Forecasts of AI's economic effect in 2026 ranged from small productivity gains to explosive growth, and they rested on different models, which made them hard to compare. This item offers one framework in which those views can be stated as different settings of a few parameters. It has two parts, published together on 9 September 2026 (the date of the first Internet Archive captures of both; the page itself shows no date):
- An interactive page on anthropic.com, headed "What will our economic future look like?" and labeled "v1.0 of the Econ Scenario Explorer, September 2026". It explains the model in plain language, shows three scenarios, and lets visitors enter their own expectations. Its credits say Santi Ruiz wrote it with the model's authors.
- A 57-page technical report, Economic Scenarios for Transformative AI, by Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter and Peter McCrory, The Anthropic Institute Working Paper No. 2026-02, dated September 2026. It states that the views are the authors' and "do not necessarily represent the views of Anthropic or The Anthropic Institute", and that the authors "used Claude as a research and writing assistant" (title page).
The technical report is the version of record for every model number in this note. The page is the source for its own framing, its survey summary and the reviewers' comments it reports.
What the item produces is of two different kinds, and this note keeps them apart. The scenarios are model outputs under assumed parameters; the report says "The scenarios are not predictions, and we attach no probabilities to them" (§1). The survey is a measurement of what US adults expect, which the authors then run through the same model.
The model
Tasks, two groups of occupations and capital
The model is a standard task-based model: production combines many tasks, each done by labor or by capital, and AI can automate a task (moving it to capital), augment it (leaving it with a worker who becomes more productive), or create new tasks for labor (§2). Workers belong to one of two groups:
- Cognitive occupations: the twelve Standard Occupational Classification major groups covering management, professional, sales and office work, 62.4% of employment in the 2025 Current Population Survey (§3.1, footnote 10). The page calls them "knowledge workers".
- All other occupations (construction, production, transportation, healthcare support and others), which the model treats as not directly exposed to AI.
AI reduces demand for cognitive labor and raises demand for capital, whose supply responds with an elasticity of 3 (Table 1), so some of the gains go to capital owners in the medium run. Displaced cognitive workers must search for jobs in the other group, and a search discount makes switching occupations harder than searching within one. A wage rigidity parameter sets how quickly the cognitive wage falls toward its market-clearing level; the more rigid the wage, the more of the adjustment shows up as unemployment rather than lower pay (§2.3, §3.4). A separate innovation block lets AI raise research input, which in the model's semi-endogenous growth setup adds little by 2030 (§1, §2.2).
What a scenario specifies
A scenario is a path from mid-2026 to 2030 for five AI objects plus two labor-market frictions (Table 1):
| Parameter in 2030 | Modest | Substantial | Extreme |
|---|---|---|---|
| Affected task mass (share of all tasks AI can affect) | 0.2 | 0.3 | 0.5 |
| Diffusion (share of affected instances done with AI) | 0.2 | 0.4 | 0.6 |
| Productivity gain per instance, in logs | 0.30 | 0.45 | 0.80 |
| Automation share of AI-performed instances | 0.50 | 0.75 | 0.90 |
| Reinstatement (new labor tasks per automated task) | 0.50 | 0.25 | 0 |
| Search discount on switching occupations | 0.17 | 0.08 | 0.04 |
| Posting speed, per month | 0.10 | 0.25 | 0.50 |
The mid-2026 starting points are measured or derived (§3.3): an affected mass of 0.14, from Anthropic's "observed exposure" measure of Claude use at work (Massenkoff and McCrory, 2026), which the report calls conservative; and diffusion of 0.10, set below the 18% of firms using AI in the Census Bureau's Business Trends and Outlook Survey because firm shares overstate the share of task instances. The normal-times search discount of 0.17 is estimated from occupational switching in the CPS, 2010 to 2019 (§3.2). The 2030 values are assumptions, chosen to span published estimates of task feasibility, field-trial productivity gains, and the share of automation-like use in the Anthropic Economic Index (about half on Claude.ai and three quarters on the API, §3.4). The wage rigidity is set at 0.5, a one-year half-life, in all three scenarios; the report says there is "no agreed value" for it (§3.4).
The report benchmarks the scenarios against outside forecasts (§1): the modest scenario lies near Acemoglu's (2025) estimate and OECD figures, the substantial scenario near financial and consulting forecasts circulating in 2023, and the extreme scenario against the literature on AGI-driven growth, including the narrative AI 2027. The page describes the extreme scenario as one that "would likely require recursively self-improving AI, adopted quickly for knowledge work."
What the scenarios produce
All figures are for the start of 2030, relative to a no-AI path, from Table 3:
| Outcome in 2030 | No AI | Modest | Substantial | Extreme |
|---|---|---|---|---|
| GDP, % above no-AI path | 0 | 1.6 | 8.3 | 32.4 |
| GDP growth, % per year | 2.0 | 2.4 | 5.4 | 15.4 |
| Average wage, % above no-AI path | 0 | 0.7 | 2.1 | 9.7 |
| Cognitive wage, % above no-AI path | 0 | 0.4 | −0.3 | −11.5 |
| Other wage, % above no-AI path | 0 | 1.1 | 5.9 | 33.6 |
| Labor share of income, % | 60.0 | 59.4 | 56.1 | 45.2 |
| Cognitive employment, % change since mid-2026 | 0 | −0.5 | −3.9 | −21.5 |
| Unemployment, cognitive workers, % | 2.9 | 2.9 | 4.5 | 17.9 |
| Unemployment, all workers, % | 3.8 | 3.9 | 4.6 | 11.9 |
The report draws these points from them (§4, §5):
- Modest. AI affects 4% of the economy's tasks by 2030; GDP gains the equivalent of about ten extra months of ordinary growth, and labor market effects are small (§4.1).
- Substantial. 12% of tasks, roughly 20% of cognitive workers' tasks, are affected. Growth of 5.4% a year exceeds the 4.7% peak of the late-1990s boom, the labor share falls four points in four years, and cognitive unemployment rises from 2.9% to 4.5% (§4.2).
- Extreme. 30% of tasks, roughly half of cognitive workers' tasks, are affected, and productivity on those tasks more than doubles. Cognitive unemployment reaches 17.9% and overall unemployment 11.9%, against annual peaks of slightly below 10% after the global financial crisis (§4.3). Total labor income is almost unchanged from the no-AI path, so all of the GDP gain goes to capital, whose income is 81% higher. The authors compute that a transfer of about 9% of GDP would hold cognitive workers' income at its no-AI level while leaving the rest of the economy more than 20% ahead, and note that transfers of that scale in response to technology have no precedent (§4.3).
- Timing. Almost all the divergence between scenarios comes after 2027 (§5).
- Innovation adds little by 2030. Through the research channel, labor productivity rises "well under one percent" in all three scenarios; the authors call this a likely lower bound, since the model cannot generate research-driven explosive growth (§1).
Robustness to the unmeasured parameters
Two parameters behind Table 3 are neither measured nor varied across scenarios, and the report reruns both:
- Capital supply elasticity (§4.5, Table 5). At an elasticity of 1 instead of 3, the extreme scenario's GDP gain falls to 21.3% and the average wage ends 9.2% below its no-AI path, with the net return to capital at 10.3%. With perfectly elastic capital, GDP is 43.3% higher and the average wage 30.1% higher. Footnote 14 constructs a case within the report's parameter ranges in which GDP rises 7.2% but labor income falls by 4.3% of no-AI GDP.
- Wage rigidity (§4.6, Table 6). In the extreme scenario, a very rigid cognitive wage (0.9, a seven-year half-life) leaves that wage 2.8% above its no-AI path but raises cognitive unemployment to 24.0%; a fully flexible wage cuts it by 42.2% with cognitive unemployment at 2.6%. The report's summary is that the cost falls on wages or on employment, and the baseline splits it.
The survey of US adults
Sample and questions
Morning Consult fielded the survey to its online panel on 11 to 23 August 2026, in eight daily samples of about 2,000; after keeping each person's first response, the sample has 10,980 unique adults, weighted to the adult population on age, gender, education, race, region and income (§3.5, Appendix B.1). Responses were screened for speed and attention. The questionnaire was developed over three pilots of about 2,000 respondents each. The report gives no response rate. Respondents were randomly assigned to answer about 2027 or 2030, and only the 2030 answers feed the model.
Five questions map onto model parameters (Appendix B.2 and B.3): when AI will match a skilled professional on eight tasks ordered from routine business emails to a Nobel-level discovery; what share of feasible tasks people will use AI for at work; whether AI will do five listed tasks alone or with workers; how much less time an AI-suited task takes; and how long a worker displaced by AI in 2030 would take to find a job in a new occupation, given that a normal search takes about three months. "Not sure" answers are dropped item by item, and the outcome distribution in Table 4 covers the 3,259 respondents who answered all five items.
Responses
- Capabilities (Figure 1, §3.5). 24% say AI can already build and maintain a software product and 19% that it can run a small online business. By 2030, 52% expect AI to run an online business with no employees and 39% a Nobel-level discovery, while 40% say such a discovery will never happen.
- Median parameters (Table 2). The median respondent expects AI to handle six of the eight tasks by 2030 (capability 0.44, between the substantial and extreme settings), use on 40% of feasible tasks, automation on about half, a log productivity gain of 0.44, and re-employment taking about eight months.
- Dispersion (§3.5). About 30% expect AI to save no time on a task suited to it; 49% expect it to at least halve the time.
- Own job versus listed tasks (Appendix B.3). Employed cognitive workers say a median 27% of their own hours go to tasks AI will do as well as they can by 2030, against 71% implied by their answers on the task grid.
What the answers imply in the model
Running each respondent's answers through the model, with the other parameters at substantial-scenario values, the median outcomes are GDP 8.6% above the no-AI path (interquartile range 3 to 19), cognitive employment 4.2% lower, and unemployment of about 4.6% both for cognitive workers and overall (Table 4, §4.4). The report concludes that the public's answers "deliver outcomes close to the substantial change scenario."
The page states the survey result in rounder and somewhat different numbers: the typical respondent's answers imply that "GDP is 10% higher by 2030" and unemployment "around 5%", and "Around 10% of respondents have views in line with the extreme scenario." The report's Table 4 gives 8.6% and 4.6%; the 10% share with extreme-scenario views is not in the report. The page may compute its figures differently (for example from the median inputs rather than the median of outcomes); it does not say. The page also shows a second respondent group, "Site visitors", with n = 33,685 when fetched on 13 November 2026; these are people who used the explorer, a self-selected and growing sample, and the report does not use them.
Reviewers' comments and stated limitations
The page names eighteen economists who commented on an early draft and reports how the model changed: two channels several reviewers raised, a higher return to capital and diverging wages between exposed and unexposed occupations, "are now part of the scenarios." It adds that "External reviewers were not asked to endorse our conclusions." It then lists criticisms still open: the model does not follow individual workers, so it paints only a coarse picture of displacement costs; some reviewers questioned whether exposed occupations will shrink at all; some felt "the most extreme scenario is better read as a thought experiment than a scenario," while others felt the modest one understates what is already visible; several asked for clarity that data-center demand effects are excluded; and more than one argued that the model may underestimate how much AI could speed up technological progress.
The report's own caveats (§1) add:
- The model omits catastrophic risks, political economy, business cycles, financial disruption, price rigidities and the demand feedback they would create, and AI's value in household production.
- Workers differ only by group and employment status. A displaced worker who moves to the other group earns that group's wage immediately, with no loss for tenure or occupation-specific skill, although re-employed displaced workers have historically taken pay losses of ten to twenty percent, "concentrated on those who change occupation" (§3.4).
- Capital is a single good; compute is not distinguished.
- Only cognitive tasks are affected. The model does not allow for rapid advances in robotics, and "largely for this reason" it stops at 2030.
- The model has no Euler equation linking the return to capital to growth; footnote 13 says adding one "could raise the return to capital substantially."
On the page, the authors write that the explorer "should be viewed as a tool for thinking" and that "actual outcomes may differ materially."
Statements of intent
The page says the model "will inform the research Anthropic funds" through its Economic Futures program and "the policy ideas we propose," linking Anthropic's Economic Policy Framework (June 2026). The report's conclusion names retraining, income support, and universal basic capital or income among mechanisms that would become "a central question of economic policy" in a scenario near the extreme one (§5). These are statements of purpose, not results of the model.
Research questions
- Which scenario do measured parameters track as data arrive? The report says each parameter can be measured. Re-estimating observed exposure, firm diffusion and the automation share from later Economic Index and Census releases, and placing them against the three 2030 paths, would show which scenario the economy is following.
- How sensitive are the conclusions to the single wage rigidity value? The report has "no agreed value" for it and shows that it moves the split between wage cuts and unemployment. Evidence on how fast wages adjust in occupations with high AI use would narrow the range.
- Does the 0.17 search discount hold for workers displaced by AI? It is estimated from 2010 to 2019 switching. Re-employment data for workers leaving highly exposed occupations, compared with earlier displaced workers, would show whether the substantial and extreme settings (0.08 and 0.04) are plausible.
- How do the page's survey figures relate to Table 4? The page reports a 10% GDP gain, about 5% unemployment and 10% of respondents with extreme-scenario views; the report gives 8.6% and 4.6% and no share. A published method for the page's figures would settle whether they come from different calculations or a different sample.
- How much do the results change with robotics included? The model holds physical tasks fixed. Adding an exposure path for physical work, such as one built from a robot exposure measure, and rerunning the extreme scenario would show how much of the unemployment result depends on displaced workers having unaffected jobs to move to.
- What does an Euler equation do to the capital share? Footnote 13 anticipates a substantially higher return to capital. A version of the model with that link would show how much the labor share results move.
How it connects
- The model's starting point for AI's reach is the observed-exposure measure of Labor market impacts of AI: A new measure and early evidence (§3.1, §3.3). Its automation-share benchmark comes from the September 2025 Economic Index report, Anthropic Economic Index report: Uneven geographic and enterprise AI adoption (Appel et al., 2025, §3.4), and its comparison of field-trial gains with gains estimated from Claude conversations from Estimating AI productivity gains from Claude conversations (§3.3), whose estimate of 1.6 in logs is about four times the range the scenarios start from.
- The page links the Economic Index, Anthropic's Economic Policy Framework and the Institute's page on recursive self-improvement. An earlier policy post, Preparing for AI's economic impact: exploring policy responses (October 2025), sorted policy ideas by how fast and severe AI's effects might be; this note connects it, and the item does not cite it.
- How well do job retraining programs work? (August 2026), from the same economics group, reviews the evidence on one of the responses the report's conclusion names. Can we predict the jobs robots will do? (September 2026) measures robot exposure of physical tasks, the channel this model leaves out. Neither is cited by the item; both connections are the note's own.
- The same 10,980-adult sample is the public side of The AI expectations gap (October 2026), which compares it with a poll of 1,334 Anthropic staff.
- Other Anthropic surveys of public expectations include What 81,000 people told us about the economics of AI, which interviewed self-selected Claude users rather than a weighted panel.
References
- Korinek, A., Jones, C. I., Sacher, S., Cotter, T., and McCrory, P. (2026). Economic Scenarios for Transformative AI. The Anthropic Institute Working Paper No. 2026-02. https://www-cdn.anthropic.com/files/4zrzovbb/website/cf58f84d46a4a76bf5a5b039ac695fba6b80041c.pdf (57 pages; section, table and appendix numbers in this note follow it). record
- Anthropic (2026, September 9). What will our economic future look like? (Scenarios for our Economic Future, Econ Scenario Explorer v1.0). https://www.anthropic.com/institute/econ-scenarios. Fetched 13 November 2026. record
- Anthropic (2026, June). Anthropic's Economic Policy Framework. https://www-cdn.anthropic.com/files/4zrzovbb/website/9ea607a5dd67c168093829b701f3a0a6d21156d5.pdf. Read only to identify it. record
- Works cited above only as the report describes them (Acemoglu, 2025; Kokotajlo et al., 2025; Eloundou et al., 2024; Bonney et al., 2026) are listed in the report's references and were not read for this note.