Anthropic's Three Economies: What AI Could Do to Jobs by 2030

America could be dramatically richer by 2030 while office workers suffer unemployment on a scale no modern boom has ever combined with rapid growth, and that paradox drives Anthropic's 57-page economic report . The paper sketches three scenarios for 2030 — modest, substantial, and extreme — while attaching no probabilities to any of them. The modeling comes from Anton Korinek and Charles Jones of Stanford , with outside review from Daron Acemoglu and David Autor , who read the draft without endorsing its conclusions. A University of Virginia publication profiles Korinek's role and Virginia connection in detail. Coverage by Euronews summarizes the same no-odds, three-scenario setup for a broad audience.

Forget industries and occupations for a moment and picture a single nurse, because the whole model is built upward from that level. Her job is a bundle of tasks : rounds, blood draws, triage, charting vital signs, ordering ward supplies, plus a dozen smaller duties. Some duties no software can touch, such as bathing a patient; others, like drafting discharge notes , are natural candidates for assistance; repetitive chores such as charting vitals can be fully automated, while oversight creates brand-new checks such as verifying the algorithm's triage. Multiply her shifting bundle across millions of workers and shifts, and the shrinking gray zone of non-adoption becomes the bridge from one hospital ward to a thirty-trillion-dollar economy.

Everything that follows turns on five dials : capability, the share of knowledge work AI can perform at professional quality; adoption, how much of that capability firms actually use; autonomy, whether the system assists a person versus replacing them ; productivity, how much faster each touched task becomes; and adjustment, how long displaced workers need to land elsewhere. Beneath them sits a sixth factor: for every automated task, history suggests roughly one new human task for every two destroyed , while the harshest run sets that replacement ratio near zero. An Anthropic research update on economic measurement provides useful context on how the lab tracks capability and autonomy in real usage.

How the model reads jobs and expectations

The dials are not just theory, because the team polled 10,980 Americans on exactly those five questions, separating the general public from readers who sought the paper out independently. The typical respondent expects AI to master most knowledge tasks yet be adopted on only about forty percent of them, to assist rather than replace roughly half the time, to cut task time by around a third, and to leave displaced workers searching for about eight months . Fed through the model, that median voter lands almost exactly on the substantial scenario : roughly nine percent more output with unemployment below five percent. About one in ten respondents gave answers implying the extreme world. A Political outlet's explainer walks through the same survey design and scenario mapping.

The three scenarios are simply three settings of the same machine. In the modest world , AI masters about a fifth of knowledge work but is used on only a fifth of that, speeds tasks by roughly a third, splits evenly between help and replacement, and creates new tasks at the historical rate — an impact resembling the internet's gradual diffusion . The substantial world lets AI handle about half of knowledge work with lagging adoption, fifty-plus percent speedups, and mostly autonomous operation, growing at twice the normal pace. The extreme world assumes near-total mastery, majority adoption, more than doubled productivity, nine-in-ten autonomous operation, almost no new tasks, and self-improving systems. Fortune's reporting lays out the same three dial settings and their very different economic flavors.

Finding one is that output rises in every world, just by wildly different amounts: about $34.1 trillion and 1.6 percent above baseline in the modest case, near $36 trillion and 8 percent in the substantial case, and roughly $44 trillion and 32 percent in the extreme case. Picture each automated task as a colored tile and each speedup as a bar: the extreme grid holds perhaps triple the tiles but a bar twenty times longer. Crucially, the paths look identical until 2027 because they share today's readings, so the next eighteen months of data reveal which line we are on. Substantial growth reaches 5.4 percent, above the dot-com peak; extreme growth hits 15 percent, doubling output every four and a half years. Axios's chart-driven coverage grounds the same GDP figures and near-term branching point.

Growth, displacement, and who gets paid

Finding two follows the people rather than the output, as office, professional, sales, and management workers — roughly sixty-two percent of employment — are pushed toward hands-on jobs the model deems sheltered. In the modest world barely three-tenths of a percent leave knowledge work; in the substantial world about two and a half percent are pushed out with seven-tenths still searching; in the extreme world over thirty percent are displaced and eight percent of the entire workforce remains stuck between occupations. Consider Dan the bookkeeper , whose small-business ledgers go autonomous, forcing a pivot toward electrician-style physical work that proves slow because pay adjusts sluggishly and outside-occupation search carries a hiring penalty. Knowledge-worker unemployment hits 17.9 percent in the extreme case versus under four percent elsewhere, while manual unemployment paradoxically falls.

Finding three shows averages concealing a split, since every scenario lifts mean pay while distributing it brutally unevenly. In the modest world everyone gains just under one percent ; in the substantial world knowledge pay sits flat to slightly down against baseline while other workers rise nearly six percent . Remember the baseline itself climbs about two percent a year, so flat means a smaller raise rather than a cut — until the extreme case, where knowledge pay drops 11.5 percent against trend and lands below today's level while manual pay surges thirty-four percent . Faster design and permitting multiply construction projects, bidding up scarce builders while displacing the architects and clerks. The lesson is that AI shifts money from words-and-numbers work toward hands-on work and, soon after, toward whoever owns the systems.

Capital, caveats, and the signal to watch

Finding four divides each output bar into labor and capital, and it is the starkest chart of all. Today about sixty cents of every dollar flows to wages and forty to owners of buildings, machines, software, and data centers — a split stable for decades after drifting only slightly since 1980. The modest case trims labor to 59.4 cents , barely noticeable; the substantial case drops it to 56 cents , compressing four decades of drift into four years; the extreme case pushes it to 45 cents , so capital takes more than labor for the first time in modern history. Strikingly, total worker pay stays near $20 trillion in all three worlds: essentially all the extra growth accrues to capital as each automated paycheck migrates columns.

The honest coda lists what the machinery omits: no robots , so capable physical automation would spread displacement into today's safe harbor; no policy response , no recessions, and no demand stimulus from data-center construction; plus reviewers who split, some calling the extreme path a thought experiment while others warn that science itself could accelerate. Capital supply is the hidden lever — easy-to-build compute pushes gains back toward wages, while scarce capital lets owners keep them. So the variable to watch is not raw model intelligence but the augmentation-versus-automation choice : whether firms use AI to help the same people produce more or to produce the same output with fewer people.