Anthropic’s economics team built a model of how AI could reshape the US economy through 2030, plus a public explorer where you plug in your own assumptions about AI capability, adoption, and how fast people find new work. The framing is simple: the economy is a pile of tasks, and AI can augment a task, automate it, leave it alone, or create new ones for people to do.

They run three scenarios:

  • Modest — AI lands roughly like the internet did. GDP ends 2030 about 1.6% higher ($34.1T).
  • Substantial — AI can do half of all knowledge work, and does most of that on its own, but adoption lags behind. Growth runs about twice the normal rate; GDP +8.3% ($36.3T). Knowledge-worker wages stay flat.
  • Extreme — AI outperforms people at nearly all knowledge work, autonomously, and creates almost no new knowledge tasks for humans. Growth reaches ~15% a year — the economy doubling every 4.5 years — and GDP ends 32.4% higher ($44.4T).

The interesting part is distribution, not growth. Today about 60 cents of every dollar the economy produces goes to workers and 40 to capital. In the extreme scenario labor’s share falls to 45.2% while capital takes 54.8% — and knowledge workers’ pay drops more than 10% even as the country gets much richer. Total labor income is roughly unchanged. (Labor share is just the slice of national income that shows up as wages and salaries rather than profits, rents, and returns to owners.)

Unemployment stays within historical ranges in the modest and substantial cases. In the extreme one it climbs past typical recession levels, because displaced knowledge workers have to move into less-exposed occupations — the team’s example is a coder retraining as an electrician or a nurse. Switching occupations is slow, which is what keeps wages under pressure in the meantime.

They also surveyed more than 10,000 Americans: the typical respondent’s expectations land near the “substantial” scenario, with about 10% describing something closer to the extreme one. The extreme case would likely require recursively self-improving AI — systems improving themselves, in a loop — adopted quickly.

The team is unusually clear about limits: no policy response, no business cycles, no financial-market shocks, no hyper-capable robots, and a model that doesn’t follow individual workers, so it can’t really price the human cost of displacement. One reviewer read the extreme scenario as a thought experiment rather than a forecast.

The conclusion isn’t “AI wrecks the economy.” It’s that in the fast scenarios growth stops being the problem — who gets the growth becomes one.