Anthropic Lets Anyone Run Its AI Economic Scenarios for 2030

The company says it built the tool not to sell a forecast but to expose the assumptions behind one. The explorer asks users to enter values themselves. Pick how far AI's capabilities reach, how widely it is adopted, whether it assists people or replaces them, how much productivity rises, and how long it takes someone who changes occupations to find a new job, and the tool produces a matching 2030 economy. Two people using the same tool can see entirely different futures, and the tool shows where the difference comes from.

The company's default scenarios, which METAL reviewed, come in three branches. In the modest scenario, AI's effect is hard to see in macro indicators. As with the internet, real gains arrive, but within the historical range of new technologies and slowly. In the substantial scenario, AI becomes capable of half of all knowledge work by 2030 and handles most of that on its own. But being capable is not the same as being used, so most knowledge-work tasks are still done without AI. The economy grows at twice its usual pace. In the extreme scenario, AI is more productive than humans on most knowledge-work tasks, handles almost all of them autonomously, and creates virtually no new tasks for people.

By the numbers, the third scenario is a different world. Annual growth reaches 15% and the economy doubles every four and a half years. Society as a whole becomes the richest in history, but jobs remaining in knowledge work shrink sharply and unemployment exceeds the level of a typical recession. The company states plainly that this scenario assumes recursively self-improving AI and rapid adoption. METAL reported Anthropic's alignment lead putting the probability of human extinction within ten years above 10%; that the same company's economic model puts the same technology at the base of its brightest growth curve is worth reading side by side.

Where does the general public land? The company surveyed more than 10,000 Americans in August. Plugging the typical respondent's answers into the model yields a result close to the substantial scenario: GDP in 2030 about 10% higher than it would be without AI, and overall unemployment rising to around 5%. About 10% of respondents held views that correspond to the extreme scenario.

The most sociological part of the model is not the growth rate but the distribution. Average wages rise in all three scenarios. But the gains cluster outside knowledge work. As demand for knowledge work falls, wages there come under downward pressure, and at the same time, when AI finishes design and permitting work quickly, there is more actual construction to be done and demand for construction labor rises. In the substantial scenario, knowledge workers' wages are effectively flat. In the extreme scenario they fall more than 10% by 2030.

The number of people who have to change occupations also varies by scenario. The model pictures coders and call-center agents moving into occupations with low AI exposure, like electricians or nurses. The problem is that moving itself is hard. People do not want to change occupations, they have to learn new skills, and even after learning them, landing a position is not easy. The more people who have to switch in a scenario, the more people end up suspended between occupations. In the extreme scenario, knowledge work is automated quickly and affected workers can remain unemployed for a long time.

The question of shares also shows up in numbers. Today, when the US economy produces a dollar, about 60 cents goes to labor and 40 cents to capital. As more tasks are automated, that ratio tilts toward capital, because capital becomes useful for more things, demand for it rises, and its price goes up. The model sees labor's share falling noticeably in both the substantial and extreme scenarios. In the extreme scenario, even as the economy grows rapidly, the slice workers collectively take home gets smaller, and total labor income in 2030 is roughly unchanged. The pie grows, but the workers' plate stays the same. Of this scenario the company wrote that "the central challenge is not achieving economic growth, but ensuring that its benefits are broadly shared and its costs are not unevenly scattered."

It is worth noting that the company wrote down what is missing. This is version 1.0, and it leaves out policy responses, business cycles, aggregate demand and financial-market shocks, and catastrophic risk. There is no scenario with highly capable robots. The list of outside economists who read the draft includes Daron Acemoglu and David Autor, and the company says it did not ask them to endorse the conclusions. It also published the criticisms that remain: that the model does not track individuals and so draws the cost of job loss only very roughly; that it is unclear whether occupations exposed to AI will really shrink or instead grow; that "the most extreme scenario is better read as a thought experiment than a scenario"; and conversely that the modest scenario underestimates changes already visible in the data. There was also a request to include the aggregate-demand effect of data center construction, which is absent.

The people who built the model span Anthropic and beyond. Anton Korinek, Chad Jones, Simon Sacher, Tess Cotter, and Peter McCrory handled the economic model and technical report. Kelsey Nanan designed the interactive interface and Kyle Turman built the explorer. The company says the model will underpin the labor-market research and policy proposals it plans to support going forward.

To sum up: Anthropic released a model that reduces the US economy to a bundle of tasks and concluded that 2030 splits three ways depending on how far AI goes. In all three branches the pie grows but knowledge workers' share shrinks, and the brightest growth curve assumes recursive self-improvement. Two things to watch from here: whether the next version the company has promised actually fills in the missing pieces, and how the company's policy proposals end up quoting this model.