Output Without Income

AI may not collapse the economy. It may keep production running while breaking the social bargain that made work legitimate.

The usual fear about AI is that it will replace workers. This is understandable, but too narrow. Replacement is only the most visible version of a deeper transformation. The more important possibility is that AI allows society to keep producing, and even to produce more, while weakening the channels that once connected production to income, status, training, and belonging.

This is output without income.

Output without income is not mass unemployment. It is a political-economic condition in which production remains functional while the labor-based mechanisms that once distributed income, dignity, and social recognition lose force.

It does not describe a world where everything collapses. It describes something more politically dangerous: a world where the surface indicators remain functional while the human bargain underneath them deteriorates. Companies still operate. Consumers still receive services. Software improves. Margins expand. Productivity rises. Markets may even reward the transition.

The optimistic version of this story is already familiar. Goldman Sachs Research has estimated that generative AI could raise global GDP by roughly 7 percent and lift productivity growth by 1.5 percentage points over a ten-year period. [1] That estimate does not prove abundance will be widely shared. It makes the paradox visible. The most destabilizing AI scenario is not necessarily one where production fails. It may be one where production succeeds while the social mechanism that distributed its rewards weakens.

That is why the AI transition cannot be understood only as a labor-market shock.

It is a legitimacy shock.

The modern economy did not merely use work to produce things. It used work to organize social meaning. Employment gave people a recognized claim on income. It gave them a place inside the system. It linked discipline to reward, contribution to dignity, adulthood to a role, and economic participation to political inclusion.

The wage was not just compensation. It was a social explanation.

Work did not merely distribute income. It explained why income was deserved.

A person with a job could say: I work, therefore I deserve income. I contribute, therefore I belong. I have a role, therefore I have standing.

This explanation was never perfectly true. Many people worked without dignity. Many contributed without being rewarded. Many were excluded from stable employment altogether. But the structure was legible enough to hold a society together. Work acted as a bridge between the productive system and the moral claims of the individual.

AI weakens that bridge.

The Economy Can Win While the Household Loses

The strongest objection to this argument is that technological panic is old. Every major wave of automation produces fear, and every time, new industries eventually appear. The spreadsheet did not eliminate finance. The internet did not eliminate commerce. Software did not eliminate work. Why assume AI is different?

The answer is not that AI must destroy all jobs. It does not need to. A system does not have to remove workers completely in order to reduce their power. It only has to make large categories of labor less scarce.

Once fewer people are needed to produce the same amount of acceptable work, the bargaining position of workers changes even before unemployment appears in dramatic form. A company may not fire an entire department. It may simply stop hiring at the same rate. It may delay replacing people who leave. It may expect one employee to do work that previously required several. It may use AI to compress junior tasks into senior workflows. It may preserve the appearance of a profession while quietly narrowing the path into it.

The IMF has warned that around 60 percent of jobs in advanced economies may be affected by AI, with some roles complemented and others facing substitution pressure. It also notes that if AI disproportionately complements higher-income workers, it may raise their labor income while productivity gains boost capital returns, both of which could worsen inequality. [2]

This distinction is critical. AI does not distribute its gains evenly merely because it raises output. It changes who is complemented, who is substituted, and who owns the systems that capture the surplus.

That is the first step toward output without income.

People may still participate in the economy, but with weaker claims on its rewards. The transition does not need to announce itself as collapse. It can arrive as a slow deterioration in leverage.

There will be exceptions. Some workers will become more valuable because AI amplifies their judgment. Some firms will expand rather than cut. Some new roles will appear around orchestration, evaluation, compliance, coordination, and synthesis. The pattern will not appear evenly. It will vary by sector, country, firm size, regulation, and the speed at which institutions redesign work around AI.

But the existence of winners does not settle the deeper question. The issue is not whether some new tasks appear. The issue is whether the economy can still produce enough stable, legible roles through which ordinary people can claim income, status, and adulthood.

Reskilling can answer a task problem. It cannot by itself answer a legitimacy problem.

The First Thing AI Breaks Is the Ladder

The best place to see the problem is not at the top of the professional hierarchy, but at the bottom.

Imagine a junior analyst entering a consulting firm, bank, startup, or corporate strategy team. Much of the first year is not glamorous. The work involves summarizing documents, cleaning slides, formatting spreadsheets, drafting memos, preparing market scans, and producing first-pass analysis that a senior person will correct. From the outside, this looks inefficient. From the inside, it is how the institution trains judgment.

The junior analyst learns what matters by doing work that is not yet fully trusted. She learns how senior people frame problems, what clients ignore, what numbers are suspicious, what language is too strong, what conclusion is unsupported, and how an argument changes when it moves from internal discussion to external presentation.

AI can now do much of that first-pass work. Not perfectly, but often well enough to change the hiring equation.

A partner may still need analysts. But perhaps not as many. A manager may still want human judgment. But perhaps not for the first draft. A firm may still value training. But training becomes harder to justify when the machine can perform many of the tasks that used to make training economically useful.

This is the ladder problem.

Many professions are not only collections of tasks. They are reproduction systems. Entry-level work trains judgment. Junior roles create senior workers. Repetition builds tacit knowledge. Organizations renew themselves by letting inexperienced people perform lower-stakes work until they become capable of higher-stakes work.

AI threatens to remove or compress precisely the tasks through which competence is formed.

Early evidence points in this direction. A Stanford Digital Economy Lab working paper using high-frequency payroll data found that since the spread of generative AI, early-career workers aged 22 to 25 in the most AI-exposed occupations experienced a 16 percent relative employment decline, while less-exposed workers and more experienced workers in the same occupations remained more stable. [3] The importance of this evidence is not that it settles the entire labor-market question. It does not. The importance is that it matches the mechanism: the first damage appears where work is most exposed, junior, and task-compressible.

That means the early AI shock may not look like mass unemployment. It may look like a thinning of ladders. Senior professionals become more productive, but fewer juniors are trained. Companies still demand experience, but fewer institutions produce it. A profession remains prestigious from the outside while becoming harder to enter from below.

The canopy remains green while the seedlings disappear.

The labor market is only the surface. The deeper damage is institutional reproduction.

Unnecessariness

The psychological force of AI anxiety is often misread. People are not only afraid of having less money. They are afraid of becoming unnecessary inside a system that still works.

Consider the same junior analyst three years later. She has not been fired. In fact, her firm is doing well. The AI tools are impressive. The partners praise productivity. Clients receive cleaner decks faster. The company no longer needs as many first drafts, first scans, first models, or first memos. Her workday is not empty, but it is thinner. She spends less time learning by doing and more time supervising outputs she did not fully create. She is told this makes her more “strategic,” but she senses something else: the system is using her less as a developing mind and more as a liability wrapper around machine work.

Her anxiety is not simply that she might lose her job. It is that the job no longer means what it was supposed to mean.

It no longer guarantees apprenticeship. It no longer guarantees a path upward. It no longer proves that the institution is investing in her future. She is employed, but the bridge between employment and becoming is weaker.

That is unnecessariness.

Unnecessariness is not the same as unemployment. It is not even the same as exclusion. Exclusion still implies a border one might cross. Unnecessariness is more corrosive because a person may remain formally inside the institution while the institution stops needing the process by which that person becomes valuable.

It can look like employment without apprenticeship. Participation without leverage. Credentials without entry paths. Entrepreneurship without control. Citizenship without productive centrality. Income without recognition.

Ordinary poverty is deprivation. Temporary unemployment is a loss of position. Unnecessariness is more destabilizing because it suggests that the system no longer requires your development, your discipline, your credentials, your loyalty, or your future.

This matters because labor was never only a production input. It was a recognition system.

In an industrial economy, even exploited workers occupied a necessary symbolic position. Production visibly depended on them. The factory, office, hospital, school, logistics network, bank, law firm, studio, and government agency all required human labor not only at the margins but near the center. That necessity gave workers political weight. It did not guarantee justice, but it made exclusion legible as a conflict. The worker could be underpaid, overworked, ignored, or abused, but the worker could still say: the system needs me.

AI changes that symbolic position.

It does not remove humans from the system. It changes how necessary many humans appear to be. That appearance matters because political legitimacy depends not only on what people do, but on whether their contribution is recognized as necessary.

A society can tolerate inequality more easily when people believe there is still a path into usefulness. It becomes more unstable when large numbers of people suspect that the system is learning to function around them.

The economy does not have to say, “you are worthless.”

It only has to say, “we can do this without you.”

Growth Is Not a Social Contract

Traditional economic language misses this because it treats wages mainly as payment for labor rather than as a recognized claim on output.

When AI reduces the need for human labor in parts of the production process, it also weakens the moral story that attached income to contribution. The problem is not only that some people may earn less. It is that the basis on which they claim income becomes less obvious.

This is where both extreme versions of the AI debate fail.

The techno-pessimist imagines a clean collapse: mass unemployment, social breakdown, machines replacing humans everywhere. The techno-optimist imagines a clean adjustment: productivity rises, new jobs appear, everyone eventually moves into higher-value work. Both frames are too simple.

Daron Acemoglu’s skeptical work is useful here precisely because it complicates the optimist’s escape. In The Simple Macroeconomics of AI, he argues that near-term productivity effects may be more modest than headline forecasts suggest, constrained by the actual share of tasks affected and the real cost savings achievable at task level. [4] If he is right, the AI transition produces disruption without the compensating productivity windfall that optimists expect to fund adjustment. If he is wrong and the productivity gains are large, distribution becomes the central question anyway. The two scenarios converge on the same problem from opposite directions.

A society does not only need tasks for people to perform. It needs roles that people can inhabit. AI may create new tasks, but tasks are not the same as careers. It may create productivity, but productivity is not the same as belonging. It may create new markets, but markets are not the same as a social contract.

That is why “people will reskill” is too small an answer.

Some reskilling will work. Some workers will become more productive. Some new roles will appear. Some companies will use AI to expand capacity rather than reduce headcount. The point is not that adaptation is impossible. The point is that adaptation cannot carry the entire legitimacy burden.

Reskilling asks whether individuals can adjust to new tools. Output without income asks whether the system can still distribute income, dignity, and stability when labor becomes less central to production.

Those are not the same question.

The Stack Is Where the Surplus Goes

When the wage-as-explanation weakens, the surplus does not dissolve. It migrates into whatever sits upstream of labor.

That is the stack.

AI is not only a tool used inside the economy. It is becoming part of the infrastructure through which economic action is organized. The more important terrain is the layered system of compute, chips, energy, cloud, models, data, distribution, payments, identity, procurement, and permission.

The stack matters not because it contains many components, but because it creates choke points.

A startup may build a useful product, but its margins depend on upstream model pricing. If the model provider changes API costs, throttles access, launches a competing feature, or alters safety rules, the startup’s economics change overnight. A creator may have talent and audience, but distribution depends on ranking systems she does not control. A hospital may want AI assistance, but procurement rules, liability frameworks, and vendor contracts determine what can actually be used. A government may announce sovereign AI, but chips, energy, data centers, and model expertise define the real boundary of that sovereignty.

This is what makes the stack political. It allocates opportunity before formal politics arrives.

If labor was the old bridge between output and claim, the stack is becoming the new bridge between capability and access.

The experience of this from below is not oppression. It is convenience. The interface improves. Access expands. Costs fall. Capabilities multiply.

This is the rise of rented intelligence environments: economic spaces where individuals and firms appear independent, but their cognition, distribution, compliance, payments, and customer access are leased from upstream systems. A startup builds on someone else’s model. A creator reaches customers through someone else’s ranking system. A seller depends on someone else’s payment rules and advertising auction. A professional increasingly works through tools whose capabilities, prices, permissions, and limits are set elsewhere.

Their formal status may be entrepreneurial.

Their practical condition is tenancy.

The old labor question was: who employs whom?

The stack question is: who controls the conditions of access?

These are not the same question, and the political institutions built to answer the first are poorly equipped to answer the second.

Stanford’s 2025 AI Index makes the material stakes visible: corporate AI investment reached $252.3 billion in 2024, with private investment climbing 44.5 percent and U.S. private AI investment reaching $109.1 billion, nearly twelve times China’s $9.3 billion and twenty-four times the U.K.’s $4.5 billion. [5] This is not a neutral diffusion of intelligence across society. It is a capital-intensive buildout of infrastructure. The more expensive the stack becomes, the more important ownership and access become.

The stack makes dependency feel like a product feature.

The Three Fracture Points

The mistake is to ask for a policy menu too early. The first task is to identify where the bargain is breaking.

Three fracture points matter: the ladder, the stack, and the claim.

The first fracture is the ladder. If AI compresses the junior tasks through which competence is formed, the institutional pipeline that produces future expertise degrades silently. This is not a personal problem for young workers. It is a civilizational maintenance problem. Societies that allow the seedling layer to thin while the canopy remains green are not preserving their institutions. They are consuming them.

The question is not whether firms can still produce today’s work with fewer juniors. Many can. The question is who produces tomorrow’s seniors when the economic logic of apprenticeship weakens. A civilization that automates away the training ground should not be surprised when it later discovers a shortage of judgment.

The second fracture is the stack. If intelligence becomes infrastructure, then access to compute, models, cloud capacity, distribution rails, identity systems, and data becomes a condition of economic participation. At that point, ownership of the stack becomes a constitutional question even when the infrastructure is formally private.

This does not require imagining some distant machine empire. A payment network is not just another private product once merchants cannot realistically participate in commerce without it. A cloud platform is not just another vendor once startups build their cost structure, compliance posture, and distribution logic on top of it. AI infrastructure is moving toward that category: formally private, but functionally constitutive of participation. The question is no longer only whether the market is competitive. It is whether the conditions of access to economic agency are themselves privately governed.

The third fracture is the claim. If labor is no longer the central mechanism through which people receive income, societies will need another explanation for why people have a rightful share in output.

This is where the income debate often becomes evasive. UBI may become necessary. Wage subsidies may become necessary. Public employment, expanded services, or hybrid income floors may become necessary. But necessity is not legitimacy. A transfer can prevent destitution without explaining belonging. If people experience income as compensation for being economically unnecessary, resentment will not disappear. It will change form.

The obstacle is not only fiscal. It is moral. People raised inside a work-based order will resist any income form that feels like payment for non-necessity. A post-labor claim has to be narrated not as pity, compensation, or pacification, but as a share in the inherited infrastructure that made synthetic productivity possible.

The political problem is not only how to fund income. It is how to narrate claim.

Some portion of income in an AI-shaped society may have to be justified by membership rather than employment. That does not mean abandoning contribution. It means admitting that contribution itself has always depended on inherited systems: public science, public law, public education, public infrastructure, accumulated data, stable institutions, and the civilizational substrate that made private innovation possible.

A claim on AI-generated surplus can therefore be framed not merely as redistribution, but as a dividend on inherited civilization.

The civilization that cannot say this clearly will say it badly — through populism, through resentment, through redistribution without rationale, through moral panic disguised as policy.

None of this is a complete program. It is a map of where pressure will build until something breaks or something is built.

The Economy Stops Explaining Why People Belong

The first task of AI political economy is not to predict the exact number of jobs lost. That number will vary by sector, country, class, and time horizon. The deeper task is to understand what happens when labor loses its role as the main bridge between output and legitimacy.

Once that bridge weakens, every society faces a harder question.

Not whether it can produce enough.

Whether people can still belong to a productive system that needs them less than before.

This is the emerging political shape of the AI economy. The issue is not automation alone, or unemployment alone, or inequality alone. It is the relocation of economic power from labor-mediated institutions into infrastructure-mediated allocation systems.

The supermarket stays stocked. The interface improves. The delivery arrives. The company scales. The model becomes more capable. The investor presentation looks better. The state reports growth.

But the individual’s relationship to the system becomes thinner, more conditional, and harder to narrate.

The economy does not collapse.

It stops explaining why people belong.

The lights stay on. The shelves stay full. The dashboards turn green.

And still, a growing number of people may stand outside the machine, watching it run, unsure what claim they still have on the world it produces.

That is output without income.

---

Sources

[1] Goldman Sachs Research estimated that generative AI could raise global GDP by 7 percent and lift productivity growth by 1.5 percentage points over a ten-year period.
https://www.goldmansachs.com/insights/articles/generative-ai-could-raise-global-gdp-by-7-percent

[2] The IMF argued that AI could affect around 60 percent of jobs in advanced economies and warned that AI may raise capital returns and disproportionately benefit higher-income workers if it complements them more than lower-income workers.
https://www.imf.org/en/blogs/articles/2024/01/14/ai-will-transform-the-global-economy-lets-make-sure-it-benefits-humanity

[3] Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen’s Stanford Digital Economy Lab working paper finds that early-career workers aged 22–25 in the most AI-exposed occupations experienced a 16 percent relative employment decline after generative AI adoption, while less-exposed and more experienced workers remained more stable.
https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/

[4] Daron Acemoglu’s The Simple Macroeconomics of AI argues that AI’s macroeconomic effects depend on the share of tasks affected and task-level cost savings, complicating stronger productivity-growth forecasts.
https://www.nber.org/papers/w32487

[5] Stanford HAI’s 2025 AI Index reports that corporate AI investment reached $252.3 billion in 2024, with private investment up 44.5 percent; it also reports U.S. private AI investment at $109.1 billion, far exceeding China’s and the U.K.’s.
https://hai.stanford.edu/ai-index/2025-ai-index-report/economy