If AI Can Do the Work, Why Are We Still Working?

Thanks for reading Anti-Capitalist Musings. It is a small operation, and I hope it offers something worth your time. There will be no premium subscriber content here: everything published will remain free to read. If you value these pieces and want to support the writing, buying me a coffee helps fund media subscriptions and the books that keep the analysis grounded. Every contribution, however modest, is genuinely appreciated.

There is something backwards about the discussion of artificial intelligence and employment. We are told that AI may become capable of doing more of the work currently performed by humans, perhaps much more. Productivity will rise. Tasks that once required teams of people may require one person, or eventually none. Then, almost without noticing the contradiction, we are told why this would be disastrous: there might not be enough work left for everyone.

It is worth dwelling on that. A technology succeeds in reducing the amount of human labour necessary to produce things, and we experience the result as a threat. The machine has done exactly what a labour-saving machine is supposed to do, but the worker who has been saved from the labour may lose their income, their home and much of their ability to participate in society. The problem is not immediately the technology. It is the relationship between work and access to everything needed to live.

A recent paper by Brett Hemenway Falk and Gerry Tsoukalas, The AI Layoff Trap, comes at this from a fairly orthodox economic direction but ends up exposing something much larger. Their argument is that firms can become trapped in an automation race in which every individual decision makes commercial sense while the aggregate result damages both workers and businesses. Each company captures the saving from replacing its workers with AI, but it bears only a small share of the economic damage caused when those former workers lose income and spend less. Most of that loss lands somewhere else.

The mechanism is easy enough to understand without the equations. A company replaces a thousand workers and saves their wages. Those thousand people respond by cutting their spending. They cancel a holiday, put off replacing the car, spend less in shops and restaurants. The company that made them redundant experiences only a fraction of that lost consumption because those workers bought things from hundreds of other businesses. From the point of view of the company doing the automating, the wage saving is concentrated while the damage is dispersed.

Every competitor faces the same calculation. A company that refuses to automate does not preserve the wider economy because its competitors can still remove workers and weaken demand. It is simply left paying higher labour costs while suffering the same deteriorating market. Falk and Tsoukalas show that, under certain conditions, this develops into a Prisoner’s Dilemma1. Companies continue automating even when they would collectively make higher profits if they exercised restraint.

The paper calls this a demand externality. An individual firm does not account for the full economic consequences of eliminating a wage because much of the lost spending hits other firms. More competition can actually deepen the problem. A monopolist has to consider most of the demand destruction it causes because ultimately those customers are its customers. Where an industry contains many competing firms, each bears only a fraction of the loss produced by its own automation. The authors’ model therefore produces the counter-intuitive result that fragmented competitive markets can automate further beyond the collectively profitable level than concentrated ones.

None of this establishes that mass unemployment from AI is imminent. The authors are clear about that. Their model identifies a structural vulnerability rather than claiming that the United States, Britain or anywhere else has already entered such a crisis. They suggest that one empirical signal would be large-scale AI-related layoffs occurring alongside declining profits, but acknowledge that displacement has not yet reached anything like the scale required to establish the effect they model.

There is another argument I am deliberately leaving to one side here. None of this settles whether AI at its present cost can produce the profits its owners expect, whether the enormous expenditure on chips and data centres can be sustained, or what the energy, water and material demands of expanding that infrastructure mean environmentally. Those are serious questions, and they may place hard limits on some of the more extravagant claims being made for AI. My argument begins one step later: if the technological claims are broadly realised, and AI really can substitute for large quantities of human labour at an economically and materially sustainable cost, what should we want that productivity to do for us?

So, for the purposes of this argument, suppose these systems really do become capable of replacing substantial amounts of human labour at an economically and materially sustainable cost. What exactly are we trying to protect when we talk about protecting jobs?

The authors are mainly concerned with preventing excessive automation during the transition. They examine retraining, profit sharing, taxation and Universal Basic Income. UBI can support household incomes and maintain demand, but in their model it does not alter the calculation facing an employer deciding whether the next task should be performed by a worker or a machine. The firm still receives the saving from automation. Giving everyone an income afterwards changes people’s living standards, but not the incentive that produced the displacement.

Their preferred intervention is therefore a Pigouvian automation tax, essentially making businesses pay for some of the wider demand destruction their decisions impose on others. There is a perfectly respectable economic argument for this. If a company imposes costs it does not itself bear, taxation can force those costs back into its decision. The authors suggest that the revenue could then help displaced workers through income support or retraining, with the tax declining if workers are successfully absorbed into new employment.

But if AI becomes capable of removing a very large quantity of necessary labour, why should our objective remain finding ways to ensure that everybody continues performing roughly the same amount of paid work?

If producing the goods and services required for a decent life takes forty hours of human labour per person today and technological development reduces that requirement to twenty, something remarkable has happened. Society has gained twenty hours. Yet there is nothing automatic about those hours being returned to the people who previously performed them. One possible outcome is that half continue working forty hours while the other half find themselves unemployed and looking desperately for another employer.

The labour has disappeared in both cases. What differs is who receives the benefit. And under the present system there is an answer to that question. AI is not socially owned machinery whose gains automatically become a common dividend. The systems used to replace labour are owned and controlled by capital, so the immediate benefit of reducing labour appears as a reduction in costs, an increase in profit or an advantage over competitors. The worker experiences automation as the loss of a wage; the owner experiences it as a productivity gain. That distribution is not a property of AI. It follows from the ownership of the productive system.

This distinction is already present inside Falk and Tsoukalas’s model, even though they do not take it in this direction. Automation shifts income away from workers and towards firm owners, while workers tend to spend a greater proportion of their income. The problem is therefore not simply that a machine has performed a task. It matters who owns the machine, who receives the saving and how access to the resulting output is organised.

Productivity is usually discussed as an increase in output per worker, but from the standpoint of human beings it could have another meaning: spending less of our lives doing things that have to be done. The possibility created by AI is therefore not merely greater production. It is potentially greater control over our own time. Private ownership separates those two things.

A society organised around that principle would have very different reasons for developing AI. We would not need to ask which occupations could be eliminated most profitably. We might ask which forms of labour people most want to escape. Dangerous industrial inspection would have a stronger claim on automation than illustration. Sewage maintenance would be a more urgent technological target than producing marketing copy. Night work, repetitive administration and physically destructive labour could be treated as problems to be removed rather than simply costs to be cut.

There would still be work. The fantasy that a sufficiently clever language model abolishes every form of human activity does not survive much contact with the real world. People need care. Infrastructure has to be maintained. Decisions have to be made about resources and priorities. Some kinds of work may remain stubbornly human, while others might be technically automatable but undesirable to hand over entirely to machines.

The remaining human work would not have to be divided on the absurd assumption that every hour is equivalent. A society capable of measuring productivity down to the second can surely recognise that an hour spent maintaining a sewer at three in the morning asks more of someone than an hour helping in a library. Instead of everyone owing an identical number of working hours, different kinds of necessary labour could carry different credits. An hour of unpleasant, dangerous or antisocial work might count for three; an hour of ordinary work for one. Care work, including looking after children or relatives, would count as contribution rather than disappearing from the calculation as it so often does now.

The point of the credits would be to distribute necessary labour, not to recreate wages under another name. Completing twenty credits would not entitle someone to better healthcare, more food or a superior basic standard of living than someone who could not work. Housing, food, healthcare, education and whatever society decided constituted its basic provision would sit outside that calculation. Illness, disability and age would require straightforward exemptions rather than another bureaucracy devoted to deciding whether somebody was sufficiently deserving.

There would also be a useful pressure built into such a system. If cleaning sewers at night carried three times the labour credits of working in a public garden, society would have a strong reason to develop machines capable of doing the sewer work. The technological priority would no longer be simply to automate whichever worker represented the largest cost to an employer. It would be to remove the forms of necessary labour people least wanted to perform. As automation improved, the number of credits society required from each person could fall. Productivity would finally have a direct relationship with free time.

This seems to me more interesting than promising that every job eliminated by AI will somehow be replaced by another one. If a machine removes ten thousand hours of genuinely necessary work, why should public policy immediately begin searching for another ten thousand hours of work for humans to do? That only makes sense if employment itself has become the objective. A system of labour credits starts from the opposite assumption: necessary labour is a burden to be shared fairly and, where possible, reduced.

There are obvious dangers in this. A supposedly liberated society could recreate the workplace through a gigantic bureaucracy measuring everybody’s contribution. The state could become the new personnel department, complete with targets, assessments and somebody checking why you only recorded eleven credits last month. Any serious argument for reducing wage labour has to take that possibility seriously because abolishing private employment does not abolish power. It merely changes the institutions through which power is exercised.

There is another difficulty with the word “free”. Removing prices does not remove scarcity. If automated agriculture can produce enough staple food for everyone, there is no compelling reason access to that food must depend upon wages. The same argument could eventually cover a large quantity of energy, transport, communications and other basic provision. But there will never be an unlimited supply of desirable land, unique objects or every form of human attention.

Something would still have to allocate scarce goods. Markets currently perform much of that allocation by giving the richest greater purchasing power, which is hardly a neutral solution. A different society would need other mechanisms and some of them would be irritating. Waiting lists, lotteries or forms of rationing would not disappear simply because production had become highly automated.

What automation could change is the amount of scarcity that is genuine. If a good can be produced abundantly at negligible additional human cost, retaining scarcity simply to preserve its exchange value becomes much harder to justify. We already encounter primitive versions of this problem with digital goods, where reproduction costs close to nothing while ownership allows access to remain restricted. Advanced automation could bring the same contradiction further into the material economy.

Ownership, then, is not an additional question which arrives after automation. It sits at the centre of the whole argument. It is difficult to imagine a society of abundant free provision if the machinery producing that abundance remains controlled by a small collection of corporations and investors. You would end up constructing an increasingly elaborate redistribution system around private ownership. Automated firms would produce much of the wealth and collect the income. The state would tax part of that income and return it to the population, who would then use the money to purchase the output of the same productive system.

There is something increasingly circular about that arrangement. Capital owns the automated means of production and receives the surplus they generate. The state taxes some of that surplus and gives it back to people whose wages have disappeared. Those people then spend the money buying the output produced by capital’s machines. At some stage the question is no longer simply how heavily the owners should be taxed, but why the productive infrastructure on which everyone depends should remain their private property.

Falk and Tsoukalas eventually reach something resembling this position. In what they call the “post-labour limit”, they argue that once automation has become overwhelmingly productive, discouraging further automation no longer makes sense. The problem changes from excessive automation to distribution. Their automation tax therefore gives way to profit taxation and UBI, which recycles some of the income generated by automated capital back towards households.

That is where their economic model stops and the political question really begins. If productive assets can operate with progressively less human labour, why should ownership of those assets continue determining who has the primary claim on what they produce? Taxing the owners and redistributing part of the proceeds may be a politically plausible compromise. But if AI becomes part of the basic productive infrastructure of society, social ownership begins to look less like an ideological addition to the argument and more like its logical conclusion. The alternative is a population increasingly dependent upon transfers in order to buy goods from privately owned machines that no longer require much of that population’s labour.

There is an old socialist argument buried inside this very modern anxiety. Technological development creates the possibility of reducing necessary labour, but private ownership turns that possibility into insecurity because people still require wages to gain access to the wealth society can produce. The worker encounters the labour-saving machine not principally as freedom from work but as competition for the means of subsistence. Capital encounters the same machine very differently: as an asset which can reduce its dependence upon labour while preserving its claim on the product.

That is why the panic over AI and unemployment needs turning around. If these systems prove less transformative than their promoters claim, much of the argument will remain theoretical. If they really do transform production, however, creating replacement jobs merely to maintain the existing relationship between work and consumption would be an extraordinary failure of imagination.

We should want machines to take work from us. There is plenty of work I would be delighted never to see another human being have to perform. The political question is whether the time liberated by those machines belongs to us collectively, or whether it appears first as a saving on somebody else’s balance sheet. That question cannot finally be separated from who owns the machines.

The AI Layoff Trap shows how an economy can become caught between those two possibilities. Individual firms keep reducing labour because competition rewards them for doing it, while the society surrounding them still relies upon wages to distribute purchasing power. The authors propose fixing that incentive. The larger question is whether, if automation proceeds far enough, we should instead begin dismantling the dependence on wages itself and the ownership structure which makes the disappearance of labour a private gain before it can become a social one.

A genuinely successful labour-saving technology should eventually give us less labour. If artificial intelligence cannot produce that result, the failure will not necessarily belong to the intelligence.

Share

Thanks for reading Anti-Capitalist Musings. It is a small operation, and I hope it offers something worth your time. There will be no premium subscriber content here: everything published will remain free to read. If you value these pieces and want to support the writing, buying me a coffee helps fund media subscriptions and the books that keep the analysis grounded. Every contribution, however modest, is genuinely appreciated.