# Economic Scenarios for Transformative AI

> Source: https://aiwiki.ai/wiki/economic_scenarios_for_transformative_ai
> Updated: 2026-09-10
> Fact-checked: 2026-09-10
> Categories: AI Research, Anthropic
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Economic Scenarios for Transformative AI." aiwiki.ai, 10 Sept 2026. https://aiwiki.ai/wiki/economic_scenarios_for_transformative_ai
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Economic Scenarios for Transformative AI** is a September 2026 working paper and interactive scenario model from [The Anthropic Institute](https://aiwiki.ai/wiki/anthropic_institute). Written by Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter, and Peter McCrory, it maps assumptions about AI capability, adoption, task-level productivity, automation, and worker reallocation into model-implied paths for United States output, wages, factor shares, employment, and unemployment through 2030.[1][2] [Anthropic](https://aiwiki.ai/wiki/anthropic) announced the accompanying scenario explorer on September 9, 2026.[3]

The authors state that the three named scenarios, modest, substantial, and extreme, are not predictions and have no assigned probabilities. Their purpose is to make the economic consequences of different assumptions comparable.[1] The figures below are therefore conditional outputs of one model, not forecasts, measured economic effects, or Anthropic's estimate of the most likely future.

## Publication and purpose

The report is The Anthropic Institute Working Paper No. 2026-02. Its companion explorer is labeled version 1.0 and lets users change the inputs to see the resulting paths.[1][2] The work is distinct from the [Anthropic Economic Index](https://aiwiki.ai/wiki/anthropic_economic_index), which measures current Claude usage. The scenario model starts from mid-2026 estimates and then explores hypothetical paths to 2030.[1]

The report addresses two connected questions. First, how much additional output could result if AI reached and diffused across more tasks? Second, how might that output be divided among cognitive workers, other workers, and owners of capital when some AI use augments labor and some automates it? The model also represents the time required for displaced workers to find jobs in a different occupational group.[1]

## Model structure

The production block is a task-based constant-elasticity model in which tasks are gross complements. Each task instance can be performed by labor or capital. AI can lower the unit cost of a task either by augmenting a worker or by letting capital perform the task outright. New labor-intensive tasks can partly offset automation, following the displacement and reinstatement framework developed by Daron Acemoglu and Pascual Restrepo.[1][4]

Workers are divided into two groups. The cognitive group consists of 12 major Standard Occupational Classification groups covering management, professional, sales, and office work. It represented 62.4 percent of employment in the authors' 2025 Current Population Survey calibration. All remaining occupations form the second group. Only cognitive work is directly exposed to AI in the model, although displaced cognitive workers can search for jobs in the other group.[1]

The labor-market block uses search and matching. Its cross-occupation search discount represents how much harder it is to find work in a different occupational group. A real-wage rigidity parameter determines how much a fall in demand for cognitive labor appears as lower wages rather than layoffs. The baseline capital-supply elasticity is 3, and the no-AI counterfactual has annual GDP growth of 2 percent.[1]

The model also includes a semi-endogenous innovation channel in which AI can increase research input and the stock of ideas. This channel contributes relatively little by 2030 because physical tasks remain a bottleneck and the model omits several feedback loops that could accelerate research and further automation.[1]

## Scenario assumptions

The affected-mass and diffusion paths share mid-2026 anchors but use different assumed 2030 values. The affected task mass begins at 0.14, the estimated economy-wide average in Anthropic's observed-exposure measure. Diffusion begins at 0.10. The authors describe the lower diffusion anchor as a deliberate adjustment because the share of firms using AI is not the same as the share of individual task instances performed with AI. The log productivity gain has different mid-2026 anchors: 0.30 in the modest scenario, 0.35 in the substantial scenario, and 0.45 in the extreme scenario.[1]

| Assumed input | Modest | Substantial | Extreme |
|---|---:|---:|---:|
| Affected task mass in 2030 | 0.20 | 0.30 | 0.50 |
| Diffusion among affected task instances in 2030 | 0.20 | 0.40 | 0.60 |
| Economy-wide task instances performed with AI in 2030, affected mass times diffusion | 4% | 12% | 30% |
| Log productivity gain per AI-performed instance in 2030 | 0.30 | 0.45 | 0.80 |
| Share of AI-performed instances that are automated | 0.50 | 0.75 | 0.90 |
| New human tasks per automated task | 0.50 | 0.25 | 0 |
| Cross-occupation search discount | 0.17 | 0.08 | 0.04 |
| Job-posting speed per month | 0.10 | 0.25 | 0.50 |

These rows are inputs selected by the authors, not estimates produced by the model. The model holds its cognitive-wage rigidity at 0.50 and its capital-supply elasticity at 3 in all three headline scenarios.[1] The affected-mass and diffusion paths are logistic curves between their common mid-2026 anchors and the assumed 2030 values, while the log productivity gain follows a linear path.[1]

## Model outputs for 2030

The following results are from Table 3 of the working paper. They apply at the start of 2030. Growth rates are log changes over the preceding 12 months, and rows labeled as above or below the no-AI path compare the scenario with the model's counterfactual rather than with observed 2026 values.[1]

| Model output | No AI | Modest | Substantial | Extreme |
|---|---:|---:|---:|---:|
| GDP above the no-AI path | 0% | 1.6% | 8.3% | 32.4% |
| Annual GDP growth | 2.0% | 2.4% | 5.4% | 15.4% |
| Average wage above the no-AI path | 0% | 0.7% | 2.1% | 9.7% |
| Cognitive wage relative to the no-AI path | 0% | 0.4% | -0.3% | -11.5% |
| Other-worker wage relative to the no-AI path | 0% | 1.1% | 5.9% | 33.6% |
| Labor share of income | 60.0% | 59.4% | 56.1% | 45.2% |
| Capital share of income | 40.0% | 40.6% | 43.9% | 54.8% |
| Cognitive employment change since mid-2026 | 0% | -0.5% | -3.9% | -21.5% |
| Cognitive-worker unemployment rate | 2.9% | 2.9% | 4.5% | 17.9% |
| All-worker unemployment rate | 3.8% | 3.9% | 4.6% | 11.9% |

Output rises in every scenario, but the distribution changes. In the substantial scenario, the model puts total labor income 1.4 percent above its no-AI path while the cognitive wage bill is 4.6 percent below that path. In the extreme scenario, total labor income is only 0.5 percent above the no-AI path even though GDP is 32.4 percent higher; the cognitive wage bill is 31.0 percent lower and capital income is 81.4 percent higher. Those results follow from the scenario's assumptions about automation, task creation, capital supply, wage adjustment, and occupational mobility. They do not establish that those assumptions will occur.[1]

The direct production channel drives most of the modeled gains through 2030. The stock of ideas is only 0.07, 0.20, and 0.61 percent above the no-AI path in the modest, substantial, and extreme scenarios, respectively.[1]

## Survey of US adults

Morning Consult fielded an online survey for the authors in eight daily samples from August 11 through August 23, 2026. After retaining each respondent's first response, the sample contained 10,980 unique US adults. Responses were weighted to the adult population by age, gender, education, race, region, and income. The questionnaire had been developed through three earlier pilots of about 2,000 respondents each.[1]

Respondents were randomly assigned questions about either 2027 or 2030. Only 2030 answers were used to create model parameters. The survey asked when AI could perform eight tasks ranging from routine business writing to a Nobel-level scientific discovery, how much capable AI would be adopted at work, whether it would automate or augment five example tasks, how much time it would save on a suitable task, and how long an AI-displaced worker would need to find work in another occupation.[1]

The median respondent expected AI to be capable of six of the eight listed tasks by 2030, producing an affected-mass parameter of 0.44 after the authors' mapping to the cognitive share of work. Median diffusion was 0.40, the automation share was 0.47, the log productivity gain was 0.44, and the cross-occupation search discount was 0.064, corresponding to a reemployment time of about eight months. Views were dispersed: about 30 percent expected no time saving on a suitable task, while 49 percent expected AI to cut the time at least in half.[1]

For the 3,259 respondents who answered all five parameter questions, the authors ran a separate model simulation for each complete response. The median model-implied outcomes were GDP 8.6 percent above the no-AI path, cognitive employment 4.2 percent lower than in mid-2026, a 57.2 percent labor share, and unemployment rates of 4.6 percent for both cognitive workers and all workers. Respondents did not directly predict those economic outcomes. The simulation supplied four parameters that the survey did not ask about, using the substantial scenario's assumptions for task reinstatement, job-posting speed, wage rigidity, and capital-supply elasticity.[1]

## Sensitivity to unresolved parameters

The headline results depend materially on values that are not directly measured. When the authors rerun the extreme scenario with a capital-supply elasticity of 1 instead of 3, GDP is 21.3 percent above the no-AI path and the average wage is 9.2 percent below it. With perfectly elastic capital supply, the corresponding results are 43.3 percent and 30.1 percent above the no-AI path. The baseline extreme values, 32.4 percent for GDP and 9.7 percent for the average wage, lie between them.[1]

Wage rigidity changes whether disruption appears mainly in wages or unemployment. In the extreme scenario with flexible cognitive wages, the cognitive wage is 42.2 percent below its no-AI path and cognitive unemployment is 2.6 percent. Under the most rigid wage setting tested, the cognitive wage is 2.8 percent above the no-AI path and cognitive unemployment is 24.0 percent. The headline setting produces a cognitive wage 11.5 percent below the counterfactual and cognitive unemployment of 17.9 percent.[1] These ranges show why the scenario table should not be read as a set of point forecasts.

## Relationship to measured evidence

The report uses measurements only as starting anchors and calibration evidence. Anthropic's observed-exposure measure combines theoretical task capability with work-related Claude usage and gives less weight to augmentative than to automated use. In March 2026, the researchers behind that measure reported no systematic increase in unemployment for highly exposed occupations since late 2022, although they found suggestive evidence of slower hiring among workers aged 22 to 25 in exposed occupations.[5] Because it is based partly on one provider's usage, observed exposure is not a census of all AI use.[1][5]

The diffusion anchor also draws on the US Census Bureau's 2026 AI supplement to the Business Trends and Outlook Survey. For November 2025 through January 2026, 18 percent of firms reported AI use in a business function, or 32 percent when weighted by employment. Workers used AI for work-related tasks in 23 percent of firms, or 41 percent employment-weighted. Among adopting firms, 57 percent used AI in no more than three business functions, and 66 percent reported using it only to augment tasks.[6] The scenario paper sets diffusion to 0.10 because firm-level adoption can cover only a small portion of a firm's task instances.[1]

Independent research illustrates why narrow workplace findings do not determine aggregate outcomes. A study of 5,172 customer-support agents found that access to a generative AI assistant increased issues resolved per hour by 15 percent on average, with larger gains among less experienced and lower-skilled workers.[7] A Danish study linking adoption surveys to administrative records found no detectable effects on earnings or recorded hours at the worker or workplace level during the first two years after ChatGPT's release, ruling out effects larger than 2 percent, while documenting task reorganization and new AI-related work.[8] These studies measure particular early settings. They neither validate nor refute the paper's conditional 2030 scenarios.

The paper's authors compare their modest scenario with Daron Acemoglu's task-based macroeconomic estimate, which put AI-driven total factor productivity gains at no more than 0.66 percent over ten years and argued that accounting for harder tasks could place the gain below 0.53 percent.[1][9] The horizons and reported outcomes are not identical, so the estimates cannot be compared as if they used the same experiment.

## Limitations

The model is deliberately aggregate. It treats workers as members of only two occupational groups, assigns one wage within each group, and does not follow individuals or represent losses of occupation-specific skill. It does not model differences across firms, regions, demographic groups, or levels of job tenure. Capital is one composite good on an exogenous supply schedule, with no separate treatment of compute and no household saving decision.[1]

Several potentially important channels are outside its scope. The framework omits policy responses, business cycles, price rigidity and aggregate-demand feedback, financial-market disruption, catastrophic risks, household production, and rapid progress in robotics or physical-task automation. The latter omission is one reason the analysis stops at 2030. Its innovation block omits feedback that could connect research progress to faster automation, so the authors describe its research-driven growth effect as a likely lower bound within the model.[1]

The working paper credits outside economists for comments and says Claude was used as a research and writing assistant. The authors state that the views are their own and do not necessarily represent Anthropic or The Anthropic Institute. The explorer separately says that external reviewers were not asked to endorse the conclusions.[1][2] Actual outcomes could differ because the assumed AI paths may not occur, the calibrated relationships may change, and important economic mechanisms are absent.

## See also

- [Anthropic](https://aiwiki.ai/wiki/anthropic)
- [The Anthropic Institute](https://aiwiki.ai/wiki/anthropic_institute)
- [Anthropic Economic Index](https://aiwiki.ai/wiki/anthropic_economic_index)
- [Artificial Intelligence](https://aiwiki.ai/wiki/artificial_intelligence)
- [AI governance](https://aiwiki.ai/wiki/ai_governance)

## References

1. Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter, and Peter McCrory. "Economic Scenarios for Transformative AI." The Anthropic Institute Working Paper No. 2026-02, September 2026. https://www-cdn.anthropic.com/files/4zrzovbb/website/cf58f84d46a4a76bf5a5b039ac695fba6b80041c.pdf
2. The Anthropic Institute. "What will our economic future look like?" Scenario explorer, version 1.0, September 2026. https://www.anthropic.com/institute/econ-scenarios
3. Anthropic. "Anthropic's Economics team is sharing a new model of how AI might affect economic growth, jobs, wages, and more by 2030." X, September 9, 2026. https://x.com/AnthropicAI/status/2097679796687769689
4. Daron Acemoglu and Pascual Restrepo. "Automation and New Tasks: How Technology Displaces and Reinstates Labor." Journal of Economic Perspectives 33, no. 2, 2019, pp. 3-30. https://doi.org/10.1257/jep.33.2.3
5. Maxim Massenkoff and Peter McCrory. "Labor market impacts of AI: A new measure and early evidence." Anthropic, March 5, 2026. https://www.anthropic.com/research/labor-market-impacts
6. Kathryn Bonney, Cory Breaux, Emin Dinlersoz, Lucia Foster, John Haltiwanger, and Aditya Pande. "The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks." US Census Bureau Center for Economic Studies Working Paper CES-26-25, April 2026. https://www2.census.gov/library/working-papers/2026/adrm/ces/CES-WP-26-25.pdf
7. Erik Brynjolfsson, Danielle Li, and Lindsey Raymond. "Generative AI at Work." Quarterly Journal of Economics 140, no. 2, 2025, pp. 889-942. https://doi.org/10.1093/qje/qjae044
8. Anders Humlum and Emilie Vestergaard. "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI." NBER Working Paper 33777, May 2025, revised March 2026. https://doi.org/10.3386/w33777
9. Daron Acemoglu. "The Simple Macroeconomics of AI." Economic Policy 40, no. 121, 2025, pp. 13-58. https://doi.org/10.1093/epolic/eiae042