What are Anthropic's three economic scenarios for AI by 2030?

Anthropic's Economics team published 'Economic Scenarios for Transformative AI' (Korinek et al., Anthropic Institute Working Paper 2026-02) in September 2026 alongside an interactive Econ Scenario Explorer. The Modest scenario treats AI like the internet: US GDP 1.6% above a no-AI baseline in 2030 ($34.1T), unemployment in its normal range. The Substantial scenario assumes AI autonomously handles roughly half of knowledge work by 2030: GDP +8.3% ($36.3T), total unemployment around 4.6%. The Extreme scenario assumes AI outperforms humans on almost all knowledge-work tasks with adoption driven by recursive self-improvement: GDP +32.4% ($44.4T), growth near 15% a year, but knowledge-worker unemployment of 17.9% and their wages down more than 10%.

What did Anthropic's survey of Americans say about AI and the economy?

In August 2026 Anthropic surveyed 10,980 Americans on five inputs — what tasks AI can do, how much people use it, how much it does by itself, how much more productive it makes people, and how long it takes to find a new job. The typical respondent's answers imply an outcome close to the Substantial scenario: GDP about 10% higher by 2030 than without AI and overall unemployment around 5%. Roughly one in ten respondents held views consistent with the Extreme scenario. Site visitors (n=22,316) can plug in their own predictions and see the economy they imply.

Does Anthropic's model say AI will cause mass unemployment?

Only in the Extreme scenario, and only for knowledge workers. In the Modest and Substantial scenarios, job reallocation and unemployment stay within historical ranges; average wages rise in every scenario, but the gains concentrate in occupations outside knowledge work such as construction, while knowledge-worker pay is flat (Substantial) or falls over 10% (Extreme). Across Substantial and Extreme, labor's share of GDP falls — from about 60 cents per dollar today to 56.1 cents in Substantial, with capital's share rising to 55% in Extreme — so the model's central warning is about distribution, not aggregate jobs.

What are the limits of Anthropic's economic model?

Anthropic calls it Version 1.0 and a 'stark simplification.' It omits policy responses, business cycles, aggregate-demand or financial-market disruptions, catastrophic risks and humanoid robotics; it does not follow individual workers, so it gives only a coarse picture of displacement costs. External reviewers — including Daron Acemoglu, David Autor, Pete Klenow, Pascual Restrepo and Emi Nakamura — commented on drafts but were not asked to endorse the conclusions, and some argued the Extreme scenario is better read as a thought experiment than a forecast.

Quick Answer

Anthropic's AI Economic Scenarios for 2030 Explained

Published: September 11, 2026

The Short Answer

Anthropic’s Economics team modelled the US economy in 2030 under three AI paths and published the results in September 2026 with an interactive explorer. The economy grows in every scenario; the question the model answers is who gets the growth. In the fastest scenario society is a third richer, knowledge workers are much worse off, and labor income is barely changed.

Scenario	Assumption	US GDP 2030 vs no-AI baseline	Unemployment	Knowledge-worker wages	Labor share
Modest	AI’s footprint ≈ the internet	+1.6% ($34.1T)	Normal historical range	Roughly flat	Little change
Substantial	AI autonomously does ~half of knowledge work by 2030	+8.3% ($36.3T)	Total ~4.6%	−0.3%; other occupations +5.9%; average +2.1%	60¢ → 56.1¢ per dollar
Extreme	AI beats humans on almost all knowledge tasks; recursive self-improvement drives adoption	+32.4% ($44.4T); ~15%/yr growth, economy doubling every ~4.5 years	Knowledge workers 17.9% — above recession levels	Down more than 10%	Capital share rises to ~55%; total labor income barely changed

GDP figures are at 2025 price levels, against a baseline of roughly $31T today.

How the model works

The paper — Economic Scenarios for Transformative AI, Anthropic Institute Working Paper No. 2026-02, co-authored by Anton Korinek and Chad Jones among others — breaks every occupation into tasks and asks four questions about each: does AI augment the task, automate it, leave it alone, or create new tasks? Layered on that are productivity gains, adoption speed, and how long displaced workers take to move occupations. Change the inputs and you get a different 2030; the three published scenarios are three input sets.

The explorer exposes exactly five dials — capabilities, adoption, autonomy, productivity, adjustment time — so anyone can enter their own expectations and see the implied economy, then compare against the survey.

The four findings

1. GDP grows in every scenario, from modestly to explosively. +1.6%, +8.3%, +32.4%. Anthropic notes that the Extreme case implies growth “faster than anything in economic history.”

2. Jobs shift from knowledge work to everything else. Occupations AI affects shrink; those it does not affect grow. In Modest and Substantial, reallocation and unemployment stay within ranges history has seen. In Extreme, knowledge-worker unemployment reaches 17.9% because automation outruns workers’ ability to retrain, relocate and get hired.

3. Average wages rise, but outside knowledge work. Anthropic’s example: if AI produces designs and permits for infrastructure faster, more construction projects start and construction wages rise. Knowledge-worker pay is flat in Substantial and falls more than 10% in Extreme.

4. The pie grows, but capital takes a bigger slice. Today about 60¢ of every dollar goes to workers. As AI makes capital useful for more tasks, its price rises; labor share falls to 56.1¢ in Substantial and further in Extreme, where the gains are “unevenly distributed” and total labor income by 2030 is almost unchanged despite a 32% larger economy.

What Americans expect

The August 2026 survey (n = 10,980) asked the public the same five questions. The median answers imply the Substantial world — GDP roughly 10% higher by 2030, unemployment around 5% — while about 10% of respondents implied the Extreme scenario. Anthropic contrasts this with its own CEO’s more dramatic public forecasts, which the model effectively frames as the outlier case.

The caveats Anthropic lists itself

Version 1.0; “a stark simplification of a complex reality.”
Excludes policy responses, business cycles, aggregate-demand and financial-market shocks, catastrophic risks, and “hyper-capable robots.”
Does not track individual workers, so displacement costs are coarse.
Reviewers (Acemoglu, Althoff, Autor, Cunningham, Freund, Hazell, B. Jones, Klenow, Lashkari, Mitman, Moll, Nakamura, Restrepo, Romer, Steinsson, Tonetti, Visschers, Wiczer) prompted two changes already in the model — higher returns to capital, and divergence between AI-exposed and unexposed occupations — but were not asked to endorse it. Some questioned whether AI-exposed occupations will shrink at all; some read Extreme as a thought experiment.

Anthropic says the model will inform the research it funds on labor-market interventions and the policy proposals it publishes.

Why it matters for anyone building with AI

The practical read for operators: the model’s own median case is not mass unemployment — it is a 4–5% jobless rate with a shrinking labor share and flat knowledge-worker pay. That is a story about margins moving from wages to software and compute. If you sell AI tooling, the Substantial scenario is your addressable market; if you employ knowledge workers, it is your retention problem.