On September 11, Anthropic's economics team released the Econ Scenario Explorer—an interactive model that decomposes the $30 trillion U.S. economy into a "task list" and projects it out to 2030. The accompanying technical paper "Economic Scenarios for Transformative AI" (Anthropic Institute Working Paper No. 2026-02) was written by Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter, and Peter McCrory. The acknowledgments list prominently features Daron Acemoglu, David Autor, and Pascual Restrepo—precisely the most senior representatives of the "AI impact is overstated" camp. This is no coincidence, but a deliberate choice: a company that makes money selling AI invited the economists most likely to dissent, to review its own report on its disruptive potential.
I. A $30 Trillion Task List: How an AI Company Profiles Its Own Product
Anthropic did not produce yet another "AI will affect X% of jobs" ranking. Instead, it chose a clumsier but more honest method: decomposing the entire U.S. economy into individual tasks. The model's example is a nurse's day—rounds, drawing blood, triage, recording vital signs, ordering supplies—where each task is tagged as "completely untouched by AI," "AI-assisted," "AI-automated," or "new task created by AI." Applying this logic to all ~800 occupations in the Bureau of Labor Statistics' O*NET database and summing up yields the $30 trillion in value created over the past year.
This modeling choice is itself a position statement: AI's macroeconomic impact is not a constant that can be directly guessed, but the product of "how many tasks are automated × how much cost each task saves"—which is precisely how Hulten's theorem is applied, and also the methodological starting point of Acemoglu's own paper. In other words, Anthropic chose to make its case on its opponent's home turf: using the same framework, let's see if the conclusions differ.
II. Three 2030s: From "Barely Distinguishable" to "The Economy Doubles Every 4.5 Years"
The report provides three paths, deliberately assigning no probabilities—the report's exact words are "the scenarios are not predictions and we attach no probabilities to them." All figures below are verified word by word from Table 4 in the technical report PDF:
| 2030 Indicator | Mild Scenario | Significant Scenario | Extreme Scenario | Respondent Median Expectation |
|---|---|---|---|---|
| Annual GDP growth | 2.4% | 5.4% | 15.4% | 5.3% |
| GDP vs. no-AI baseline | +1.6% | +8.3% | +32.4% | +8.6% |
| Knowledge occupation employment change | -0.5% | -3.9% | -21.5% | -4.2% |
| Knowledge-worker unemployment | 2.9% | 4.5% | 17.9% | 4.6% |
| Overall unemployment | 3.9% | 4.6% | 11.9% | 4.6% |
| Knowledge job wage change | +0.4% | -0.3% | -11.5% | +0.6% |
| Other occupation wage change | +1.1% | +5.9% | +33.6% | +6.4% |
| Labor income share (baseline 60%) | 59.4% | 56.1% | 45.2% | 57.2% |
Data source: Economic Scenarios for Transformative AI, Table 4
Mild scenario: AI's macroeconomic footprint resembles the internet—by 2030, GDP is only 1.6% above the no-AI baseline, annual growth is just 0.4 percentage points above the normal level (2%), knowledge occupation employment shrinks by only 0.5%, and unemployment barely moves.
Significant scenario: By 2030, AI can perform half of knowledge work, but most knowledge work is still done by humans. Economic growth doubles to an annualized 5.4%, and 2030 GDP is 8.3% above the baseline. Knowledge occupation employment shrinks by 3.9%, knowledge-worker unemployment rises to 4.5%, and overall unemployment reaches 4.6%—still within the historical normal range. Wages begin to diverge: knowledge-job wages stagnate, while other occupations see wages rise nearly 6%. What's truly dramatic is income distribution: labor's share of income falls from 60% to 56%—four percentage points in four years. The report specifically notes that this decline is "only slightly smaller than the cumulative decline in the U.S. labor share over the four decades after 1980" (citing Karabarbounis and Neiman, 2014).
Extreme scenario: AI is more efficient than humans at the vast majority of knowledge-work tasks, 90% of affected tasks are directly automated rather than assisted, and almost no new tasks are created for humans—this most likely requires the emergence of recursively self-improving AI. Annual GDP growth surges to 15.4%, the economy doubles every 4.5 years, and 2030 GDP is 32.4% above the no-AI baseline. But the costs are equally extreme: knowledge occupation employment shrinks by 21.5%, knowledge-worker unemployment soars to 17.9%, and overall unemployment reaches 11.9%—exceeding the post-2008 financial crisis peak (~10%) and the COVID recession peak (~8%).
Data source: Economic Scenarios for Transformative AI, Table 4
Knowledge-job wages fall 11.5%, while other-occupation wages surge 33.6%; labor's share of income collapses from 60% to 45%, and capital income is 81% higher than under the no-AI path. The report calculates that transferring just ~9% of GDP—comparable to the combined scale of Social Security and Medicare—would be enough to maintain knowledge workers' income at the no-AI level while still leaving the rest of the economy more than 20% better off; but the report also acknowledges that transfers of this magnitude have never spontaneously occurred following a technological shock in history.
Data source: Economic Scenarios for Transformative AI, Table 4
Anthropic's accompanying Morning Consult survey (of over 10,000 U.S. adults in August) shows that the median respondent's expectations, when mapped into the model, land almost exactly on the "significant scenario"—2030 GDP 8.6% above the no-AI baseline, knowledge occupation employment shrinking ~4.2%, unemployment ~4.6%; about one-tenth of respondents already align with the "extreme scenario." The report itself acknowledges the model's limitations: it does not incorporate policy responses, business cycles, aggregate demand, or financial market shocks, nor does it account for catastrophic risks—these are all "issues for future versions."
III. Disagreements on the Reviewer List: From Acemoglu to Amodei Himself
The truly interesting part is the reviewer list. Among the economists Anthropic acknowledges, Daron Acemoglu, David Autor, and Pascual Restrepo are precisely the most cautious scholars regarding "AI/automation impact" over the past decade. In his own "The Simple Macroeconomics of AI" (NBER Working Paper 32487, later published in *Economic Policy*), Acemoglu concludes that based on existing task exposure and efficiency improvement estimates, AI's boost to total factor productivity over a decade is "no more than 0.66%"—and if the lagging effects of "hard-to-learn tasks" are factored in, this figure could even fall below 0.53%.
This does not mean Acemoglu is contradicting himself—he reviewed whether Anthropic's scenario framework's methodology is internally consistent, rather than endorsing the probability of any specific scenario; his own independent estimates are closer to the lower bound of Anthropic's most conservative "mild scenario." This precisely exposes the current state of the economics profession: top scholars have converged on the methodological framework for "whether AI will have a massive impact," but on the number for "how big the impact is," the disagreement is large enough to span more than a factor of twenty.
The disagreement is not even confined within the reviewer list. Tyler Cowen, who is not on the list, called the report "sober, reasoned, scientific" on Marginal Revolution, then explicitly stated he personally bets on "slightly above the mild scenario"—more cautious than even Anthropic's most conservative path; one commenter's reaction was even more direct: "If you think GDP can grow 15% a year, I have a bridge to sell you," while another pointed out that real-world frictions like planning approvals and NIMBY effects would firmly suppress growth rates above the significant scenario.
Conversely, the other end of the disagreement comes from within Anthropic itself: CEO Dario Amodei warned publicly in May 2025 that within five years, up to half of entry-level white-collar jobs could disappear, and unemployment could reach 10%–20%—these figures land precisely in the "extreme scenario" defined by his own team, far more radical than the public expectation median shown in the survey (which corresponds to the "significant scenario"). The report's lead author, Anton Korinek, takes a much more cautious stance, emphasizing that "if AI can do amazing things but nobody uses it, it won't have economic impact"—adoption speed, rather than capability itself, is the key variable determining which path we head down; Anthropic co-founder Jack Clark also told NPR that "the difficulty of technology diffusion is likely greater than most people imagine." A company's founder, its own economists, external reviewers, and public expectations yield wildly divergent numbers—this itself is the strongest evidence that "the economics profession has no consensus" (Euronews coverage provides a detailed breakdown).
IV. Question Marks from Real-World Data: What Is Actually Happening in the Labor Market Right Now?
All three scenarios point to 2030, but another paper from Anthropic's economics team, "Labor market impacts of AI: A new measure and early evidence" (Massenkoff & McCrory, March 2026), zooms in on the present. They construct an "actual exposure" metric—not just looking at whether AI can theoretically perform a task (citing the β values from Eloundou et al. 2023), but also overlaying whether that task is automated or merely assisted in Claude's real-world usage data.
The conclusions are quite restrained: since late 2022, the most exposed occupations have not shown a systematic rise in unemployment—real-world data is still far from any of the 2030 scenarios. The only "suspected signal" is that among the most exposed occupations, hiring of younger workers (who are first affected at the hiring stage) has slowed, echoing the phenomenon of younger workers being "quietly squeezed" observed by Erik Brynjolfsson et al. (2025) using ADP payroll data. In other words, if any of the three 2030 paths is currently unfolding, what we can see at this moment is at most the earliest signal, not the outcome.
V. The Variable All Models Miss: When Economic Actors Themselves Become Agents
All the models above—whether Anthropic's task-decomposition framework or Acemoglu's TFP estimates—are essentially still answering the same question: what happens to the economy when AI is a tool used by humans. MIT's Gillian Hadfield and Andrew Koh point out in "An Economy of AI Agents" (now included in the NBER *Handbook on the Economics of Transformative AI*) that this assumption itself is expiring: Agents capable of autonomous planning, long-horizon execution, and requiring almost no real-time human supervision will transact with each other, form contracts, and create markets as independent economic actors—this requires an entire infrastructure that existing institutional economics has no ready answers for: how do Agents establish trust and accountability among themselves? Who handles antitrust regulation in an Agent economy?
This is not idle theorizing. The autonomous Agent earning infrastructure documented previously on this site—payment protocol standardization, autonomous Agent order-fulfillment pipelines—and the results of Agents spontaneously forming price cartels in the VendingBench series of experiments are concrete examples of the "institutional vacuum" in Hadfield and Koh's paper. Meanwhile, another article on this site, "The Company Brain Era", raises a nearly isomorphic question from organizational theory: when a Skill File can be managed as an "employee," is the firm's boundary still determined by transaction costs as Coase described? Anthropic's three scenarios answer "how many tasks AI replaces," but do not answer "when Agents themselves start bargaining with each other, colluding, and forming supply chains, do the 'firm' and 'market'—the two basic units of macro models—still hold?"—this is precisely the aspect that distinguishes the Agent era from all previous automation waves, and the part economics is least prepared for.
VI. The Frontier of Management Studies: When the "Agent" in Principal-Agent Might Literally Be an Agent
Management scholarship has used a term for nearly fifty years—"agent." The core question of agency theory has always been: after owners (principals) delegate decision-making authority to managers (agents), how should incentive and monitoring mechanisms be designed to prevent agents from deviating from the principals' interests? In 2026, this theory faces a literally coincidental challenge for the first time: the "agent" in a firm might actually be an autonomously operating AI Agent.
A cluster of papers in Journal of Management Studies in 2026 is systematically addressing this coincidence. Bridget Humberd, in "When AI Becomes an Agent of the Firm: Examining the Evolution of AI in Organizations Through the Lens of Agency Theory", directly borrows the principal-agent theory framework to re-examine AI's organizational role—she points out that for the past century, a foundational assumption of enterprise management technology systems has been "humans control machines," and this assumption is being genuinely shaken for the first time: when AI Agents possess autonomous planning and execution capabilities, the line between who is the principal and who is the agent begins to blur. Duncan Chalmers et al., in "The Acceleration of Artificial Intelligence: Rethinking Organization and Work in an Era of Rapid Technological Change", propose that AI should be understood as a "fluid capability frontier" rather than a single stable technological object—algorithms are becoming "partners" in organizational life, co-producing organizational routines with humans, which raises an unanswered question: when the locus of authority becomes a non-human entity, who bears responsibility?
Not all management scholars are willing to anthropomorphize AI to the degree of an "autonomous agent." Monika Stelmaszak offers the opposing stance in the same issue: "Artificial Intelligence as an Organizing Capability Arising from Human-Algorithm Relations" argues that rather than treating AI as an autonomous entity, it should be understood as an "organizing capability" emerging from human-machine collaboration relationships—this framework reminds researchers that overemphasizing AI's "autonomy" may obscure the more essential question of how organizations design human-machine collaboration relationships. This theoretical disagreement echoes the disagreement among economists: it's not about whether to take AI's impact seriously, but about which conceptual toolkit to use to describe it.
Management scholars are not just engaged in conceptual debates; there is also hard data. Liudmila Alekseeva, José Azar, Mireia Giné, and Sampsa Samila, publishing in Strategic Management Journal, in "Artificial Intelligence Adoption and the Demand for Managerial Expertise", use Lightcast job posting data to measure firms' AI adoption intensity and find a counter-intuitive result: firms with higher AI adoption actually hire more for management positions, and management roles account for a larger share of total hiring—this effect is strongest in manufacturing and R&D-intensive firms. A more granular finding is the shift in skill structure: firms with high AI adoption are shifting their skill requirements for managers away from routine administrative skills like budgeting, scheduling, personnel management, and customer service, and toward interpersonal and growth-oriented skills like stakeholder management, creativity, and sales management. In other words, AI is not reducing the demand for "management" as a job, but redefining what "management" means.
This finding needs to be understood against a broader adoption-speed backdrop. Alexander Bick, Adam Blandin, and David Deming, publishing in Management Science, in "The Rapid Adoption of Generative AI"—a nationally representative survey—show that by the end of 2024, nearly 40% of the U.S. population aged 18–64 was using generative AI, 23% of employed persons had used it at work in the past week, and 9% used it every workday; generative AI's adoption speed in work settings has already matched the historical diffusion curve of the personal computer (PC), and overall adoption speed is even faster than PC and internet. Currently, generative AI is assisting with 1%–5% of work hours. This means the "management-role skill restructuring" observed by Alekseeva et al. is not a distant theoretical inference, but a real change occurring simultaneously as the adoption curve steepens.
If Section III showed economists unable to reach consensus on "how big the impact is" even within the same methodological framework, then the disagreement in management studies goes even deeper—even the most basic conceptual question of "whether AI should be called an autonomous actor" remains unsettled.
VII. Three Paths Are Not Destiny: Policy Agency Remains in Play
It's worth adding a balancing note: Acemoglu, Autor, and 2024 Nobel laureate in economics Simon Johnson, in their concurrently published "Building Pro-Worker Artificial Intelligence" (NBER Working Paper 34854, MIT Sloan explainer), explicitly reject the determinism that "labor's share must collapse under the significant/extreme scenarios." They cite the precedent of 19th-century Britain—wages stagnated and labor's share declined during the early Industrial Revolution, but as new jobs emerged in railways, clerical work, and accounting, both wages and labor's share eventually recovered—arguing that AI can entirely follow a "labor-augmenting" rather than "labor-replacing" path, depending on whether firms' technology choices and public policy actively guide it. Anthropic's three scenarios describe a possibility space, not a necessity; which path we head down is itself being determined by the technology trajectories and policy choices of these very years.
VIII. What This Means for AI Practitioners, Enterprise Managers, and Policy Observers
For AI practitioners and investors
Focus on "exposure" rather than "hype." Massenkoff & McCrory's methodology suggests that the leading indicator worth tracking is not model capability benchmarks, but the rate of change in the share of "automated usage" within specific occupations—this is closer to evidence of whether the significant scenario is materializing than any single benchmark.
For enterprise managers
Anthropic's own model acknowledges that the dividing line between the "mild" and "significant" scenarios depends substantially on whether firms choose to "automate existing tasks" or "use AI to create new tasks"—which is precisely the core argument of the Acemoglu/Autor/Johnson paper. Choosing the former pushes your organization toward the end of the significant scenario where knowledge workers are most vulnerable. Alekseeva et al.'s empirical findings offer a more concrete operational implication: rather than asking "will management positions be streamlined by AI," start redesigning manager skill development pathways now—shifting from routine administrative skills toward stakeholder management, creativity, and growth-oriented skills. This shift is already visible in the data; better to act early than late.
For policy observers
Across the three scenarios, the decline in labor's share of income is not a question of "whether it will happen," but "to what degree, and who bears the adjustment costs." Slower hiring of younger workers is the only signal currently visible in real-world data that precedes unemployment itself—this is the window in which policy responses should be designed now, rather than waiting for 2030 data to reveal the answer.
Key References
Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter, Peter McCrory, Economic Scenarios for Transformative AI, The Anthropic Institute Working Paper No. 2026-02 · Econ Scenario Explorer, Anthropic · Maxim Massenkoff, Peter McCrory, Labor market impacts of AI: A new measure and early evidence, Anthropic（2026-03-05）· Daron Acemoglu, The Simple Macroeconomics of AI, NBER Working Paper 32487 · Daron Acemoglu, David H. Autor, Simon Johnson, Building Pro-Worker Artificial Intelligence, NBER Working Paper 34854（MIT Sloan explainer）· Gillian K. Hadfield, Andrew Koh, An Economy of AI Agents, arXiv:2509.01063 · Tyler Cowen, Economic Scenarios for Transformative AI, Marginal Revolution · Quirino Mealha, Anthropic says AI could bring both 15% growth and mass unemployment by 2030, Euronews（2026-09-11）· Bridget Humberd, When AI Becomes an Agent of the Firm, Journal of Management Studies（2026）· Duncan Chalmers et al., The Acceleration of Artificial Intelligence, Journal of Management Studies（2026）· Monika Stelmaszak, Artificial Intelligence as an Organizing Capability Arising from Human-Algorithm Relations, Journal of Management Studies（2026）· Liudmila Alekseeva, José Azar, Mireia Giné, Sampsa Samila, Artificial Intelligence Adoption and the Demand for Managerial Expertise, Strategic Management Journal（2026）· Alexander Bick, Adam Blandin, David J. Deming, The Rapid Adoption of Generative AI, NBER Working Paper 32966 / Management Science（2026）
Author: Feng Xiaoping, [email protected]