In September 2026, Anthropic released an "Economic Scenario Explorer" tool that estimates how AI progress could affect the U.S. economy through 2030. By changing assumptions about capability advances and enterprise adoption, users can see how GDP, wages, and unemployment might shift. Though the research comes from a company whose CEO, Dario Amodei, has warned of large-scale job losses, what stands out is not whether his pessimistic forecast is right or wrong, but a sharper question: who actually captures the gains from growth?
Even when the economy expands rapidly, workers' total income may barely rise. Reading further into the paper reveals another path to worker losses—one where unemployment stays low, but wages fall instead.
Was the CEO's Pessimistic Scenario Judged "Low Probability"?
In a May 28, 2025 interview with Axios, Amodei warned that within one to five years, half of entry-level white-collar jobs could disappear, pushing unemployment to 10–20%. The Decoder, which covered this new model, suggested that because this figure resembles the "extreme" scenario, the CEO's forecast has effectively been recast as the least likely outcome.
However, the paper states upfront that these are scenarios, not predictions, and no probabilities are assigned to any of them. "Extreme" describes the magnitude of economic change, not a statistical judgment of likelihood. Presenting three scenarios—modest, large, and extreme—doesn't mean the middle case is the most probable.
The unemployment rate's denominator also matters. The extreme scenario's 17.9% figure applies specifically to workers who originally held knowledge-work jobs. Across the whole economy, the figure is 11.9%. The CEO's overall unemployment estimate and his figure for entry-level job losses cannot be directly compared to this paper's knowledge-worker unemployment rate.
The research was compiled by five authors, including Anton Korinek and Charles I. Jones, and published as an Anthropic Institute working paper—not a peer-reviewed academic article. While it received comments from notable economists, the publication page explicitly states that it did not seek external researchers' endorsement of its conclusions. Nor does it serve as evidence that the company has retracted the CEO's remarks.
Between AI Capability and Real-World Adoption
At the core of the paper is a framework that treats occupations as bundles of discrete tasks. It distinguishes between AI assisting human work and AI replacing human tasks, while also incorporating the effect of new jobs created for humans after automation. Adding the time required for job transitions, the model captures a situation where productivity rises even as unemployment increases.
Capability gains alone don't mean every company's operations change overnight. In the large-change scenario, by 2030 AI is assumed to be technically capable of handling 30% of tasks across the economy, with 40% of those capable tasks actually put into use. Multiplying these—30% × 40%—yields 12%. This doesn't mean 12% of jobs disappear; rather, it represents the share of task-units across the economy where AI is actually deployed.
Of that AI usage, the share functioning as automation (versus augmentation) also varies by scenario: 50% in the modest scenario, 75% in the large scenario, and 90% in the extreme scenario. In the extreme scenario, the effect of new human jobs being created is set to zero, and job-searching for career changes becomes harder. It's a calculation where technological advancement and conditions that make worker adaptation difficult compound each other.
| Early 2030 Outcomes | Modest | Large | Extreme |
|---|---|---|---|
| GDP boost (vs. no-AI baseline) | 1.6% | 8.3% | 32.4% |
| Economy-wide unemployment rate | 3.9% | 4.6% | 11.9% |
| Knowledge-worker unemployment rate | 2.9% | 4.5% | 17.9% |
| Knowledge-worker wage gap (vs. no-AI baseline) | +0.4% | −0.3% | −11.5% |
| Labor share of income | 59.4% | 56.1% | 45.2% |
Source: Table 3 of the technical paper. All figures are model outputs for the United States. GDP and wage comparisons are against a "no-AI economy" at the same point in time, not a change from the present. In the no-AI baseline, overall unemployment is set at 3.8% and the labor share of income at 60.0%.
In the large-change scenario, GDP rises 8.3% above baseline, yet knowledge-worker wages fall slightly below the no-AI baseline. As companies become able to handle knowledge work with fewer people, demand for such workers weakens. Meanwhile, demand for jobs like construction may increase. But a software engineer can't necessarily move quickly into electrical work, and the slower this cross-occupation mobility, the more unemployment accumulates.
Tracking AI performance benchmarks alone misses this time lag.
When Low Unemployment Still Means Falling Wages
The extreme scenario's knowledge-worker unemployment rate can range anywhere from 2.6% to 24.0%, depending on how readily wages adjust. Table 6 of the paper compares a case where wages respond immediately to shifts in labor demand against a case where wages are slow to adjust. Even without changing assumptions about AI capability or adoption, the way labor markets adjust can dramatically shift outcomes.
| Wage-Adjustment Settings, Extreme Scenario | Instant Adjustment | Baseline Setting | Very Slow Adjustment |
|---|---|---|---|
| Knowledge-worker wage gap (vs. no-AI baseline) | −42.2% | −11.5% | +2.8% |
| Knowledge-worker unemployment rate | 2.6% | 17.9% | 24.0% |
Source: Excerpted from three columns of the extreme scenario in Table 6 of the technical paper. Values are for early 2030; all parameters other than wage rigidity are held constant. Wage gaps are compared to the no-AI economy at the same point in time; unemployment rates apply within that group.
Even when unemployment stays low, workers may still bear losses through steep wage declines. Under instant wage adjustment, demand for now-cheaper labor persists, so employment tends to be preserved. Under sticky wages, however, reduced demand tends to manifest as layoffs instead.
This isn't an argument in favor of wage cuts. It's a comparison meant to separate the burden of losing a job from the burden of losing income while remaining employed. The observation that "unemployment didn't rise even as AI spread" is not sufficient on its own to fully assess the impact on workers.
Indeed, the policy recommendations Anthropic published in June also call for monitoring wages, underemployment, and the labor share of income alongside unemployment. Whether a job survives and whether that job provides adequate income are two separate things that need to be verified independently.
Who Receives the Gains in GDP?
In the extreme scenario, GDP is 32.4% larger than the no-AI baseline, yet total labor income rises by only 0.5%. Capital income, meanwhile, rises 81.4%. The labor share of income falling from 60.0% to 45.2% means that a substantial portion of the increased output is being distributed through channels other than compensation for work.
As AI-driven automation advances, demand for capital needed for production increases. But over the timeframe this model covers—through 2030—the supply of capital doesn't expand without limit to match that demand. As a result, returns to capital rise, and more income flows to owners of capital. Workers who hold capital could also benefit, but the model doesn't detail who holds how much capital.
Rising average wages, too, don't mean improvement for everyone. In the extreme scenario, average wages rise 9.7% above baseline, but knowledge-work wages fall 11.5% while wages in other occupations rise 33.6%. And of course, the unemployed receive no wages at all. Total wage income for knowledge workers as a group ends up 31.0% smaller than in the no-AI baseline.
This is where the June policy proposals become concretely relevant. For temporary gaps during job transitions, unemployment insurance or wage insurance—which partially compensates income for workers who move to lower-paying jobs for a set period—are candidate responses. Policies that lower barriers to credentialing and occupational mobility also fit the goal of speeding reemployment.
But if demand for human labor weakens over a longer period, repeated retraining alone may not be enough to restore income. Anthropic lists universal basic income, sovereign wealth funds built on AI investment stakes, and mechanisms for workers to hold equity in companies as candidates for further research. It also notes that it isn't yet prepared to endorse any of these as specific policy—these are not settled recommendations.
Between growth generating more income to distribute and that income actually reaching people who've lost their jobs, there remains the work of building institutions.
Losses After Switching Jobs, and Crises Outside the Model
In this model, anyone who successfully switches occupations immediately earns the same wage as their new occupational group. It doesn't account for the temporary income loss that often follows a job change, caused by losing seniority or specialized skills from a previous role. Because occupations are also grouped into just two broad categories, the model can't precisely capture the burden borne by specific age groups or regions.
Business cycles and financial market turmoil also fall outside the model. It doesn't reproduce a downward spiral where unemployment and falling incomes cool consumption, which in turn worsens employment further. On the other hand, it also excludes a scenario where robotics advances rapidly enough to displace physical labor. The result showing rising wages in "other occupations" shouldn't be taken as a guarantee that such jobs will remain safe going forward.
The accompanying public opinion survey also doesn't verify the accuracy of any forecast. It gathered responses from 10,980 U.S. adults about their expectations for AI, and for the individual-level economic outcome tally, fed the model inputs from the 3,259 respondents who answered all five relevant questions. Respondents didn't predict GDP or unemployment figures themselves—the equations converting responses into economic figures were supplied by the researchers.
Given these constraints, the model's value lies less in pinpointing a future unemployment rate down to the decimal and more in preparing appropriate responses for each set of assumptions. Since most of the differences between scenarios emerge after 2027 in the paper, the fact that near-term employment effects appear small doesn't rule out greater disruption later.
Will companies' AI adoption move toward having the same number of people do more work, or toward reducing headcount? Are the incomes of workers who've changed occupations recovering? By tracking this progression alongside productivity gains, and by distinguishing stages where transition support suffices from stages requiring changes to income-distribution systems, policymakers can better choose approaches that translate AI-driven growth into real improvements in workers' lives.