Anthropic has released a research paper, Economic Scenarios for Transformative AI, that tries to turn the debate over AI-driven productivity and job losses into a quantified economic model. The framework covers 2026 through 2030 and examines how AI capability, adoption speed, automation, and productivity gains could affect GDP, wages, unemployment, and labor’s share of income.
The paper finds that AI makes the economy larger in every scenario it models. The open question is where the added wealth ends up: with workers or with capital owners.
Three paths: modest, substantial, and extreme
Anthropic breaks the outlook into three cases.
The first is Modest Change. In that setup, AI looks more like past major technological shifts such as the internet: productivity keeps rising, but macroeconomic indicators do not move dramatically. By 2030, U.S. GDP is about 1.6% above a no-AI baseline, annual GDP growth rises from roughly 2% to 2.4%, and unemployment edges up from a normal 3.8% to about 3.9%.
The second is Substantial Change. Anthropic assumes that by 2030, AI can handle about half of the capability range involved in knowledge work, and can complete most tasks autonomously, though business adoption is still not universal. In this case, GDP is 8.3% above the no-AI scenario, and annual economic growth in 2030 reaches 5.4%, more than double the normal pace.
The third is Extreme Change. Here, AI is more effective than humans across most knowledge work, can operate with little human input, and creates few new knowledge-work tasks for people. Anthropic says this outcome may require recursively self-improving AI plus rapid adoption by companies.
Under that extreme case, U.S. GDP in 2030 is 32.4% above the no-AI baseline. Annual GDP growth climbs to about 15%, implying the economy doubles in size roughly every 4.5 years.
Knowledge workers take the first hit
Anthropic broadly divides jobs in its model into cognitive or knowledge work and other occupations. AI has the clearest direct effect on the first group, including software engineering, administration, sales, and professional services. Jobs that require substantial physical work — such as construction, electrical work, and nursing — see less direct impact in the near term.
In the extreme scenario, knowledge-work employment in 2030 is down 21.5% from mid-2026. The unemployment rate for workers who previously held knowledge jobs rises to 17.9%, close to one in five. Economy-wide unemployment reaches 11.9%.
The paper notes that engineers, customer service staff, or administrative workers displaced by AI may not be able to switch immediately into electrical work, nursing, or other occupations where demand rises. Skill conversion, retraining, and friction in the job search process all extend unemployment spells.
Capital captures more of the upside
Another central result in the model is that AI could shift income distribution more clearly toward capital. Anthropic assumes that for each $1 of output in the U.S. economy today, about 60% goes to labor income and 40% to capital. As automation expands, the importance of compute, data centers, software, equipment, and other capital rises, lifting returns to capital as well.
In the modest scenario, labor’s share slips only slightly, from 60% to 59.4%. In the substantial scenario, it falls to 56.1%. In the extreme scenario, labor’s share drops to 45.2%, while capital income rises to 54.8%.
Put differently, capital income overtakes labor income in the model. Anthropic also says that even if the economy is about one-third larger than it would be without AI, total labor income in 2030 barely increases, with almost all of the added GDP converted into capital income.
Average wages rise, but not for many knowledge workers
The study also challenges the idea that stronger productivity automatically lifts wages across the board. In the substantial scenario, average wages are about 2.1% higher than in the no-AI case, but wages for knowledge workers are about 0.3% lower. Wages for non-knowledge occupations rise as demand increases.
In the extreme scenario, average wages are still about 9.7% above the no-AI baseline. But wages for knowledge workers are 11.5% lower, while pay in other occupations rises by about 30%.
Anthropic’s explanation is that AI sharply reduces business demand for cognitive labor, while the productivity surge it creates also drives more activity in the physical economy, including infrastructure, manufacturing, and services. Jobs that AI struggles most to perform directly may become scarcer and more valuable.
Public expectations line up more closely with the substantial case
Anthropic also worked with Morning Consult on a survey of 10,980 U.S. adults. Respondents were asked about expected AI capability, adoption, autonomy, productivity, and how long it would take to find work again after displacement.
The median public expectation was closest to Anthropic’s substantial-change scenario. Based on those typical responses, U.S. GDP in 2030 would be about 8% to 10% above a no-AI baseline, knowledge-work employment would decline about 4%, and the overall unemployment rate would be around 5%.
Anthropic added that about 10% of respondents gave expectations that were already close to its extreme scenario.