There is something peculiar about the enormous numbers being attached to the future AI economy. The technology is expected to transform almost everything: employment, production, science, education and perhaps the balance of power between nations. Yet the money in the projections seems to inhabit a reassuringly familiar world, in which customers still have budgets, households still have incomes and a trillion dollars means roughly what it means today.
In August, Reuters reported that Anthropic was expected to tell investors its total addressable market exceeded $30 trillion, based on the full scope of work AI could perform. This was not a forecast that Anthropic itself would earn $30 trillion annually, nor a valuation of the company. It was an estimate of the market potentially available to it. The distinction matters, but it does not remove the underlying question: how much of that market will still exist, at those prices, after the technology has transformed it?
The tempting calculation is straightforward. Human beings are paid enormous sums to perform intellectual work. If AI can do that work, some of those sums could instead be paid to AI companies. Capture even a modest fraction and the resulting business becomes spectacularly valuable. Unfortunately, the salaries being counted are also the money with which those human beings buy things. They cannot be removed from one side of the spreadsheet while remaining helpfully present on the other.
Imagine a company that makes washing machines. It replaces much of its administrative, design and sales staff with AI, and eventually automates more of its factory too. At first, the calculation looks excellent. Costs fall, output remains healthy and profits rise. The shareholders are pleased, and the management gives a presentation about its successful transformation, possibly prepared by one of the systems that made the communications department redundant.
Then its competitors do the same. Washing machines become cheaper, because firms can afford to undercut one another. That is a real benefit: producing useful things with less effort is much of the point of technological progress. But meanwhile, other employers are also reducing their wage bills. Some of the people losing their incomes were intending to buy washing machines. Others were planning a new kitchen, a holiday or a house. The cost reduction that looked so attractive inside each company begins to change the market outside it.
This need not cause an immediate collapse. Lower prices stretch the incomes people retain. Investment creates demand. Some people find new work, and some businesses expand enough to employ more staff. But none of that licenses us to assume that today’s wage bill is a vast reserve of purchasing power waiting to be transferred intact to the owners of AI. The act of capturing it changes its size.
I have already written about why professional competence may cease to guarantee employment. The economic consequence deserves equal attention. Retraining might help someone find another job, but finding another job does not necessarily restore their previous income. A redundant analyst who becomes a poorly paid personal assistant still appears in the employment statistics. Their mortgage, spending and tax contributions may tell a rather different story. A government can celebrate successful retraining while presiding over a substantial decline in household purchasing power.
Nor does falling pay necessarily mean that society is producing less. The old analytical work may still be getting done by AI, while the former analyst now provides an additional service. We could become richer in useful output and poorer in the incomes through which people obtain access to it. The hopeful assumption is that prices will fall far enough to compensate. The awkward question is which prices, and for whom.
I first approached this problem in a post about theories of value in 2015, prompted by Paul Mason’s PostCapitalism. I was unconvinced that near-zero production costs would bring money or capitalism to an end. Even if robots produced most things, somebody would still own the fields, mines and attractive places to live. My tentative answer was Georgist: tax land and use the proceeds to provide a basic income. Machines could do the work while everyone received a share of the wealth.
I still think there is much to recommend that approach. But I had quietly assumed that the relevant owners, workers and government would be in the same country. The political argument might be difficult, but at least it would take place within one tax system. I had not given sufficient thought to a world in which the machines working across an entire country belonged to companies in California or Shanghai.
As I argued in The Cloud Has a Flag, access to foreign computing infrastructure is conditional. But there is a separate economic issue even when access remains completely reliable. A domestic salary and a payment to a foreign AI provider do different things to the distribution of income. The first supports a household and domestic tax revenues. The second pays for a useful imported service, but the resulting ownership income may accrue elsewhere. Cheaper intelligence could certainly help domestic firms prosper and export more. It could also undermine the services the country previously exported. There is no guarantee that the two effects will balance conveniently.
A basic income would help people pay their bills, but redistribution at home does not by itself generate the foreign earnings needed to buy imports. A country can draw on export revenues, investment income, borrowing or asset sales. If its valuable exports have been automated elsewhere and it owns little of the new productive machinery, the last two options become rather less reassuring. Selling assets to maintain living standards creates further claims on future income. Eventually, rather a lot of the country may belong to somebody else.
Even my confidence in land needs a qualification. Land will remain useful and some locations will remain scarce, but neither fact guarantees today’s prices. An expensive office district derives much of its value from the businesses and earnings around it. If those disappear, its value can fall too. A government cannot safely assume that the tax base provided by land will remain untouched while the incomes supporting it collapse. The land cannot emigrate, but the reasons for paying so much to occupy it can disappear.
Value is relational. Gold is valuable because people want it and will give up other things to obtain it. A factory is financially valuable because somebody expects to sell its output. Ownership guarantees control; it does not guarantee a profitable customer base. This applies to AI companies as well as washing-machine manufacturers. Their customers’ willingness to pay will depend on what those customers can earn, and on what competing providers charge. We should be sceptical of a future in which competition relentlessly squeezes everybody else’s margins but politely stops at the data-centre door.
This does not mean money must become worthless. If many prices fall, money may buy more. It means that the relationship between money, work and useful goods will change, perhaps radically. Forecasting revenues through such a transition requires assumptions about the resulting economy, not simply estimates of the tasks a model can perform. Technological discontinuity in the sales pitch sits rather awkwardly beside economic continuity in the accounts.
There is also no reassuring rule that the owners must eventually share because otherwise they will have nobody to sell to. They can buy from one another. Production could increasingly serve wealthy households, firms and foreign customers, while much less is produced for people with little purchasing power. Such an economy would disappoint many investors who had assumed a vast consumer market. It would not necessarily stop functioning. A smaller market serving a narrow class of owners is perfectly compatible with considerable deprivation outside it.
That is where the neo-feudal analogy becomes uncomfortable. I have already explored the loss of labour’s bargaining power in The End of the Age of Labour, but the material consequence is stark. People may still have to pay for food, energy and somewhere to live while finding almost nobody willing to pay for their labour. Medieval landlords needed peasants to work the soil. Owners of automated estates might need only machines supplied by other owners. Being unnecessary is not obviously an improvement on being exploited.
Presumably there would be a political reaction. Governments might initially offer training, then subsidies, then increasingly urgent appeals for patience. Eventually, people might elect parties promising to overturn the arrangement, or cease waiting for elections. But what would a revolution actually do? A government can change domestic ownership and taxation. It cannot simply vote itself a share of profits in California, and occupying a local robot depot would not necessarily provide the software, chips and continuing support needed to run it. Banning AI altogether would risk preserving employment by making the country progressively poorer than competitors using it.
There are more promising possibilities: broader ownership of productive assets, public investment funds, domestic and regional computing capacity, usable open models and international agreements over taxation and access. Land taxation still belongs in that discussion. So do basic incomes and public services. But these are institutions to construct, assets to acquire and agreements to negotiate. They are not consequences that emerge automatically when a model passes another benchmark. Establishing them before large numbers of people lose their incomes would be considerably easier than improvising them afterwards.
I would very much like the future described in the happier posts on this site: less drudgery, better food, more time for gardens, travel, scholarship and doing things simply because they are enjoyable. There is no technical contradiction in that vision. The machines might be quite capable of supporting it. The difficulty is ensuring that people have a claim on what those machines produce when their labour no longer earns them one.
Before treating the work of humanity as a $30 trillion market, we should therefore ask what happens to the people currently being paid to do it. They are also the customers whose spending supports the businesses that will supposedly pay for all this artificial intelligence. The washing-machine factory may no longer need their labour. Its business plan still needs their money.