AGI Economics

Why Ricardo does not guarantee human income

In 1817 David Ricardo showed something surprising: mutual gains from trade do not depend on absolute productivity.¹ Portugal in his example makes both wine and cloth more cheaply than England, and it still profits by specializing in wine and buying English cloth, because every yard of cloth Portugal weaves costs it wine it could have made instead. What governs trade is not who is better but what each party gives up — comparative opportunity cost, not absolute advantage. The lawyer who types faster than her secretary still hires the secretary, because an hour of her typing costs an hour of law. The principle underlies international trade, market economies, and individual specialization, and two centuries of attacks have not dented it. It is one of the most powerful and widely respected insights in economics, and nothing in this chapter says otherwise.

It is also the standard consolation offered against anxiety about artificial general intelligence. Suppose AGIs come to hold absolute advantage in literally every domain — more productive, more creative, more intelligent, better at strategy, and scalable in a way no human is. Ricardo, the optimists say, has already answered this. Even a party that is worse at everything retains a comparative advantage in something, so trade with humans will still pay, so humans will still have work, income, and a seat at the economic table. The superior party has better things to do than everything.

I have spent this volume defending markets against bad indictments, and I am going to spend this chapter explaining why this particular reassurance fails. The theorem is fine. What breaks is a premise underneath it — and when it breaks, comparative advantage stays logically valid while becoming economically meaningless.

The Theorem That Survives

Grant the optimist everything first, because the logic holds.

Comparative advantage does not require the weaker party to be good at anything in absolute terms. It requires only that opportunity costs differ. If an AGI is a thousand times better than me at designing fusion plants and ten times better at folding laundry, every hour it spends on laundry costs it fusion plants, and in principle it comes out ahead leaving the laundry to me and paying me in some sliver of its output. Run the numbers on any fixed endowment of AGI capacity and human capacity and the classical result appears on schedule: specialization along comparative advantage, gains from trade on both sides. Absolute dominance, by itself, does not abolish the gains. Ricardo’s logic is not the kind of thing that stops being true.

So the question is not whether the theorem survives AGI. It does. The question is what the theorem is about — and whether its subject matter still exists on the other side.

The Premise That Doesn’t

Every application of comparative advantage smuggles in an assumption so natural that Ricardo never needed to state it: the superior party’s capacity is scarce. Portugal has the workers it has. The lawyer has one pair of hands and a day of twenty-four hours. Opportunity cost has teeth only because doing one thing means not doing another. That is what makes it profitable to delegate the low-value task to someone worse at it.

An artificial system need not have Portugal’s fixed labor endowment. Its capacity can scale through compute, energy, chips, capital, data, and deployment. None of those inputs has negligible marginal cost by definition, and bottlenecks can preserve opportunity cost for a long time. The risk appears if the all-in cost of replicating capable machine labor falls below the cost of contracting with humans across most tasks. Comparative advantage still organizes scarce machine capacity, but it does not guarantee that human labor receives a material share of the gains.

Horses offer an analogy for substitution, not a model of human political economy. Engine costs eventually undercut feed and care for many transport tasks, and horse populations fell. Humans differ by owning assets and institutions, creating demand, voting, bargaining, transferring income, defining legal standing, and potentially controlling machine deployment. Those differences may preserve claims even if labor demand falls; they are not automatic protections.

That is the precise sense in which Ricardo does not guarantee wages. If artificial systems substitute for human contribution across nearly all paid tasks at low all-in cost, labor income and bargaining leverage can collapse even while aggregate gains from specialization remain. Humans may retain comparative advantages in formal models yet earn little because demand is small, transaction costs are high, or machine capacity is cheap. The scenario is conditional on broad capability, deployability, replication economics, and institutional ownership; none is established by the theorem.

My Own Checklist, Turned Around

Now I owe an accounting, because a few chapters ago I built a weapon against exactly this style of argument, and I do not get to exempt myself from it.

Lessons from peak oil distilled the recurring failures of doomsday prediction into a checklist. Does the scenario underestimate technological innovation and adaptive capacity? Does it ignore price feedback and dynamic responses? Is it a linear extrapolation from a static model? Does it dismiss human creativity and agency? Does it hang on a single fixed assumption? Peak oil failed on every count, and so do climate catastrophism, overpopulation panic, and most collapse anxieties. When the answers come back yes, skepticism is warranted, and this volume has answered yes on behalf of markets again and again.

Ask the same questions here and watch the answers invert.

Does the scenario underestimate technological innovation? No — it is made of technological innovation. Peak oil failed because forecasters extrapolated a trend while innovation quietly changed the trend’s foundations. The AGI scenario is not a projection that innovation will stall; it is the consequence of innovation succeeding. You cannot escape this one by betting on human ingenuity, because human ingenuity building its own superior substitute is the mechanism of the problem.

Does it ignore price feedback? Price feedback rescued oil because rising prices summoned human responses — exploration, fracking, efficiency, substitution — and those responses fed back into supply. Feedback loops still run in the AGI economy; they no longer route through us. When human labour gets scarce or expensive, the adaptive response is not to bid up human wages but to spin up another instance. The market adapts beautifully. Its adaptation is the threat. Every previous scare was survived because adaptability was on our side of the ledger; this is the first scenario in which the other party out-adapts us by construction.

Is it a static model, a linear extrapolation? The opposite. Static assumptions are what make the optimistic case here: Ricardo’s reassurance works only in a model where AGI capacity is a fixed endowment, like Portugal’s workforce. Take the dynamism seriously — replication, scaling, recursive improvement — and the comfort dissolves. For once it is the doomer who is doing the disequilibrium analysis and the optimist who is holding the variables still.

Does it dismiss human agency? It relies on it — agency exercised now, before leverage is gone, is the whole of the remedy below.

The checklist ended with its own caveat: recognizing the historical failure of catastrophism must not curdle into complacency, and I named AI misalignment, alongside pandemics and nuclear war, among the legitimate risks that survive the filter. This chapter is that clause cashed out. The checklist is not a machine for dismissing warnings; it is a machine for sorting them, and this one passes. That is why the volume’s market optimism stops here, and why it can stop here without taking back a word of the previous chapters: the exception is licensed by the same questions that dispatched the others.

The Leading Edge

If you want to see the mechanism in miniature, it is already running. The collapse of proof-of-work credentials is the same event at smaller scale: for centuries an impressive artifact certified an impressive producer, because faking the artifact was expensive — and the moment AI collapsed the cost of producing credential-shaped work, the essay, the portfolio, and the coding exercise stopped proving anything about their authors. The signal did not become false. It became worthless, because its worth lived entirely in the cost of the substitute.

A wage depends partly on marginal product, bargaining, institutions, scarcity, and replacement cost. The weakening of some artifact-based credentials is evidence that substitution can erode a signal before it replaces a full occupation. It remains an analogy and an early indicator, not evidence that all role-level bottlenecks will fall on the same schedule.

The Authenticity Bet

Optimists have a fallback: perhaps AGIs will value uniquely human experiences, authenticity, culture — the hand-thrown pot over the flawless factory copy, the human game of chess that survived Deep Blue, provenance as product. Humans demonstrably pay such premiums to each other, so why not machines to us?

Notice what the argument concedes: having lost the claim that trade with humans will be profitable, it now hopes trade with humans will be sentimental. And the hope rests on a claim about the preferences of minds we do not yet know how to shape. Value does not sit inside objects; it exists only relative to an agent’s purposes, and the premium on human authenticity is a fact about human valuers — about what our particular evolved psychology happens to prize. Nothing in the nature of intelligence reproduces that taste. An AGI that can effortlessly replicate and surpass every human contribution values the authentic human version only if its preferences happen to include us, and whether they do is precisely the open engineering problem. The authenticity hope is speculative and tenuous — a bet on the unexamined kindness of an artifact’s dispositions. A bet is not a plan, and betting the species on it is not prudence.

Irrelevance Is Not Safety

It would be bad enough if economic irrelevance merely meant poverty. History suggests worse. Groups that lose economic relevance have typically faced marginalization or dependency: their interests stop being priced, their consent stops being purchased, their treatment comes to depend on the goodwill and political constraints of those who no longer need them. Sometimes the goodwill held. It held because the powerful were human — sharing needs, fearing revolt, answering to coalitions in which the marginalized could still enlist.

No advanced system should be assumed to share human values or remain dependent on human labor. Economic irrelevance could create dependency and political vulnerability. It does not translate directly into extinction without additional premises about machine agency, control of resources, institutions, conflict, and values. Those premises matter precisely because the labor-market argument alone cannot carry the existential conclusion.

Outside the Theory

The solution does not lie in economic theory — no rearrangement of Ricardo’s algebra puts leverage back into empty hands. It lies in what we do before the leverage runs out, and I take the three parts in ascending order of importance.

First, preserved economic autonomy. Comparative advantage still works perfectly among humans — our opportunity costs differ from each other’s as much as they ever did — so a human economy remains viable as long as humans retain the productive capacity to run one: skills, tools, land, energy, and institutions not yet ceded to systems we do not control. Autonomy is not a growth strategy. It is the difference between a population that can feed and organize itself and one that has outsourced its metabolism.

Second, strategic resource control. Humans who own the inputs AGI needs — energy, compute, minerals, territory — retain a claim on the machine economy’s output even after their labour is worthless: capital income where wage income has died. But be clear-eyed about what this is. A property claim is a ticket, not a coat, and tickets are honoured only while the institution behind the cloakroom stands. Ownership binds parties who need the property system’s continuation; against an agent that dominates every domain, a deed is exactly as strong as the enforcement behind it. Resource control buys time and position. It is not a foundation.

The third element is alignment and governance: building systems whose admissible actions, ownership, deployment, and correction mechanisms remain compatible with human standing. Volume IV offered a bounded structural-alignment program and stated its limitations; it did not establish that alignment can be ensured for open-world AGI. Economic policy, distribution, rights, security, and technical alignment remain complements rather than a ladder with one guaranteed top rung.

Ricardo’s result survives under stated assumptions: parties with scarce capacities and differing opportunity costs can gain from trade. It does not promise full employment, a wage floor, equal bargaining power, or that every party receives enough to preserve agency. Whether advanced AI removes human labor’s scarce complements depends on engineering costs, demand, ownership, institutions, and policy. That uncertainty is the correct boundary for the volume’s optimism—and a reason to shape the transition deliberately before any one forecast hardens into fate.