Post-AGI Economics: What Happens to Wages
4 min read · updated August 3, 2026
Predictions about wages after transformative AI range from “broadly unchanged” to “approaching zero”. The range is not mostly about how capable AI gets. It is about two or three specific parameters in the same underlying model, and it is easier to follow once you can see which one a given argument is disputing.
What sets a wage
In a competitive labour market, a worker’s wage tends toward the value of what they add at the margin. An employer pays up to that and no more, because above it hiring loses money and below it a competitor bids. So the question “what happens to wages” is the question “what happens to the marginal product of human labour”, and that depends on which tasks humans do and what the alternatives cost.
The framework most of this literature uses is task-based, developed by Daron Acemoglu and Pascual Restrepo. Production is a set of tasks. Each can be done by labour or by capital. Automation moves tasks from the first column to the second, which has three effects that pull in different directions:
- Displacement. Automating a task removes demand for the labour that did it. Pushes wages down.
- Productivity. Cheaper output means more of it, and more demand for whatever labour remains complementary. Pushes wages up.
- Reinstatement. New tasks appear that only humans do, historically the largest offsetting force. Pushes wages up.
Historically these have roughly balanced over long periods, which is why two centuries of automation did not produce permanent mass unemployment. Every serious disagreement about the AI case is a disagreement about whether they still balance.
The comparative advantage argument
The most common reassurance is comparative advantage, and it is a real theorem rather than a hand-wave. Even if machines are absolutely better at every task, gains from specialisation exist as long as their relative costs differ: a machine that is a thousand times better at engineering and twice as good at gardening should be doing engineering, leaving gardening to humans. Trade is mutually beneficial regardless of absolute advantage.
The theorem holds. What it delivers is the conclusion that humans still have something to do. It says nothing about what that work pays.
Where the floor actually comes from
This is the step most discussions skip, and it is where the real disagreement lives. Comparative advantage tells you the pattern of specialisation. The price of human labour in the remaining tasks is set by what it costs to do those tasks the other way. If an AI system plus its compute and energy can do a task for some cost per hour, nobody pays a human more than that for it, no matter how the tasks are allocated.
So the human wage floor is tied to the cost of the substitute — which means to the cost of compute, energy and whatever physical capital the task needs. That reframes the question usefully. If the alternative remains expensive because compute and energy are scarce, human labour retains value in the tasks where it is cheaper. If the alternative becomes very cheap and scales without limit, the ceiling on human wages falls with it.
This is the point of the horse analogy, and also the point at which the analogy is usually pushed too far. Horses lost their labour market because a substitute became cheaper at every task horses did, and horses could not move into new ones. The disanalogy is that humans are also the consumers, the voters and the owners, and the horse story has nothing corresponding to that. The analogy illustrates the wage mechanism and predicts nothing about the rest.
What the growth models say
Economists have modelled this and reached different conclusions from different assumptions, which is exactly what makes the assumptions worth reading.
| Work | Description |
|---|---|
| Aghion, Jones & Jones | 'Artificial Intelligence and Economic Growth'. Applies Baumol's cost-disease logic: if some tasks resist automation, they become an ever-larger share of cost and the overall growth rate is governed by them, not by the automated part. Explosive growth requires nearly complete automation, including of the production of ideas. |
| Nordhaus | 'Are We Approaching an Economic Singularity?'. Proposes supply-side and demand-side tests that could detect an approaching acceleration in existing data, and treats the question as empirically checkable rather than purely theoretical. The method — say in advance what the data would look like — is the durable contribution. |
| Acemoglu & Restrepo | The task framework itself, plus the argument that the tax treatment of capital relative to labour already tilts firms toward automating tasks where the productivity gain is small. On this view the distribution of gains is partly a policy choice rather than a technological inevitability. |
| Korinek & Stiglitz | Focus on distribution rather than aggregate growth: innovation can raise total output while making a large group worse off, and whether it does depends on redistribution that does not happen automatically. |
| Hanson | 'The Age of Em' takes the extreme case seriously — if labour itself becomes reproducible at the cost of copying, wages tend toward the cost of running a copy. An explicit exploration of an assumption, not a forecast, and useful for showing what that assumption implies. |
The bottleneck assumption
Almost every disagreement above reduces to one question: are there tasks that resist automation, and how large a share of production do they hold?
Candidates that get proposed, with the reason each is contested: physical work in unstructured environments, where robotics has moved slower than software; tasks where a human is legally or contractually required, which is a policy variable rather than a technical one; work whose value depends on being done by a person, such as care and some performance, where demand is a matter of preference and preferences change; and the production of genuinely new ideas, which is precisely what is in dispute.
Note the structure of the disagreement. If bottlenecks are durable, the cost-disease logic applies and wages in the bottleneck tasks may rise substantially. If they are not, the models that predict very large effects apply instead. Nobody can settle it now, and the useful move is to notice that a given forecast is an assumption about bottlenecks wearing a conclusion’s clothing.
Wages are not income
One clarification that removes a lot of confused argument: wages are the return to labour, and income also includes the return to capital. In a scenario where labour’s share of output falls, total output can rise while the share flowing to people who own nothing falls. Whether that produces widespread hardship depends on ownership and on redistribution, not on the production technology.
That is why the wage question and the welfare question come apart, and why the policy discussion moves quickly to transfers — taken up in UBI and automation. It is also why some researchers argue that income transfers address only half of the issue: the political leverage that comes from labour being needed is not restored by a payment, which is the argument in gradual disempowerment.
Everything on this page is conditional. Whether the labour-substituting capability arrives, and when, is a prediction on which the field disagrees; the value of the framework is that it tells you what to watch — the price of the substitute and the size of the bottleneck — rather than which scenario to believe.