Anthropic built a tool that shows AI adding $10 trillion to the US economy by 2030 -- and cutting knowledge workers' pay by double digits in the same scenario

Anthropic published an interactive model on September 9 called the Econ Scenario Explorer, built by its in-house Anthropic Institute to let anyone adjust assumptions about AI capability, adoption, and autonomy and watch the implied 2030 US economy change in real time. The underlying working paper -- authored by economists Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter and Peter McCrory -- models occupations as bundles of tasks and tracks four things AI can do to each one: leave it unchanged, augment it, automate it, or create new work in its place. Anthropic is explicit about the caveat that matters most: these are scenarios, not forecasts, deliberately simplified to isolate a few forces while leaving out policy responses, business cycles, and financial-market disruption.

The tool sets out three named paths. The modest scenario -- AI's impact roughly comparable to the internet's -- puts 2030 GDP at $34.1 trillion, just 1.6% above a no-AI baseline. The substantial scenario, where AI autonomously performs roughly 50% of knowledge work but adoption stays incomplete across sectors, reaches $36.3 trillion, an 8.3% lift. The extreme scenario -- AI outperforming humans at most knowledge tasks and doing nearly all of them autonomously, with recursive self-improvement accelerating the whole process -- reaches $44.4 trillion, a 32.4% lift, with annual growth near 15% and the economy effectively doubling every four and a half years.

The task-bundle framework is what makes the tool more than a GDP slider. Every occupation is treated as a set of individual tasks, and the model tracks what AI does to each one: leave it unchanged, augment it (a worker still does the task, faster or better), automate it (AI does the task without them), or create an entirely new task that didn't previously exist. The modest and substantial scenarios lean on augmentation and incomplete automation -- jobs change shape but mostly persist. The extreme scenario is the one where automation dominates and, per the model's own reported figures, the "create" category stays close to zero -- a detail that reads as the mechanical reason extreme growth and extreme labor displacement arrive together: almost no new human work is being generated to absorb the workers automation displaces.

The fastest growth and the worst wage outcome are the same case

This is the finding Anthropic itself puts front and center, not something a critic had to dig out: the extreme scenario's exceptional growth arrives alongside the steepest decline in labor's share of that growth. In the substantial case, knowledge-worker wages are roughly flat, unemployment stays near typical levels around 5%, and labor's share of national income slips from a 60% baseline to 56.1%. In the extreme case, that share falls further, to 45.2% -- and non-AI-exposed workers see earnings grow more than 33% while knowledge workers see wages fall more than 10% and, under a version of the model that holds wages sticky rather than letting them fall freely, knowledge-worker unemployment specifically can reach the low-to-mid twenties in percentage terms. The paper is candid that this isn't a side effect the model failed to avoid -- it's the mechanism: automation genuinely replacing cognitive labor is what produces the extreme case's growth rate in the first place, and replaced labor is exactly the labor that stops collecting a wage.

Anthropic paired the tool's release with a survey of 10,980 Americans, asking people to estimate where the economy is actually headed. The typical respondent's answer landed closest to the substantial scenario -- the middle path, not the extreme one -- while roughly one in ten respondents' expectations matched the extreme case. That's a real data point about public sentiment sitting inside a company-authored economic model, worth reading as exactly that: what a broad public expects, gathered and published by the same lab whose product is the technology being modeled.

In the extreme scenario, the gains from a rapidly expanding economy are unevenly distributed.

It's worth stating the obvious tension plainly rather than skating past it: the company modeling how much of your job AI could take, and how much of your wage could go with it, is also the company selling the AI. That doesn't make the model dishonest -- Anthropic disclosing its own worst-case labor outcome, in public, with the actual percentages attached, is a harder thing to do than staying quiet about it, and the sticky-wage and public-survey additions read as genuine attempts to show the model's assumptions rather than bury them. But a reader weighing how much to trust the specific numbers should weigh them as one lab's own economists' scenario analysis, not as a neutral government projection -- which is exactly the distinction Anthropic's own "scenarios, not forecasts" framing is asking for.