit paid in equity rather than in tax, accruing to funds controlled by the people the robots displaced.

The cooperative tradition, Mondragon being the standing example, puts ownership in the hands of the workers of a particular firm, which resolves the question inside the enterprise and leaves it open across the economy, since a cooperative that owns a fleet still competes against everyone else on the same terms and has the same reasons to automate. The market socialist line, Oskar Lange in the calculation debate, Yugoslav self-management in practice, John Roemer’s coupon schemes more recently, tries to hold the capital stock socially while keeping market allocation, and it is the tradition that has thought hardest about how you would own a fleet in common without also having to plan its every movement from a ministry.

What none of these currently have is an instance. There is no fleet of general-purpose robots anywhere in the world that is socially owned along with the data it produces, and there is no serious proposal in front of any legislature to create one.

The case that is already running

The complication sitting inside the region I write about is that the largest robotics buildout in history is happening in a country that calls itself socialist. China accounts for more than half of the world’s annual industrial robot installations, its domestic manufacturers have been taking share from the Japanese and European incumbents inside its own market, and the buildout is steered through industrial policy, provincial funds, procurement targets and directed credit in a way no market economy attempts.

Read against the test above, it is a study in how far state direction of investment can travel without becoming social ownership of the surplus. The capital is allocated politically, and the profit is very largely appropriated privately, by firms whose shareholders are not the workers displaced and whose fleets are not held in common. Ownership of the data those fleets generate has not, as far as I can tell, been raised as a public question at all. The lesson I draw is narrow and I think defensible: the label a system gives itself settles nothing about who owns the machines, and an industrial policy powerful enough to build the world’s robot base is not the same instrument as one that would socialise it.

Where that leaves the question

I do not have a programme to offer. What I have is a test that I find clarifies almost everything written about robots and fairness, which is to ask of any proposal whether it changes who owns the fleet and the demonstrations that train it.

Run the current field through it. Open weights do not. A robot tax does not. Ethics boards and safety standards, which are worth having for their own reasons, do not. Sovereign funds partly do, at the level of the return without the level of control. Meidner’s mechanism did, which is a reasonable part of why it was defeated.

And then go back to the person in the headset with their hands in someone else’s kitchen. They are producing, in the most direct form the economy has yet arranged, the fixed capital that will stand against them. Marx’s claim was that this arrangement is a property relation rather than a fact about machinery, that the identical robot returns time to a person under one arrangement and takes an income from them under another, and that nothing in the engineering decides which. On that reading, the whole redistribution argument comes down to whether anybody proposes to change the relation, and at the moment the loudest proposals on offer are about the licence file.