Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter, and Peter McCrory published Economic Scenarios for Transformative AI this month as The Anthropic Institute Working Paper 2026-02. The authors write that the views are theirs and do not necessarily represent Anthropic. Scott Horsley at NPR covered the public explorer on September 9. The model is on anthropic.com/institute/econ-scenarios.
They split the workforce into two groups. Cognitive occupations are management, professional, sales, and office jobs. Those jobs can be touched by current systems. Construction, the trades, and other work sit in the second group and are treated as unexposed through 2030. Artificial intelligence raises output on the first group, pushes some of those workers to look for jobs in the second group, and raises the return on capital because machines do more of the work that used to be paid as wages.
Five inputs drive the paths: how much of knowledge work the systems can do by 2030, how widely firms actually use them, how much faster a touched task gets done, how often the tool replaces a person instead of sitting next to one, and how long a displaced person takes to land in another occupation. Change those five and the model writes new lines for gross domestic product, wages, the labor share, and unemployment. The authors attach no probabilities. They call the three labeled paths modest, substantial, and extreme.
Modest change barely moves the labor market
On the modest path, about 4 percent of tasks in the economy have been touched by 2030. Capability and use both sit at 20 percent of the cognitive stack. Half of those uses replace a person and half assist one. For every two tasks that disappear, one new cognitive task shows up.
GDP in 2030 is 1.6 percent above a world with no artificial intelligence, and growth is 2.4 percent a year instead of 2 percent. The average wage is 0.7 percent higher. Cognitive employment is 0.5 percent below its mid-2026 level. Economy-wide unemployment is 3.9 percent against a normal 3.8. The labor share slips from 60.0 percent of income to 59.4. That is an extra ten months of ordinary growth spread over four years, with almost no change in who has a job (Korinek et al., 2026).
Substantial change is the path the survey median lands on
On the substantial path, 12 percent of tasks in the economy have been touched by 2030, about a fifth of the work cognitive occupations did in 2025. Use reaches 40 percent of what the systems can do. Productivity on a touched task is up about 57 percent. Three quarters of those uses replace a person. For every four automated tasks, one new cognitive task appears.
GDP is 8.3 percent above the no-AI path. Growth in the twelve months to 2030 is 5.4 percent a year. The fastest calendar year of the 1990s internet boom, the authors note, was 4.7 percent in 1999. The average wage is only 2.1 percent higher than the no-AI path. Cognitive wages sit 0.3 percent below that path. Wages in other occupations sit 5.9 percent above it. The labor share falls from 60 percent to 56.1 percent in four years, a move the paper compares with the entire U.S. labor-share decline across four decades after 1980.
Cognitive employment is 3.9 percent below its mid-2026 level. Employment in other occupations is 4.6 percent higher. Cognitive unemployment rises from 2.9 percent to 4.5 percent because people cannot switch occupations on the same timetable the software ships.
Output can rise while the knowledge-work wage bill falls, if capital is slow to expand and people are slow to change jobs.
Extreme change moves a large slice of income off wages
On the extreme path, 30 percent of tasks in the economy have been touched by 2030, about half of 2025 cognitive work. Use reaches 60 percent of what the systems can do. A touched task more than doubles in productivity. Ninety percent of those uses replace a person. No new cognitive tasks are put back into the model.
GDP is 32.4 percent above the no-AI path. Growth in 2030 is 15.4 percent a year. The average wage is 9.7 percent higher than the no-AI path, but that average hides the split. Cognitive wages are 11.5 percent below the no-AI path and slightly lower in 2030 than in mid-2026. Wages in other occupations are 33.6 percent higher. The labor share falls from 60 percent to 45.2 percent. The capital share rises to 54.8 percent. Fifteen points of national income leave the wage bill and show up as a return on investment. The net return on capital moves from 6.5 percent to 8.3 percent.
Cognitive employment is 21.5 percent below its mid-2026 level. Cognitive unemployment is 17.9 percent. Economy-wide unemployment is 11.9 percent. Annual U.S. unemployment peaked a little under 10 percent after the financial crisis and about 8 percent in the Covid recession. Total labor income on this path is almost the same as the no-AI path because the labor share shrinks about as much as GDP grows. All of the extra output accrues as capital income, which sits 81 percent above its no-AI path. The cognitive wage bill, counting the unemployed at zero, is 31 percent lower. The authors write that a transfer of about 9 percent of GDP would hold cognitive incomes at the no-AI level and still leave the rest of the economy more than 20 percent ahead. They also write that transfers of that size after a technology shock have no precedent.
The public’s answers are mixed. The outcomes they imply are not.
The team asked 10,980 U.S. adults five questions: when systems will match a professional on each of eight cognitive tasks, how widely those systems will be used at work, whether they will run a task alone or with a person, how much time they save on a suited task, and how many months a displaced worker will need to find work in a new occupation. Table 4 in the paper runs the 3,259 people who answered every item through the model.
The median adult expects the systems to handle six of the eight tasks by 2030, which is stronger than the substantial path on capability. The same person expects use on only 40 percent of what the systems can do, assistance half the time, about a one-third time cut on a suited task, and roughly eight months to land a new occupation. Run those medians and GDP in 2030 is 8.6 percent above the no-AI path. Cognitive employment is 4.2 percent below mid-2026. Unemployment, for cognitive workers and for the whole labor force, is about 4.6 percent. That cluster sits next to the substantial path, which is why the authors describe the public as landing there.
The spread under those medians is wide. About 30 percent of respondents expect no time savings on a suited task. Forty-nine percent expect the task to take at most half as long. The 25th-to-75th percentile band on implied GDP in 2030 runs from 3 percent above the no-AI path to 19 percent above it. Cognitive unemployment in that band runs from 3 percent to 8 percent.
One result that is easy to miss: the ideas channel, the part where systems speed up research and therefore long-run growth, stays small through 2030 in this model. The stock of ideas is 0.61 percent above the no-AI path even on the extreme line. The authors follow a semi-endogenous growth setup in which research still waits on physical work. They say that omission likely understates the research channel.
If capital is hard to add, wages can fall even while output rises. In a robustness run with an inelastic capital supply, the extreme path’s average wage sits 9.2 percent below the no-AI path and the net return on capital hits 10.3 percent. If cognitive wages are sticky instead, unemployment takes more of the hit. At the stickiest wage they test on the extreme path, cognitive unemployment reaches 24 percent.
The model leaves out household work that never shows up in GDP, regional differences, firm-by-firm adoption, and any demand slump that follows a wave of layoffs. It also leaves out accidents, politics, and financial stress. Treat it as a worksheet, not a forecast.
Write down the five numbers your 2027 workforce plan silently assumes: share of knowledge tasks the tools can do, share of those tasks you will actually put them on, time saved per task, share of uses that remove a seat, and months for a displaced person to land elsewhere. Put those five into the explorer. If the implied cognitive headcount in 2030 is more than 4 percent below today’s plan, the budget you approved this quarter is on a different path than the one your hiring managers are staffing.
Sources
Korinek, Anton, Charles I. Jones, Szymon Sacher, Tess Cotter, and Peter McCrory. "Economic Scenarios for Transformative AI." The Anthropic Institute Working Paper No. 2026-02, Sept. 2026, https://www-cdn.anthropic.com/files/4zrzovbb/website/cf58f84d46a4a76bf5a5b039ac695fba6b80041c.pdf.
Anthropic. "Scenarios for our Economic Future." Sept. 2026, https://www.anthropic.com/features/econ-scenarios.
Anthropic. Economic Scenario Explorer, https://www.anthropic.com/institute/econ-scenarios.
Horsley, Scott. "A new Anthropic model seeks to test how AI could impact the U.S. economy." NPR, 9 Sept. 2026, https://www.npr.org/2026/09/09/nx-s1-5961443/ai-anthropic-economy.
