Anthropic modeled AI's 2030 economy: enter your guess, see the outcome

Anthropic's economics team has published a model of how AI might affect growth, wages and employment through 2030. It is not a report but a tool you run: you set five dials with your own assumptions, the outcome appears on screen, and it sits next to the answers of more than 10,000 Americans.

Three scenarios

Anthropic has worked out three end points itself. Without AI, US gross domestic product in 2030 would be $33.5 trillion; the scenarios stack on top of that.

• Modest: AI has roughly the impact the internet did. Output rises 1.6% to $34.1 trillion. Labour's share slips from 60% to 59.4%, which is barely a move at all.
• Substantial: AI can do half of all knowledge work by 2030, most of it on its own, but adoption stays incomplete. Output rises 8.3% to $36.3 trillion, about double the normal growth rate. Labour's share is 56.1%. Pay for knowledge workers is essentially flat.
• Extreme: AI is more productive than humans at the vast majority of knowledge-work tasks, spreads fast, and improves itself. Output rises 32.4% to $44.4 trillion, growth of 15% a year. Labour's share falls from 60% to 45.2%, and capital takes the 15 points in between. Knowledge-worker wages drop by more than 10% and unemployment goes past typical recession levels.

The striking figure is not the growth rate but who the growth goes to. In the extreme case the economy expands by a third while labour's share of it shrinks. Record output and falling pay appear in the same column.

What the model actually counts

The model treats jobs not as occupations but as bundles of tasks, built on the US Department of Labor's O*NET taxonomy. AI can do one of four things to a task:

• Leave it alone. Some tasks stay human, such as bathing a patient.
• Augment it. A person does the same work faster or better.
• Automate it. The task gets finished without a person.
• Create new ones. What technology has historically done.

As the bundle changes, the job changes with it, and so do productivity and demand for labour.

The five questions you answer

Running the tool means setting five dials yourself: what share of tasks AI can perform, how widely it is used across the economy, how much of the work it does autonomously, how many times more productive it makes a worker (1.5x to 10x), and how long a displaced worker takes to find new work (one to two months, out to more than three years). Five answers, one curve.

What 10,000 Americans said

Anthropic put the same questions to more than 10,000 Americans through Morning Consult. The median set of answers lands near the substantial scenario: output in 2030 about 10% higher than it would have been without AI, unemployment around 5%. Roughly 10% of respondents expect the extreme case.

Anthropic's own caveat

The company states plainly that the model is a stark simplification of a complex reality. Left out: policy responses, business cycles, aggregate demand and financial market disruptions, catastrophic risks, highly capable robots, and the demand effects of the data centre buildout. Tracking is at a coarse occupational level, not per worker. In Anthropic's words, actual outcomes may differ materially.

Co-founder Jack Clark told NPR that the technology will keep developing quickly but that diffusion will be harder than people expect. Anton Korinek, who heads the company's work on transformative AI economics, makes the same point from the other side: AI that nobody uses has no economic effect, however capable it is.

What this actually is

Most AI companies either stay quiet on economic impact or offer a single optimistic number. What is different here is that the company building the model also calculated and published the scenario where wages fall, and handed the assumptions to the reader. It moves the argument from whether this happens to which assumptions you hold, which is where the real disagreement was all along.

Sources: Anthropic scenario explorer, NPR.