A $44 Trillion Economy Where Labor Loses 15 Points: Inside Anthropic's Interactive AI Scenario Explorer

There is a particular genre of AI forecasting that everyone has learned to discount: the think-tank slide deck with three arrows on it, none of which ever get audited again. Anthropic’s Economics team has just done something categorically different. On September 9, alongside a technical report titled Economic Scenarios for Transformative AI (Korinek et al., 2026), the company shipped an interactive scenario explorer — a working model of the US economy in 2030 that anyone can poke at. You enter your own assumptions about how capable AI will become, how widely it gets adopted, how much of its work it does autonomously, and how productive it makes the humans who remain. The model then computes the economy those beliefs imply: GDP, unemployment, wages, and the split of income between labor and capital.

The headline result is not that AI grows the economy — every scenario does that. It is that in the fastest-growth scenario, the share of national income going to workers collapses from roughly 60% today to 45.2%, while knowledge-worker wages fall by more than 10%. The pie gets dramatically bigger. Most workers’ slice gets smaller.

The model: an economy made of tasks

The intellectual core of the explorer is a task-based representation of work, rather than an occupation-based one. Every job in the US economy is modeled as a bundle of tasks drawn from the Department of Labor’s O*NET taxonomy. Anthropic illustrates this with the day of a nurse: she does rounds, draws blood, triages patients, charts vitals, orders supplies. AI can leave each of those tasks alone, augment them (drafting discharge instructions, monitoring patients remotely), fully automate them (charting vitals, ordering supplies), or create entirely new ones (auditing how well an AI triages patients).

Aggregated across the millions of instances of every task performed every day in the US — which together make up the $30-trillion-plus American economy — these micro-level choices mechanically produce macro-level outcomes. Whether a task gets augmented or automated, how much productivity AI adds, how fast adoption spreads, and how long displaced workers need to find new jobs: these are the explorer’s input dials, and GDP, labor share, and unemployment are its outputs. It is the same general approach pioneered in the academic literature on automation and tasks, but packaged as a public instrument that a policy analyst, a union researcher, or a curious software engineer can actually run.

Three scenarios, one uncomfortable gradient

The explorer highlights three illustrative scenarios, and the gradient between them is the story.

Modest. AI’s macroeconomic impact is comparable to the internet’s: real, but within the historical norm for general-purpose technologies, and it arrives gradually. 2030 GDP lands 1.6% above the no-AI baseline at $34.1 trillion. Annual growth edges up from about 2% to 2.4%. Unemployment sits at 3.9%. Nobody’s world changes very much.

Substantial. AI becomes capable of doing half of all knowledge work by 2030, the majority of it autonomously — but adoption is incomplete, and most knowledge-work tasks still get done without AI. The economy grows at roughly twice its normal rate. GDP comes in 8.3% higher at $36.3 trillion. Labor’s share of income drifts down from about 60% to 56.1%. Knowledge-worker wages go essentially flat, while non-knowledge workers — electricians, nurses, construction crews — see pay rise, because AI-accelerated design and permitting feed demand for physical work that models cannot do.

Extreme. AI outperforms humans at the vast majority of knowledge-work tasks, does nearly all of them autonomously, and creates essentially no new knowledge tasks for people. Anthropic is explicit that this path likely requires recursively self-improving AI adopted quickly across knowledge work. Annual GDP growth reaches 15% — the economy doubles every 4.5 years — and 2030 GDP hits $44.4 trillion, 32.4% above baseline. But knowledge-worker wages fall more than 10%, unemployment rises beyond typical recessionary levels (with cognitive-occupation unemployment hitting double digits), and labor’s share collapses to 45.2%, handing capital an additional 14.8 points of national income.

Read in sequence, the scenarios make a point that is easy to miss in single-number headlines: the growth case and the distributional case are not separate debates. The very mechanism that produces 15% annual growth — autonomous AI doing knowledge work without creating new human tasks — is the same mechanism that strips labor of bargaining power. Fast growth and broad-based prosperity are not automatically the same future.

What ten thousand Americans think

Anthropic paired the model with a Morning Consult survey of 10,980 US adults, fielded in August, asking five questions that map directly onto the explorer’s inputs: what tasks AI can do, how much people use it, how much it does autonomously, how productive it makes people, and how long re-employment takes after an occupation change.

The median respondent’s answers imply outcomes close to the substantial scenario: GDP about 10% higher by 2030 than it would otherwise be, and an overall unemployment rate around 5%. Roughly 10% of respondents hold views consistent with the extreme scenario. Public expectations, in other words, cluster well away from both dismissal and doom — the typical American already prices in a world where AI handles a large fraction of knowledge work, and where the labor market absorbs a real shock without collapsing.

Honest limits, stated by the model’s own authors

What distinguishes this exercise from lab marketing is the candor of the limitations section. The explorer excludes policy responses — no retraining programs, no transfers, no wage insurance appear in the model, even though any real 15%-growth world would surely produce them. It excludes business cycles, aggregate-demand effects, and potential financial-market disruptions. It ignores hyper-capable robotics, so physical work is treated as a refuge even though that assumption may not age well. It does not track individual workers through displacement, and — as outside reviewers noted — it omits the aggregate demand effects of the current data center buildout itself.

The technical report was circulated to economists including Daron Acemoglu, David Autor, and Emi Nakamura before release, and Anthropic says two of their critiques visibly changed the model: rising returns to capital, and diverging wages between exposed and unexposed occupations are now in the results.

Why a frontier lab would publish this

The most interesting question is motivational. Why would a company building frontier models publish an instrument that quantifies how its own product could hollow out knowledge-worker wages? The answer fits Anthropic’s institutional strategy: the model’s outputs will feed the grants it makes through its Economic Futures program and inform its policy recommendations. The implicit argument is that the live question is no longer whether AI grows the economy — on that, the scenarios agree — but who ends up holding the larger pie, and whether policy, training, and distributional institutions can move faster than the task displacement does.

For practitioners, the practical takeaway is narrower than the headlines. Task-level augmentation dominates in the modest and substantial scenarios; autonomous execution dominates only in the extreme one. The gap between the futures turns almost entirely on how quickly agentic systems become reliable enough to run without a human in the loop. That is not an abstraction. It is the roadmap question every AI team is currently shipping against — and now there is a public model that converts each answer into a GDP figure, a wage curve, and a labor share, before 2030 arrives to grade them.