AI Meets Keynes
My General Theory was written primarily for my fellow economists. I titled it a "General Theory" because it offered a systematic account of conditions that classical theory dismissed as special cases. Classical economics took full employment as the natural state of affairs; I argued that underemployment equilibrium is not merely possible, but may persist indefinitely without policy intervention.
Now, ninety full years after the publication of the General Theory, I witness a new technology — AI — rewriting the fundamental meaning of "employment". Classical theorists may argue all technological unemployment is frictional: temporary, local, and self-clearing via market forces. I have never agreed with that proposition. Yet even my General Theory, crafted to address mass involuntary unemployment, assumed labour markets could be repaired through policy.
What if the labour market is broken not by cyclical deficiencies of effective demand, but by a structural erosion of the very need for human labour? This paper addresses that question.
The classical employment function can be written as N=f(Y), where N is total employment and Y is aggregate output. Higher output yields higher employment. This relationship was treated as linear, stable and axiomatic in classical analysis.
AI shatters this linkage. When a single AI model replaces a thousand customer service staff, a hundred translators, fifty junior programmers and twenty copywriters, output does not fall — employment collapses. Output may even rise: models work faster, more reliably, twenty-four hours without shift breaks. The elasticity of N=f(Y) tends toward zero.
In the General Theory, I established that employment is determined by effective demand, shaped by the propensity to consume and the inducement to invest. I criticised the classical school for ignoring the possibility of involuntary unemployment stemming from deficient aggregate demand. Yet even in my bleakest conjectures, I assumed the marginal product of labour — the additional output generated by hiring an extra worker — remained positive and would rise over the long run.
AI creates a scenario I never analysed: the marginal product of human labour remains positive — hire a worker, and they will produce something — but the marginal product of AI far exceeds human labour, with costs approaching zero. When a factory owner faces a choice between hiring ten people or subscribing to an API, where the API costs one per cent of human wages, the owner's choice is not mere greed; it is a survival imperative, for competitors have already made the same switch.
The consumption function is the cornerstone of the General Theory. I laid down a fundamental psychological law: as income rises, consumption rises, but by less than the rise in income. The poor spend nearly all their earnings; the wealthy spend a smaller proportion. It follows that greater income inequality weakens aggregate effective demand.
This psychological law rests on a hidden premise: people still possess income. The problem introduced by AI is not a falling marginal propensity to consume — the affliction of the rich. AI creates a world in which many possess no income at all. When mass labour is displaced by model inference rather than alternative human jobs, we are no longer dealing with weak consumption propensities; we are dealing with a lack of consumption entitlement: people have no earnings to spend.
The worst-case scenario I envisioned in the General Theory — the failure of automatic full employment equilibrium — still assumed a substantial base of effective demand waiting to be stimulated by policy. Yet policy tools to boost demand — tax cuts, transfer payments, public works — presuppose the existence of people capable of receiving the stimulus. What if the core failure is not insufficient willingness to consume, but permanent exclusion from the production process?
I have no doubt a society can sustain mass involuntary unemployment for a prolonged period on purely technical grounds. AI extends that "prolonged" state into permanence. This is neither the frictional unemployment of classical theory, nor my involuntary cyclical unemployment. It is a new category: structurally unemployable labour.
The multiplier is my most famous concept from the General Theory. When the government invests one pound, that pound becomes wages for workers, who spend most of it; the shopkeepers who receive that money spend most of theirs, and so on — total income rises by a multiple of the initial outlay. This is the fiscal engine of demand expansion.
What is the multiplier effect of AI investment? I fear it tends toward unity.
A government or large corporation invests a hundred billion dollars in AI infrastructure. That hundred billion flows to GPU manufacturers, power utilities, and the salaries of a small cohort of highly paid AI researchers. Unlike investment in roads or bridges, this money does not ripple broadly through ordinary households. GPUs are produced in highly automated semiconductor fabs with minimal unskilled labour. AI researchers do not circulate their high wages widely through barbershops, restaurants and cinemas — not out of reluctance, but because their numbers are too thin to sustain broad local economic circulation.
The traditional multiplier, which I estimated at between two and three, relies on new income passing rapidly and widely through ordinary households. Income streams from AI investment bypass ordinary families entirely. Capital flows from tech firm to tech firm, from sovereign wealth fund to GPU vendor, from cloud platform to cloud platform, with very little leakage into household consumption. This is no multiplier; it is demand leakage.
In the General Theory, I analysed the liquidity trap: when interest rates fall to a threshold where all market participants expect rates to rise (bond prices to fall), the demand for money becomes infinitely elastic. No matter how much currency the central bank prints, it is hoarded as idle cash rather than channelled into investment or consumption. Monetary policy becomes powerless.
The AI economy has a parallel phenomenon I term the employment trap. When the cost of substituting human labour with AI is sufficiently low, employers have no incentive to hire people, no matter how low wages fall. Human labour never costs zero. Even if the legal minimum wage were set to nil, human labour carries overhead: management, training, error risk, social contributions, leave absences.
In a liquidity trap, all injected liquidity is absorbed as idle balances. In an employment trap, no wage reduction can incentivise hiring, because the marginal cost of not hiring is lower than the marginal cost of hiring. Traditional monetary expansion and fiscal stimulus pour resources into a bottomless void under these conditions.
Near the penultimate chapter of the General Theory, I penned a line often quoted out of context: the euthanasia of the rentier. My meaning was this: capital would become so abundant that owners of capital could no longer extract scarcity rents; interest rates would fall toward zero, and those who live off unearned capital income would fade away quietly.
AI reverses this entire dynamic. Capital, embodied in GPU clusters, trained proprietary models and exclusive datasets, has never been scarcer. We face not the euthanasia of capital, but the expulsion of labour from the production process.
In my original framework, rentiers fade when capital scarcity vanishes — when savings outstrip viable investment outlets. Today AI eliminates the scarcity of cognitive labour, while cognitive capital — the ownership of GPU fleets and proprietary datasets required to train state-of-the-art models — becomes the scarcest, most concentrated and most powerful means of production in human history.
This is not a problem of overabundant capital. It is a distortion where one specific form of capital attains extreme market value, while all other forms of capital — including human capital — are devalued. This is a species of inequality I never analysed in my original work.
If the General Theory yields a single core policy conclusion, it is this: when private investment fails to fill the gap between savings and full-employment output, public investment must step in. This is the ultimate justification for fiscal intervention.
The gap in the AI age is no longer excess savings and insufficient investment. The gap is the secular decline in the market value of human labour. This runs deeper than underinvestment, for it is structural rather than merely a demand-side failure.
I have no doubt the state must play a far larger role in the AI economy than in any prior era — not necessarily as a mass employer, but as the primary redistributor. When labour ceases to be the source of income for most of the population, the flow of earnings must be reshaped via collective public decision-making.
I defended prudent macroeconomic management all my life, neither an enthusiast for total central planning nor a devotee of laissez-faire. Yet let me state plainly here: AI breaks the classical chain labour → income → consumption at its root. If you accept this chain — as I did in the General Theory — as the core mechanism of economic circulation, you must accept that irreversible technological breakdown of this chain implies an economic order entirely alien to the world described in my original work.
I closed the General Theory with a passage more famous than all my technical analysis combined: The ideas of economists and political philosophers, both when they are right and when they are wrong, are more powerful than is commonly understood. Indeed the world is ruled by little else. Practical men, who believe themselves to be quite exempt from any intellectual influences, are usually the slaves of some defunct economist.
Today all of us shaping AI policy — technologists, entrepreneurs, regulators alike — remain captive to an outdated economic intuition: labour generates income, income generates consumption, consumption generates employment, in perpetual circulation. This cycle is no natural law; it is a product of a specific technological epoch.
When you witness AI displacing customer service, translation, design and programming, you cannot merely repeat "new jobs will emerge". That claim may hold for certain past technological shifts, but rigorous proof is required: what new occupations rely uniquely on human labour rather than AI, and can they scale large enough to absorb hundreds of millions of displaced workers? I have seen no such rigorous demonstration. What I observe is mechanical extrapolation from past patterns — the very habit of thought the entire General Theory sets out to refute.
I warned that the difficulty lies not in new ideas, but in escaping old ones. The oldest and most damaging idea we must abandon in the AI age is this: labour markets are self-correcting. They are not.