Anthropic's economics team this month released a task-based economic model and an accompanying scenario browser projecting AI's impact on US growth, employment and wages through 2030. The tool is labeled version 1.0, and the underlying paper is credited to Korinek and colleagues. The model lays out three scenarios rather than a single central forecast.
| Scenario | 2030 GDP | Relative gain | Unemployment | Knowledge-worker wages |
|---|---|---|---|---|
| Modest | $34.1 trillion | +1.6% | Within historical range | Flat or up depending on sector |
| Significant | $36.3 trillion | +8.3% | Rises to about 5% | Roughly flat |
| Extreme | $44.4 trillion | +32.4% | Rises to historic highs | Down more than 10% by 2030 |
The modest scenario assumes AI's impact is comparable in scale to the internet's, with 2.5% of knowledge workers displaced between 2026 and 2030. The significant scenario assumes AI can handle half of knowledge work by 2030, most of it running autonomously — under that path, 59.7% of knowledge workers stay in their current roles, 39.6% move into other work, and wages for non-knowledge workers actually rise. The extreme scenario requires AI capable of recursive self-improvement and rapid adoption, pushing annual GDP growth to 15% and doubling the size of the economy every 4.5 years.
Distribution shifts faster than the total
The biggest swing across the three scenarios is in distribution. Labor's and capital's shares of GDP run 59.4% to 40.6% in the modest scenario, 56.1% to 43.9% in the significant scenario, and 45.2% to 54.8% in the extreme scenario — capital's share climbs 14.8 percentage points above where it stands today, crossing the halfway mark.
A rough calculation from the three GDP figures shows the extreme scenario produces about $10.3 trillion more in annual output than the modest one, close to a third of the current size of the US economy. In the same model, that extra output comes paired with knowledge-worker wages falling more than 10% and unemployment rising to historic highs. Total output and its distribution don't move in the same direction in this projection, which is the main point of laying the three scenarios out side by side.
How the model is built
The framework represents the economy as bundles of tasks carried out by different occupations, with tasks classified using the US Department of Labor's O*NET system. AI's effect on a task is split into four types: no effect, augmentation (doing the task faster or better), automation, and the creation of entirely new tasks. The model also incorporates task-level productivity changes, the cost and time required to switch between occupations, differences in adoption speed across industries, and capital substituting for labor.
The accompanying survey covered 10,980 respondents, whose median expectation aligns with the significant scenario; about 10% gave answers close to the extreme scenario.
"The main challenge is not achieving economic growth, but making sure the benefits are broadly shared and the costs aren't unequally dispersed."
What the model leaves out
This list is worth reading more closely than the scenarios themselves. The model excludes policy responses, business cycles, aggregate-demand or financial-market shocks, catastrophic risks, and highly capable robots, and it doesn't track individual workers — only aggregates at the occupation level. It also leaves out potential growth in AI-exposed occupations themselves, demand effects from data-center construction, and any acceleration in the pace of technological progress.
As the authors put it: "Like every model, it is a stark simplification of a complex reality: it isolates a few key forces and omits many others that may become relevant."
Sources: Anthropic's official economic-research page, CocoLoop; the three scenarios' GDP figures, unemployment and wage ranges, labor/capital share percentages and survey sample size were checked against the model's documentation, and the cross-scenario gap is this site's own rough calculation.