All walkthroughs · Paper briefs

Economic Scenarios for Transformative AI

Anton Korinek, Charles I. Jones, Szymon Sacher, Tess Cotter and Peter McCrory
The Anthropic Institute Working Paper No. 2026-02, September 2026 (released 9 September 2026); also CEPR Discussion Paper 21939

Paper · Brief and rating · Dashboard entry

About 21 minutes · Download MP3

AI-generated walkthrough with synthesised voices: a host explains the paper and a discussant questions it. It covers claims, methods, key assumptions, citations and use so far, decision relevance, and what an evaluator should check. Every direct quotation was checked word for word against its source. Public sources only. Not an Unjournal evaluation or decision.

Transcript

Host
This is a walkthrough of Economic Scenarios for Transformative AI, by Anton Korinek, Charles Jones, Szymon Sacher, Tess Cotter and Peter McCrory. It is Working Paper twenty twenty-six dash zero two of the Anthropic Institute, released on September ninth, twenty twenty-six, and also circulated as CEPR Discussion Paper two one nine three nine. This is an AI-generated walkthrough with two voices, a host and a discussant, and it is not an Unjournal evaluation or decision. The paper builds a small macro model that turns assumed paths for AI capability and adoption into paths for US GDP, wages, the labor share and unemployment to twenty thirty. It runs three scenarios, modest, substantial and extreme change, and feeds a survey of about eleven thousand US adults through the same model.
Discussant
So it is a mapping from assumptions to outcomes. How do the authors themselves describe the status of the numbers?
Host
The authors write: "The scenarios are not predictions, and we attach no probabilities to them; their purpose is to make the consequences of different assumptions comparable." A public scenario explorer lets anyone change the parameters. Under modest change, GDP in twenty thirty is about one point six percent above the path without AI, and unemployment rises by a tenth of a point. Under substantial change, GDP is about eight percent higher, the labor share falls from sixty to fifty-six percent, which the authors compare to the whole decline in the US labor share over the four decades after nineteen eighty, and cognitive unemployment rises from about three to four and a half percent. Under extreme change, GDP is thirty-two percent above the no-AI path, GDP growth reaches about fifteen percent a year, the labor share falls from sixty to about forty-five percent, and nearly one in five workers who started in cognitive occupations is unemployed.
Discussant
Before the mechanics, what counts as a cognitive occupation, and why split the economy in two?
Host
Cognitive means the twelve standard occupation major groups covering management, professional, sales and office work, which employ about sixty-two percent of workers in the twenty twenty-five Current Population Survey. Everything else, construction, food service, transport and so on, is the other group, and it is not exposed to AI. The paper leaves out robotics, which is also its stated reason for stopping at twenty thirty.
Discussant
Walk through the production side.
Host
It is a task-based model in the tradition of Zeira, Acemoglu and Autor, and Acemoglu and Restrepo. Output is a constant elasticity of substitution aggregate over task instances. An instance is one performance of a task, one contract reviewed, one patient history taken. The authors write: "AI arrives instance by instance: some contracts are reviewed with AI while others are not." The elasticity of substitution across instances is set to one half, so tasks are gross complements. Within an instance, labor and capital are perfect substitutes, and cost minimization assigns each instance to the cheaper factor.
Discussant
And how is a scenario defined in that structure?
Host
By five objects. First, the affected mass: the fraction of the economy's tasks, weighted by their base-period wage bill, that AI can do. Second, diffusion: the fraction of instances of those tasks actually done with AI. Third, the gain: the log cost reduction per AI-performed instance. Fourth, the automation share: of AI-performed instances, the fraction done by AI outright, which moves the instance to capital, rather than augmenting a worker. Fifth, the reinstatement ratio: new labor tasks created per automated task. Two labor-market frictions are also scenario-specific: a discount on searching outside one's own occupation group, and the speed at which expanding occupations post vacancies.
Discussant
What carries the aggregate result?
Host
Hulten's theorem. To first order, the rise in measured total factor productivity equals the labor share times affected mass times diffusion times the gain. In the substantial scenario in twenty thirty that is zero point six times thirty percent times forty percent times zero point four five, roughly three percent. That product does not depend on how much is automation versus augmentation. The split between labor and capital comes from the factor price frontier. If the rental rate on capital does not move, the wage rises by the whole cost saving. If capital is slow to adjust and the rental rate rises, each percent rise in the rental rate takes away two thirds of a percent from the wage, two thirds being the ratio of the capital share to the labor share.
Discussant
So the supply of capital is doing a lot of work. How is it modeled?
Host
As a reduced-form upward-sloping schedule: relative to the no-AI path, the capital stock rises by epsilon percent for each percent rise in the rental rate. Moll, Rachel and Restrepo imply about six in the long run; the authors take three, arguing that the relevant capital is largely compute, financed on world markets and built in a year or two. There is no explicit saving decision and no Euler equation linking the return to capital to growth.
Discussant
Now the labor share. Why does it fall so much in the extreme case?
Host
Three channels. The first and largest is reassignment. The authors write: "The entire wage bill of an automated instance moves to capital, however small the cost saving that motivated the switch." Net of reinstatement, the labor share falls by one minus the reinstatement ratio, times the automation share, times the fraction of instances done with AI. The second channel pushes the other way: because tasks are complements, the remaining labor tasks become the weak link and their expenditure share rises. The third is that a higher rental rate shifts spending toward capital's tasks. In the extreme scenario the automation share is ninety percent and reinstatement is zero, so the first channel dominates.
Discussant
And can the average wage fall?
Host
Yes, when the cost saving per automated task is small relative to the wage bill transferred and capital is scarce. That is so-so automation. At the baseline elasticity of three, wages rise in every scenario. In the robustness section, at an elasticity of one, the average wage falls below the no-AI path in both the substantial and the extreme scenarios.
Discussant
What about innovation? A lot of the transformative AI literature is about AI speeding up research.
Host
The ideas block is semi-endogenous, in the tradition of Jones, calibrated to Bloom, Jones, Van Reenen and Webb. AI raises research input only because it raises GDP. The authors write: "One surprising finding of our model is that the growth speedup from AI in innovation is relatively minor, even in the extreme change scenario." In the extreme scenario the ideas stock is only about zero point six percent above its no-AI path in twenty thirty. They call this likely a lower bound: gains from recursive self-improvement, meaning AI improving AI research, are treated as already built into the assumed path of the gain.
Discussant
So the extreme growth rate comes almost entirely from the assumed task paths. Now the unemployment block.
Host
It is a search and matching model at a monthly frequency. Each month, employers in the other group post vacancies to close a fraction of their shortfall relative to target employment. In the cognitive group, the wage is sticky relative to the economy-wide wage, following Blanchard and Galí: the gap to the market-clearing wage closes by half each year. Firms lay off the workers they do not want at that wage. Unemployed workers search in both groups, but search outside one's group of origin is discounted. Hiring uses the den Haan, Ramey and Watson matching function.
Discussant
How is the search discount calibrated? That seems like the hinge for the unemployment numbers.
Host
From Current Population Survey flows for twenty ten to twenty nineteen. About one in seven job-finders changes group, and multiplying the two directional odds lets them back out a discount of zero point one seven in normal times. The modest scenario keeps that. The substantial and extreme scenarios cut it to zero point zero eight and zero point zero four. The argument is that managers and professionals cross less often than sales and office workers, and that crossing gets harder when many search at once. The authors write: "Neither can be estimated precisely, given that AI has not yet displaced a significant share of workers."
Discussant
And the AI parameters themselves. Where do the starting values come from?
Host
All scenarios share a mid twenty twenty-six anchor. Affected mass starts at fourteen percent, from the observed-exposure measure of Massenkoff and McCrory, built from work-related use of Claude, Anthropic's own assistant, which the paper treats as a proxy for AI use generally. Diffusion starts at ten percent, from the Census Bureau's Business Trends and Outlook Survey. The gain starts between zero point three and zero point four five log points. The twenty thirty values are assumptions: affected mass of twenty, thirty or fifty percent, diffusion of twenty, forty or sixty percent, and a gain of zero point three, zero point four five or zero point eight. The automation share is fifty, seventy-five or ninety percent, motivated by the Anthropic Economic Index finding that about half of consumer Claude use and three quarters of programmatic use looks like delegation. Reinstatement is one half, one quarter, or zero.
Discussant
Give me the extreme scenario in full, since that is what gets quoted.
Host
GDP thirty-two percent above the no-AI path and growing at about fifteen percent a year. The average wage is nearly ten percent above its no-AI path, but the cognitive wage is eleven and a half percent below it, while wages in other occupations are about thirty-four percent above it. The labor share falls to about forty-five percent. Cognitive employment is about twenty-one percent below its mid twenty twenty-six level, cognitive-origin unemployment is about eighteen percent, and economy-wide unemployment is about twelve percent. Total labor income ends up almost exactly where it would have been without AI, so all of the GDP gain goes to capital income, which is about eighty-one percent higher. The paper computes that a transfer of about nine percent of GDP would hold cognitive workers at their no-AI income. It then adds: "Transfers of that scale in response to technological change have no precedent", citing the China shock as evidence that compensation does not happen on its own.
Discussant
Now the survey. How do you turn a lay respondent into five model parameters?
Host
Morning Consult fielded it to ten thousand nine hundred and eighty US adults in August twenty twenty-six. Capability comes from asking when AI will do each of eight tasks as well as a skilled professional, from writing routine business emails to making a Nobel-level discovery. The count of tasks expected by twenty thirty is placed on a linear scale of knowledge work, capped at ninety-five percent, and multiplied by the cognitive wage-bill share. Adoption, automation, speed-up and re-employment time come from direct questions. The median answers put capability at forty-four percent, adoption at forty percent, automation near fifty percent, and re-employment at about eight months. Running complete respondents through the model gives a median GDP gain of about eight and a half percent and unemployment around four point six percent, close to the substantial scenario.
Discussant
That mapping has a lot of discretion in it.
Host
It does, and the appendix shows some of the tension. Employed cognitive workers say a median twenty-seven percent of their own working hours will be doable by AI by twenty thirty, against seventy-one percent implied by the task grid. And about thirty percent of respondents expect AI to save no time at all on a task suited to it, while forty-nine percent expect it to at least halve the time.
Discussant
What do the robustness checks show?
Host
Two exercises. On capital supply, the extreme scenario at an elasticity of one gives GDP twenty-one percent higher, an average wage about nine percent below its no-AI path, and a net return above ten percent. A footnote constructs a so-so case, combining substantial capability with extreme disruptiveness and inelastic capital: GDP rises about seven percent but labor income falls, and compensating cognitive workers would take most of the gain. On wage rigidity, the authors write: "The effects of AI are felt in the cognitive labor market in either wages or unemployment; our baseline scenario shares the cost across both margins." With fully flexible wages, the extreme scenario's cognitive wage falls by about forty-two percent and cognitive unemployment is under three percent. With very rigid wages, cognitive unemployment reaches twenty-four percent.
Discussant
Let's turn to how it has been used and cited.
Host
It is about a month old. Tyler Cowen posted the abstract on Marginal Revolution the day after release and described the authors as sober and reasoned, adding that his own expectation is a little above the modest case. Alex Tabarrok followed with a post on redistribution. He argued the paper underplays adjustment margins: retirements and new entrants can shift labor between sectors without unemployment, and governments have mobilized large transfers before. Axios and Quartz reportedly covered it. The most substantive follow-up is an independent reproduction by Alan Lujan, packaged as an Econ-ARK REMARK and deposited on Zenodo. It reimplements the model and reproduces all two hundred and twenty-six checkable published numbers. It also reports twelve implementation details the paper does not state. The reproduction is explicit that it checks the arithmetic, not the assumptions. OpenAlex does not yet list the working paper itself with citing works, so academic citation is effectively unmeasured for now.
Discussant
Who would use this for decisions?
Host
Three groups. Funders and researchers working on redistribution, universal basic income or universal basic capital will likely use it as a quantitative anchor. Regulators and statistical agencies could use the framework to decide what to measure, since each parameter maps to something observable: exposure, firm adoption, per-task gains, re-employment times. And AI labs and their critics will use it in public arguments about labor impacts. The authors write: "The lasting contribution, we hope, is the framework itself: a common way to state disagreements about the economic effects of AI as disagreements about a few measurable parameters." For an Unjournal evaluation, the arithmetic has already been checked, so evaluators can focus on assumptions, calibration and interpretation.
Discussant
Then let me give the critiques. First, authorship. This is a lab-produced model of a lab's own technology, and two key anchors, the exposure measure and the automation share, come from Claude usage data. An evaluator should ask how sensitive the scenarios are to those anchors and whether independent measures would give different starting points.
Host
The paper does say the exposure measure lines up with those other measures, and that it is conservative because assistive use counts half. But the automation share is a sharper issue. Delegation in a chat or an API call is not the same as an instance of work moving permanently from a worker to capital, and the labor-share result moves almost one for one with that parameter.
Discussant
Second, the reassignment rule. Moving the whole wage bill of an automated instance to capital is a convention from the task literature. Combined with an automation share of ninety percent and zero reinstatement, it largely determines the fifteen-point fall in the labor share. An evaluator should ask whether zero reinstatement over four years is plausible, and whether the extreme scenario is better read as a bound than a scenario.
Host
Third is the capital block. The extreme case has the capital stock about fifty-six percent above its no-AI path by twenty thirty, with the return to capital rising less than two points. Without a saving decision or an Euler equation, it is hard to check whether the investment flows are feasible or how they are financed. The authors acknowledge that an Euler equation could raise the return substantially, which would also feed back into wages through the factor price frontier.
Discussant
Fourth, the unemployment block routes all reallocation through the unemployment pool. The paper itself cites evidence that historically much adjustment happens through entry and exit from the labor force. That is Tabarrok's point, and it suggests the unemployment numbers may be on the high side for a given displacement path. Against that, the model has no aggregate demand channel, no price rigidities beyond the sticky cognitive wage, and no earnings losses for displaced workers, who earn the new group's wage immediately. Those omissions push the other way. An evaluator should ask which side dominates.
Host
Fifth, the search discount values for the substantial and extreme scenarios are assumptions with a heuristic argument, and the posting speeds are pure assumptions. Since unemployment is what most readers take away, a sensitivity table over those two frictions would help.
Discussant
Sixth, the survey. The linear mapping from a count of eight heterogeneous tasks to a share of the wage bill, the ninety-five percent cap, and coding the median item by item mean the public's median scenario is a composite no single respondent gave. The gap between twenty-seven and seventy-one percent on own-work exposure suggests the grid may overstate expected capability. And the outcome distribution rests on the three thousand or so respondents who were asked about twenty thirty and answered all five items, so who drops out on each item is worth checking.
Host
Finally, interpretation. The paper gives no probabilities, but the extreme case's numbers are what circulates in coverage. An evaluator could check whether the explorer and press materials keep that caveat visible.
Discussant
So the summary for an evaluator: the model is transparent and its arithmetic has been independently reproduced. The open questions are the automation share and reinstatement assumptions, the capital supply elasticity, the routing of reallocation through unemployment, and the survey mapping, plus the conflict-of-interest question that comes with lab authorship.
Host
This has been an AI-generated walkthrough of Economic Scenarios for Transformative AI, and not an Unjournal evaluation or decision.

Sources