Who will capture the gains from AI growth: Anthropic’s scenario for wages, employment, and capital

Anthropic’s scenario shows that rapid economic growth does not guarantee higher earnings for most workers. In the extreme calculation, by 2030, capital owners’ income increases by 801%, while Americans’ total labor income rises by only 0,5%.

This is not an exact forecast or an official assessment of the future labor market. Anthropic uses a mathematical model to examine how different rates of AI adoption could affect production, employment, wages, and the distribution of national income.

Why economic growth does not equal growth in household income

Economic growth shows how many goods and services have been produced, while household income depends on who receives the value created. If companies increase output through automation, profits and asset values may grow faster than the payroll.

This distinction lies at the heart of Anthropic’s scenario. In the extreme case, US GDP reaches 44,4 trillion dollars in 2025 prices by 2030, which is 32,4% above the baseline forecast. Annual growth reaches 15,4% by the end of the period in the model, yet labor income barely changes.

Capital owners benefit for several reasons: companies reduce the cost of performing intellectual tasks, productivity rises, and profits and business values increase. Workers earn income mainly by selling their time, skills, and responsibility. Their position depends on demand for specific tasks and their ability to move into another field.

High GDP by itself does not answer whether an individual specialist will become wealthier. To determine that, we need to look at wages, employment, workers’ share of income, and access to ownership.

How Anthropic’s model describes task automation

The report’s authors view an occupation as a set of individual tasks rather than an indivisible job. This approach makes it possible to separate the automatable part of work from functions requiring physical presence, communication, oversight, or the acceptance of responsibility.

In medicine, AI can prepare documents, create a schedule, place an order, or fill out an electronic medical record. Helping a patient bathe remains a physical task that the system cannot perform on its own. Automation also creates new responsibilities for medical staff: specialists review AI decisions, monitor the system, and correct errors.

The model accounts for the share of tasks available for automation, the scale of adoption by companies, productivity gains, fully automatable processes, and new responsibilities for people. This list does not turn the calculation into an observation of the real economy: it establishes a set of assumptions on which the final result depends.

The analysis uses O*NET data. According to the report’s authors, intellectual occupations account for 62,4% of employment in the United States. This group includes managers, professionals, office workers, salespeople, and other workers who primarily process information.

Three scenarios: from moderate adoption to near-total automation

The authors compare the moderate scenario with the spread of the internet. In it, AI adds 1,6% to GDP by 2030, while annual growth accelerates from 2% to 2,4%. This effect is noticeable for the economy, but it is not enough to change the incomes of every group of workers.

The high scenario assumes that by 2030, AI performs roughly half of all intellectual work, with the system solving some tasks independently. Gradual adoption softens the blow to employment: companies continue to use people, and the market has time to restructure some processes.

The extreme scenario assumes very rapid technology adoption and possibly recursive AI self-learning. The authors build in near-total automation of intellectual labor, a doubling or greater increase in the productivity of the processes involved, and an almost complete absence of new tasks for people.

What will happen to the wages of intellectual workers

In the extreme scenario, wages for intellectual workers are 11,5% below the level expected without AI. Their real incomes by 2030 may be slightly below the level of the middle of 2026, even if the economy’s total output increases sharply.

The reason is an oversupply of human labor. When AI takes on most analytical, administrative, and programming tasks, employers can perform the same amount of work with fewer specialists. Demand for people does not disappear completely, but the price of their labor falls.

In this model, physical labor occupies the opposite position. Wages for construction workers, electricians, and other manual workers are 33,6% above the baseline forecast because AI increases the number of projects requiring installation, maintenance, and on-site work.

This does not mean that all physical labor will become highly paid. The result reflects a specific assumption: digital planning increases the number of projects faster than their physical execution is automated. If robotics reaches the same scale, the advantage of physical labor will diminish.

Flexible wages and unemployment: two different risks

Anthropic’s scenario distinguishes between employment and earnings problems. With inflexible wages, unemployment among intellectual workers could reach 24% because employers cannot quickly reduce pay to match the new level of demand.

With fully flexible wages, unemployment in this group remains around 2,6%, according to the model’s calculations. The price of this adjustment is high: specialists’ earnings fall by more than 42% relative to the baseline scenario.

What is worse for a worker? The answer depends on which harm is considered the main one. High unemployment deprives people of labor income and makes it harder to move into another profession. A sharp decline in wages preserves employment but reduces the standard of living of those who continue working.

Moving between professions also takes time. Under normal conditions, roughly one in every seventh person who finds a new job switches from an intellectual field to a physical one or vice versa. Among managers and highly qualified specialists, such a transition occurs in about one in ten cases, so large-scale displacement will require training and new jobs.

Why capital receives a larger share of the gains

In the extreme scenario, workers’ share of national income falls from approximately 60% to 45%, while the capital share rises from 40% to 55%. Total labor income increases by only 0,5%, whereas capital owners’ income, according to the model’s calculations, rises by 801%.

The figure 801% indicates growth relative to the initial level of capital income in the scenario, not the share of capital in national income. These are different quantities: the capital share is 55%, while the change in asset owners’ income reflects an increase in the income stream itself.

Capital gains an advantage when software systems replace the labor component of the production process, while rights to infrastructure, computing capacity, data, companies, and intellectual assets remain with a limited group of owners. Productivity rises across the economy, but how the gains are distributed depends on ownership and market rules.

GDP growth does not guarantee evenly distributed prosperity. Without mechanisms for workers to participate in profits, additional output may accrue to owners of companies and financial assets.

What scale of redistribution does Anthropic consider

The authors estimate the necessary scale of compensation at approximately 9% of US GDP. In the model, this amount would make it possible to preserve the incomes of people harmed by automation, while other participants in the economy would become more than 20% wealthier.

This is a calculated amount, not a ready-made budget program. Redistribution could take the form of transfers, tax changes, income insurance, retraining support, or worker participation in company income. The choice of instrument affects incentives and consequences, so one figure cannot replace a policy decision.

The scale of 9% of GDP shows the difficulty of the task. There are almost no historical examples of a government compensating for the consequences of a technological shift on a comparable scale. In the report, the authors compare this scenario with the consequences of a sharp rise in imports from China: some American workers lost their jobs without full compensation.

Even having the funds does not solve every problem. People need a new profession, access to training, health insurance, social connections, and time to make the transition. If new jobs appear in other regions or require different qualifications, a cash payment merely softens the blow.

Model limitations and reasons to read it cautiously

The extreme scenario is not a forecast. It describes a limiting case in which AI spreads rapidly, performs intellectual tasks almost entirely, substantially raises productivity, and creates almost no new functions for humans.

Anthropic’s model largely does not account for an acceleration in scientific discoveries. According to the estimate provided, the cumulative volume of new knowledge by 2030 in the extreme scenario is only 0,6% above the level without AI. The authors explain this by noting that new discoveries become more difficult, while scientific progress depends on physical experiments and other processes that cannot be automated instantly.

Recursive self-improvement is also not modeled as an independent dynamic. Its possible effect is incorporated into the productivity figures, so the model does not reproduce a technological singularity or describe the consequences of the emergence of a huge number of digital researchers.

Another limitation concerns infrastructure. If companies do not manage to build facilities, data centers, and other infrastructure, average wages could be 9,2% below the baseline forecast even amid rapid economic growth. The economy needs algorithms, energy, equipment, construction, communications, regulation, and implementation specialists.

The US labor market has so far seen relatively small changes associated with the spread of neural networks. However, current trends do not rule out subsequent automation. Anthropic’s scenarios are useful not as a promise of a specific future, but as a way to see the main risk: overall productivity growth could be distributed between workers and capital owners extremely unevenly.

What the scenario means for workers and managers

• Workers should assess not only their profession but also the set of tasks within it. Automation usually begins with individual operations rather than the instantaneous disappearance of a job.
• Managers need to calculate AI’s impact together with the costs of verification, security, training, infrastructure, and responsibility for errors.
• A person who can check AI’s output, connect digital tools with a physical process, and take responsibility for the result retains more human value.
• Government policy must take unemployment, falling wages, and concentrated ownership into account: these risks require different measures.

The main conclusion of the Anthropic scenario is simple: AI can increase the size of the economy, but income distribution is not determined by productivity alone. It is influenced by ownership structure, wage flexibility, the speed of transitions between occupations, and the scale of compensation for those whose tasks are automated.

Frequently asked questions

Is the increase in capital income of 801% an exact forecast?

No. It is the result of an extreme scenario from Anthropic’s mathematical model. It depends on assumptions about the speed of AI adoption, productivity, task automation, and the emergence of new roles for people.

Will all knowledge workers lose their jobs?

No. The calculation applies to a group of occupations and sets out a possible average trajectory. Within a single occupation, some tasks may be automated, while demand for oversight, communication, responsibility, and work involving the physical world may persist.

Why is unemployment lower when wages are flexible?

Lower pay allows employers to preserve more jobs. But the model shows the other side of the issue: the earnings of specialists could fall by more than 42% relative to the baseline scenario.

What does redistribution amounting to 9% of GDP mean?

This is an estimate of the scale of transfers that, according to the authors’ calculations, could offset the consequences of automation. It does not represent an approved program or specify a particular tax or payment.