Anthropic: 15 percent growth, 18 percent knowledge-work unemployment

Anthropic has published an interactive model of how AI could reshape the US economy through 2030. It is not a forecast. It is a map of three futures, built by the company’s economics team and released as working paper 2026-02.

For boards, CIOs and CHROs the operational point is blunt: growth and pay do not have to move together. In the most transformative path, society is much richer while knowledge workers take lower wages, higher unemployment — and every extra dollar of GDP accrues to capital.

What Anthropic actually released

On 9 September Anthropic opened “Scenarios for our Economic Future” at anthropic.com/institute/econ-scenarios. Behind the explorer sits Economic Scenarios for Transformative AI, by Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter and Peter McCrory at The Anthropic Institute.

The model treats jobs as bundles of tasks. AI can augment a task, automate it, leave it alone, or create new tasks. Paths for GDP, wages, unemployment and labor’s share of income then follow from how much of the economy is affected, how fast firms adopt, how large the productivity gain is, and how much of the affected work is automated rather than assisted.

The authors say the views are theirs, not necessarily Anthropic’s. The explorer is version 1.0. It omits policy responses, business cycles, demand from the data-center buildout, humanoid robots and catastrophic risks.

Three paths to 2030

Table 3 in the working paper is the board pack.

In the modest scenario, AI looks like the internet. Four percent of tasks in the economy are affected. GDP sits 1.6 percent above a no-AI path. Annual growth is 2.4 percent against 2.0 without AI. Labor’s share barely moves, from 60.0 to 59.4 percent. Unemployment rises from 3.8 to 3.9 percent. Cognitive employment is 0.5 percent below mid-2026. Visible, but inside the historical range for a new general-purpose technology.

In the substantial scenario, AI is larger than the internet or the railroad. Twelve percent of tasks are affected, roughly 20 percent of knowledge work. On those tasks AI lifts productivity by about 57 percent, and three quarters of affected instances are automated. GDP is 8.3 percent above the no-AI path. Growth in 2030 is 5.4 percent a year. Average wages rise 2.1 percent, but cognitive wages sit 0.3 percent below the no-AI path. Labor’s share falls to 56.1 percent; capital’s share rises to about 44. Cognitive unemployment moves from 2.9 to 4.5 percent. The authors note that a four-point shift in factor income in four years almost matches the entire multi-decade decline in the US labor share.

In the extreme scenario, AI performs almost half of today’s cognitive work. Thirty percent of tasks in the economy are affected, diffusion is 60 percent, the productivity gain more than doubles affected tasks, and 90 percent of instances are automated. GDP is 32.4 percent above the no-AI path. Annual growth hits 15.4 percent — the economy doubles about every 4.5 years. Cognitive employment is 21.5 percent below mid-2026. Cognitive unemployment hits 17.9 percent; economy-wide unemployment 11.9. The cognitive wage is 11.5 percent below the no-AI path; wages in other occupations are 34 percent above it. Labor’s share falls from 60 to 45.2 percent.

That last number is the governance issue. In the extreme path, total labor income in 2030 is almost exactly what it would have been without AI: 0.45 times 1.32 is about 0.60. The entire GDP increase becomes capital income, 81 percent above the no-AI path. The authors write that a transfer of about 15 percent of GDP — roughly Social Security and Medicare combined — would be needed to hold knowledge workers’ income.

The public’s expectations sit in the middle

In August Anthropic surveyed more than 10,000 US adults on what AI will be able to do, how widely it will be used, and how long occupational switching takes. The median respondent’s answers imply GDP 8.6 percent above the no-AI path and overall unemployment around 4.6 percent — close to the substantial scenario. About 10 percent of respondents map onto the extreme path.

Anthropic assigns no probabilities. The explorer is a risk map, not a forecast. Dario Amodei has previously warned that up to half of entry-level office jobs could vanish and that unemployment could hit 10–20 percent. Those numbers match the extreme scenario, not the middle case. Boards should hear that gap: lab leaders can talk about the tail while the published model treats the tail as one of three tracks.

Why this matters outside the United States

The calibration is US data. The mechanism is not. Knowledge-heavy economies — software, consulting, finance, energy services, public administration — are exactly the cognitive occupations in the model. A Nordic welfare state, tripartite bargaining and a sovereign wealth fund change how shocks are absorbed. They do not repeal the arithmetic: if AI automates knowledge tasks faster than new tasks appear, pay and employment in those occupations come under pressure, while demand can rise in physical work that builds the infrastructure around the productivity gain.

The paper’s example is faster design and permitting feeding more construction. For European CIOs that is power, grids, housing and public digitalisation. It is also a vendor question. If surplus accrues to model and cloud capital rather than payroll, AI ROI has to be measured as distribution, not only as hours saved.

Decisions for the CIO, CHRO and board

Run your own scenarios, not the lab’s slides. Ask for three 2030 paths on your own knowledge tasks: share affected, share automated versus augmented, and time to re-employment. If the middle case is the base, the extreme case is the stress test.

Separate productivity from headcount. A 57 percent gain on affected tasks is not a 57 percent cut in FTEs. In the substantial path GDP rises more than wages. In the extreme path GDP rises while knowledge workers’ total income stands still.

Put surplus ownership in the contract. If agents take over casework, code, analysis and customer operations, price, data rights and exit terms should reflect that labor’s share of income can fall quickly.

Plan occupational switching, not just courses. The model says moving from coder or contact-center work into electrician or nurse is slow and costly. That is licensing, apprenticeships and public capacity, not an e-learning module.

Do not read the model as a vision. It leaves out policy. It leaves out demand from the data-center boom. External reviewers including Daron Acemoglu and David Autor commented on a draft without endorsing the conclusions. Some called the extreme path a thought experiment. Some said the modest path already understates what is visible in the data.

What is published is an open framework with numbers. For a European board the question is not whether the US hits 15 percent growth. The question is which of the three mechanisms — a small macro effect, doubled growth with flat knowledge-work pay, or growth that accrues entirely to capital — will govern their own workforce over the next three to four years.

Sources and media

Anthropic, “Scenarios for our Economic Future”, 9 September 2026, https://www.anthropic.com/institute/econ-scenarios

Korinek, Jones, Sacher, Cotter and McCrory, “Economic Scenarios for Transformative AI”, The Anthropic Institute Working Paper No. 2026-02, September 2026, https://www-cdn.anthropic.com/files/4zrzovbb/website/cf58f84d46a4a76bf5a5b039ac695fba6b80041c.pdf

Anthropic on X, scenario explorer announcement, 9 September 2026, https://x.com/AnthropicAI/status/2097679799829307588

The Decoder, “Anthropic built an economic model that frames its CEO's bleakest job forecasts as an outlier scenario”, 9 September 2026, https://the-decoder.com/anthropic-built-an-economic-model-that-frames-its-ceos-bleakest-job-forecasts-as-an-outlier-scenario/

Thumbnail: OpenAI Image 2 / hogby.ai