Artificial intelligence can make the economy produce much more money and, at the same time, a significant part of those who today work in front of a computer will end up losing.
That is probably the most uncomfortable conclusion of the new economic study by Anthropic. It does not suggest that AI will necessarily impoverish us. It suggests something quite different: we could live in an extraordinarily richer economy, but with less employment in certain professions, much more unequal wages, and a growing proportion of that wealth ending up in the hands of capital owners.
The work is new. Economic Scenarios for Transformative AI appears as Working Paper 2026-02 from the Anthropic Institute and is dated September 2026. Anthropic also released this week an interactive tool based on that model to explore what the U.S. economy could look like in 2030.
Its authors are economists Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter, and Peter McCrory.
And there is a necessary warning before delving into the numbers: they are not predicting what will happen. They have constructed three possible futures and do not assign a probability to any. The study serves to observe what would happen if the capacity and adoption of AI advance at different speeds.
Three completely different economies in just four years
The model starts in 2026 and ends in 2030.
The modest scenario assumes that AI advances and spreads, but without causing a historical break. The economic effect would be similar to that of other recent major technologies.
The substantial scenario implies a much greater transformation of intellectual work. By 2030, AI would be capable of performing approximately half of the cognitive work considered by the model, although companies would not use it for everything it could technically do.
The extreme scenario enters another territory. AI would be better than humans at most cognitive tasks, would take on almost all of them autonomously, and very few new intellectual tasks reserved for people would appear. To get there, it would probably be necessary for the AI systems themselves to accelerate the development of new generations of artificial intelligence.
And the three economies that emerge from there have very little in common with each other.
GDP always grows: from an additional 1.6% to 32.4%
The first thing that Anthropic finds is that AI grows the economy in all its scenarios.
In the most moderate scenario, the U.S. GDP in 2030 would be 1.6% higher than in an equivalent economy without artificial intelligence. It would reach about 34.1 trillion dollars, measured at 2025 prices. In the substantial scenario, the difference increases to 8.3% and the GDP reaches 36.3 trillion.
And at the extreme, the leap occurs: 44.4 trillion dollars, 32.4% more than without AI. The annual growth would then reach 15.4%. At that rate, the economy could approximately double its size every four and a half years.
The problem begins when it is asked who gets that growth.
A 32% richer economy can have almost 12% unemployment
More productivity does not automatically mean more employment. In the modest scenario, the labor impact is barely noticeable. The overall unemployment rises from the 3.8% used as a reference by the model to 3.9%. In the substantial scenario, it reaches 4.6%.
At the extreme, it rises to 11.9%. Among workers starting in cognitive occupations, it reaches 17.9%, practically one in five.
It is not that all that work disappears. Part is automated. Part changes. Some of the displaced people seek employment in other professions. And precisely there appears one of the central problems of the study.
A programmer may lose part of the demand for their work because AI writes code. That does not mean they can immediately become an electrician, nurse, or construction worker.
Changing professions takes time, requires different skills, and does not always work. The greater the number of workers trying to do so simultaneously, the greater unemployment may be during the transition.
Programmers, analysts, administrative staff, and professionals are more exposed
Anthropic does not simply divide the economy into jobs that will disappear and jobs that will survive.
It divides each occupation into tasks. A nurse may have administrative tasks that an AI automates and other physical or human tasks that continue to depend on a person. A lawyer may delegate the initial review of documents while retaining other functions. A programmer may produce more code with the help of an AI or see part of their work directly automated.
The model concentrates direct exposure on management, professional, commercial, and administrative occupations. Together they represent around 62% of the U.S. employment used to calibrate the study.
Physical professions are much less exposed in this model because Anthropic deliberately does not incorporate an accelerated revolution of robotics before 2030.
This generates one of its most striking conclusions: in certain scenarios, being an electrician, nurse, or working in construction can have economically more potential than developing a purely intellectual activity in front of a computer.
Wages are split in two
Here appears probably the result with the greatest everyday impact.
In the modest scenario, practically everyone improves. The average salary is 0.7% above what it would have been without AI. Cognitive workers earn an additional 0.4% and the rest 1.1%.
The substantial scenario already begins to separate the paths.
The average salary increases by 2.1%, but that of cognitive occupations remains practically frozen, 0.3% below the scenario without AI. In the rest of the professions, it increases by 5.9%.
The extreme scenario breaks the labor market in two.
The average salary is 9.7% higher. But that hides a huge difference.
The salaries of cognitive workers are 11.5% below what they would have reached without artificial intelligence. Those of other occupations are 33.6% higher.
In other words: an economy can pay on average higher salaries and, at the same time, pay less to a huge part of the people who today hold some of the positions considered more qualified.
How an electrician can earn more if AI automates programmers
The explanation of the model has quite a bit of economic logic. If AI allows designing a building faster, processing documentation in less time, calculating structures at a lower cost, or automating a good part of the planning, more projects can be initiated.
Then the demand for workers whose jobs benefit indirectly from the productivity of AI but that AI itself cannot physically perform.
Anthropic provides examples such as construction, electricity, or nursing. At the same time, the opposite happens with certain intellectual jobs.
If a company can produce the same amount of software, documents, analysis, or customer service using fewer workers, the demand for people to perform those tasks falls.
AI would not have to completely replace a profession to exert pressure on its salary. It would be enough to reduce the number of workers needed.
The big change could come after 2027
The study's timeline is also relevant. The authors calculate that almost all the difference between their three scenarios appears after 2027. Until then, economic trajectories remain relatively close.
This helps explain an apparent contradiction in the current debate. The use of artificial intelligence has increased rapidly, but there is still no evidence in the aggregated U.S. data of massive job destruction attributable to it.
Another study by Anthropic published in March also did not find a systematic increase in unemployment among workers in the most exposed professions since late 2022. It did detect signs of a slowdown in the hiring of young workers in those sectors, but not a general break in the labor market.
The new work presents two possible explanations. One is reassuring: perhaps the economy is truly moving toward the moderate scenario. The other is much less so: perhaps we are still in the early stages of a transformation that will accelerate when AI can perform many more tasks and companies begin to use it on a large scale.
The pie grows, but it changes who gets it
The deeper issue in Anthropic's work is not unemployment. It is the distribution of wealth.
Currently, the model starts from an economy where 60% of income corresponds to labor and 40% to capital. In the modest scenario, it barely changes: workers retain 59.4% and capital rises to 40.6%. In the substantial scenario, labor's share falls to 56.1% and capital increases to 43.9%.
At the extreme, a reversal occurs: labor receives 45.2% and capital 54.8%.
For the first time in this simulation, more than half of the income would end up rewarding capital instead of labor.
Capital owners can become the big winners
The data becomes even more compelling when looking at how much money each group receives. In the extreme scenario, GDP is 32.4% higher than in the economy without AI.
However, total income derived from work is only 0.5% higher. Capital income increases by 81.4%.
That is the heart of the most disruptive scenario. Society produces much more. The problem is not in creating wealth.
It is in how that wealth reaches the population if an increasingly smaller proportion is distributed through wages.
For cognitive workers, the difference is even more pronounced. The total wage mass of that group falls by 31% compared to the trajectory it would have followed without AI. Two blows combine: wages are 11.5% lower and there is 21.5% less employment in those occupations.
The paradox: the average wage can rise while workers lose weight
There is a result that seems contradictory until the variables are separated. In the extreme scenario, the average wage increases by 9.7%.
At the same time, the participation of workers in the economy falls from approximately 60% to 45.2%.
Both things can happen at the same time. The economy has become so large that certain workers earn much more in absolute terms. But the owners of companies, infrastructure, data centers, chips, and other assets necessary to produce with AI capture an even larger part of the growth.
The size of the pie increases much faster than the portion allocated to work.
Which of the three scenarios do people think will occur?
Anthropic added to the study a representative survey conducted by Morning Consult of 10,980 American adults during August.
They asked them how far they think AI will go, how much companies will use it, what productivity it will bring, how much work it will automate, and how long it would take a displaced person to find another job.
When the researchers introduced the median response of the population into their model, a very similar economy to the substantial scenario appeared.
The GDP would be 8.6% above the trajectory without AI in 2030. Unemployment would be around 4.6%. Cognitive employment would be approximately 4.2% lower.
The wage of cognitive workers would hardly change, while that of other occupations would be 6.4% higher.
Only about 10% of respondents gave answers compatible with the extreme scenario.