Anthropic's economists built a model of how AI could reshape US jobs, growth and wages by 2030. We stripped it to the essentials, made the projections legible, and checked them against every other major forecast, from the sceptics to the explosive-growth camp.
By Christos Papadimitriou · House of 47 Research
Last updated:
The economy is a pile of tasks. AI can ignore a task, speed it up, automate it, or invent new ones. How far that goes, and how fast, decides which future arrives. In the calm cases AI looks like the internet. In the extreme one it is unlike anything in economic history, and hardest on knowledge workers.
AI shows up in tasks, not yet in the macro data. The whole economy is a bundle of tasks AI can leave, augment, automate, or expand.
$34.1T. Real but gradual, within historical norms. Jobs stay in the usual range. Pay flat to up. Capital takes 40.6% of the pie.
$36.3T. AI can do half of knowledge work, mostly on its own. Growth ~2× normal. Unemployment ~5%, knowledge pay flat, other pay rises. Capital 43.9%.
$44.4T. 15%/yr growth, the economy doubles every 4.5 years. Needs recursively self-improving AI, adopted fast. Knowledge pay −10%+, unemployment spikes. Capital 54.8%.
From +1.6% to +32% of 2030 GDP. The disagreement is about magnitude, never direction. AI grows the pie.
Coding, clerical, support and analysis are hit first. Manual trades and care work are least exposed.
Faster designs and permits mean more construction; manual pay climbs while knowledge pay stalls or falls.
Today ~60¢ of each dollar goes to workers. As AI automates tasks, more flows to the owners of the machines, a wealthier, more concentrated class that is 47HNWI's beat.
ILO, OECD and the World Bank stress augmentation over destruction, few jobs are fully automatable soon.
In the big scenarios the pie grows but the worker's slice thins. The challenge shifts from growth to sharing.
Headline uplift from each forecast. Mixed geographies (US vs global), placed by magnitude.
Share of jobs touched by AI. "Exposed" ≠ automated, definitions differ by source.
Is AI another tool, or a new super-exponential growth regime?
Acemoglu 2024 (~+1% GDP/decade) · Goldman 2023 (+7% global GDP, 300M jobs exposed) · McKinsey 2023 ($2.6–4.4T/yr) · PwC 2017 (+$15.7T/+14%) · IMF 2024 (40% exposed, 60% advanced) · WEF 2025 (net +78M jobs) · OECD 2023 (27% high-risk) · ILO 2025 (3.3% top bucket) · Epoch GATE 2024 (20–30%/yr) · Davidson 2021 (≥10% chance of >30%/yr).
Synthesis estimates across the forecasts above, scenario estimates, not market forecasts. Anthropic's own survey of 10,980 Americans lands, on average, close to the substantial scenario (GDP +10%, unemployment ~5%); about 1 in 10 hold extreme-scenario views.
Average pay rises in every scenario, but the gains collect outside knowledge work. The more transformative the future, the harder knowledge workers are hit and the more people must cross into trades AI barely touches.
| Group | Modest | Substantial | Extreme |
|---|
Anthropic reports results by group (knowledge vs. other), not by individual occupation; example roles are shown for orientation. "Spikes" / "much higher" are directional. Crossing occupations is slow, which is why transformation raises unemployment.
In every scenario, capital's share of the pie rises. The durable move is to own the systems that produce value, not only to sell the hours. In practice that means building the agents that do the work, the business of theagency47.
If agents become the first point of contact, the channel and the audience are the moat. Owned media and audiences compound while ad costs don't.
The structured knowledge that feeds AI, facts, pricing, policies, proof, is the asset agents run on. Build it once, reuse it everywhere.
As the tasks themselves get automated, the money moves to whoever owns the result, the system that produces it, and the client relationship. Sell outcomes, not hours.
Plan for the substantial scenario. The median of every serious forecast, Anthropic's respondents, Goldman, McKinsey, the IMF, the WEF, is a real, internet-plus shock where the economy grows faster and knowledge work is automated in large chunks.
Stress-test for the extreme. A ~1-in-10 tail where growth explodes and knowledge work is hollowed out. Cheap to prepare for; catastrophic to ignore.
The question was never whether AI grows the economy. It's who owns the growth.
Anthropic, Scenarios for our Economic Future · Technical report, Korinek et al. 2026 (PDF)
Goldman Sachs 2023, Gen AI could raise global GDP 7%
McKinsey 2023, Economic potential of generative AI
PwC 2017, Sizing the Prize
Accenture 2017, Why AI is the Future of Growth
IMF 2024, Gen-AI and the Future of Work
WEF 2025, Future of Jobs Report 2025
OECD 2023, Employment Outlook 2023
ILO 2025, Generative AI and Jobs
World Bank 2025, AI exposure in developing countries
Acemoglu 2024, The Simple Macroeconomics of AI
Trammell & Korinek 2023, Growth under Transformative AI
Davidson / Open Phil 2021, Could Advanced AI Drive Explosive Growth?
Epoch AI 2024, AI and Explosive Growth · GATE
Replaced the “Graduate the ladder” takeaway with “Own the outcome”.
First published.