AI research review
Evaluator guide (Unjournal – Korinek et al., 2026)
Which conclusions about AI’s economic effects should readers rely on? We’d value your assessment of the model, its assumptions, and the evidence used to choose its inputs.
Please focus on the questions closest to your expertise and most important to the conclusions. Nobody is expected to cover everything; a few well-chosen checks may be enough. You’re welcome to question the paper’s emphasis as well as its answers.
What we need from you is a written report, a short summary, and ratings, all submitted through our evaluation form. An independent reproduction covers the arithmetic; we’re mainly asking about the economics. How the questions relate to the form.
The paper: Korinek, Jones, Sacher, Cotter and McCrory, Economic Scenarios for Transformative AI, Anthropic Institute Working Paper 2026-02 and CEPR Discussion Paper 21939, September 2026 (PDF · scenario explorer)
Why this paper? How does it connect to AI safety?
The paper offers a model and an interactive explorer for turning assumptions about AI progress into estimates of growth, wages, and unemployment. Anthropic says it plans to use the model to inform research funding and policy proposals.
People are likely to use it: the explorer invites journalists and policymakers to pick a scenario and quote its numbers, and the paper has already had press coverage and a presentation to a policy audience. Independent scrutiny matters more than usual here, since the authors are at, or work with, a frontier AI developer’s research institute. The arithmetic is already checked: an independent reproduction by Econ-ARK matches all 226 checkable numbers in the paper, so the open questions are about the assumptions and the economic structure.
There are possible connections to the risks from powerful, misaligned or out-of-control AI. Expectations of large economic gains could strengthen incentives to invest in capabilities and deploy systems quickly. Who receives those gains could affect who funds development and who has influence over it. The results also bear on power: how much economic and political weight AI developers gain, and whether large job losses make opposition to AI a major political force, which would shape the room for regulating it. A fast economic transition could also leave institutions less time to prepare. These are reasons to examine the economic assumptions carefully; they are not findings of this paper.
The model doesn’t estimate the risk of losing control of AI. An evaluation could help establish how much of the economic reasoning used in debates about AI development is supported, and which important questions its choice of model leaves unanswered.
An AI-drafted take on the paper’s main claims
Which claims matter most, which does the evidence support, and which mainly restate the scenario inputs? This AI-drafted list is a starting point; please amend it.
- The model and its explorer give a useful and reliable mapping from assumed paths for AI capability and adoption to US GDP, wages, workers’ share of income, and unemployment through 2030. This one is broad, and the paper frames it as a tool more than a claim.
- If AI advances very quickly (the ), US GDP in 2030 is about 32% above a world without AI, and workers’ share of income falls from 60% to about 45%. Growth reaches about 15% a year, cognitive employment falls about 21.5% from mid-2026, cognitive unemployment reaches about 18%, and cognitive wages are about 11.5% below the no-AI path while other wages are about 34% above it (Table 3).
- Even then, the gains could compensate the losers: a transfer of about 9% of GDP would hold cognitive workers’ income at its no-AI level (p. 35).
- In the , GDP is about 8.3% higher by 2030, cognitive wages are roughly unchanged, and workers’ share falls to about 56%. The typical survey respondent’s answers imply outcomes close to this scenario (Table 3; pp. 4 and 36).
- How readily new capital is supplied largely decides whether workers or the owners of capital gain. In the extreme scenario, moving from a low capital-supply elasticity (1) to an unlimited supply raises the 2030 GDP gain from about 21% to 43%, and moves average wages from about 9% below the no-AI path to 30% above it (Table 5).
- Whether disruption shows up as unemployment or as lower pay depends mainly on how rigid wages are and how easily workers change occupations. In the extreme scenario, cognitive unemployment ranges from about 2.6% with flexible relative wages to about 24% with very sticky ones (Table 6).
- AI-driven research adds relatively little to growth by 2030, even in the extreme scenario. The fast capability path is an input to the model, so this says little about how large feedback from AI research into AI itself could be.
List drafted by AI from the paper and two earlier AI reviews; the figures are the paper’s Tables 3, 5, and 6, which the Econ-ARK reproduction matches. “What it may inform” is our judgment, drafted with AI. Partly reviewed by our team.
1. Questions about the paper
We’re particularly interested in whether the main conclusions survive reasonable modeling alternatives. We can probably help run combinations of alternatives (a ). The model alternatives page offers starting points and one pilot calculation. Suggestions for which comparisons matter most would help.
Were the AI scenarios well chosen, and do the headlines depend on how extreme they are?
The authors supply three paths for AI capability and adoption, then calculate their economic consequences. Are the paths reasonable for 2030? Do the headlines depend on scenarios that are too extreme, or not extreme enough?
Which assumed productivity gains, adoption rates, and investment requirements have a sound basis?
The paper assigns no probabilities to the three scenarios. What can readers learn by setting them side by side, and how should the headline numbers be described: as forecasts, as bounds, or as illustrations?
Paper: Sections 1–3 and Tables 1–3. Explanation of the model.
Would other reasonable modeling choices change the main conclusions?
This is one of the questions we most want evaluators to take up.
Capital responds to returns through a supply rule, without explicit saving or investment decisions. Cognitive wages adjust gradually, which can leave workers unemployed when demand for their work falls.
Would construction delays, financing constraints, or a different wage rule change the conclusions through 2030? Since adoption is an input, is rapid adoption consistent with firms’ incentives and the investment it requires? The alternatives page also covers job search, new tasks, and substitution between human and AI work.
For a suggested comparison, please explain why it is plausible and which conclusion it could change.
Capital supply · Wage adjustment · Model alternatives and pilot results.
How well does the evidence support the inputs?
The paper chooses its inputs () using , labor-market evidence, and a public survey. Which choices are well supported? For example, do observed rates of moving between occupations tell us how easily workers displaced by AI could find different work?
Could hiring or wage changes in exposed occupations distinguish the scenarios as AI develops, allowing for other economic changes? A good fit to the calibration data may not carry over to a rapid transition.
Some survey data and calibration code are unavailable. Please assess what the published methods support; we can ask the authors for missing materials.
Available replication material · Survey questions below · Evidence for distinguishing scenarios.
How convincing are the results about who gains and who loses?
In the model, large output gains can accompany lower income for cognitive workers and a larger share going to capital. Workers are divided into just two occupation groups. How much does that simplification miss about differences in pay, skills, and the cost of changing careers?
A simple model can still give useful insights. Would a finer division of workers materially change these conclusions? Could market power and rents from models, scarce compute, or proprietary data change who gains?
The authors also calculate that gains could cover workers’ losses. How much is available after paying for extra capital, whose returns may go abroad? Does the discussion explain what compensation would cover? The optional accounting check below examines this more closely.
Paper: Section 4 and Table 3. Compensation discussion · Related AI criticism.
Does the timing of AI progress change the disruption?
Reaching a given level of AI capability in 2030 could involve steady progress or a late surge. Those paths could give investment and workers very different amounts of time to adjust.
Would paths with the same 2030 endpoint produce different peak unemployment, time out of work, or wage losses? We could help compare them using AI assistance on the public reproduction, as in the labeled pilot calculation.
Paper: Sections 2.3 and 3–4. Labor-market results.
What should we learn from the small effect of AI on innovation?
The paper reports that AI-assisted research contributes relatively little extra growth by 2030. Research funding follows output, and new ideas take time to affect production. The model also leaves out feedback from research into further AI capability; rapid capability growth is supplied as an assumption.
How much of the small innovation result follows from these choices? Does it receive the right emphasis, particularly for readers interested in AI accelerating its own development?
Paper: Section 2.2, especially 2.2.1 and footnote 6. Innovation discussion · AI issue 10.
What important questions does the model leave unanswered?
The model focuses on the US through 2030 and on automation of cognitive work. That makes its results harder to apply to countries with different industries, capital ownership, or opportunities for displaced workers. Advances in robotics could also affect the other occupations that absorb workers in these scenarios.
Which omissions most limit an important conclusion or leave a central question unanswered? Are the outcomes and time horizon well chosen? Feasible extensions that address these gaps would be useful.
Paper: Section 1, including the caveats. Model mechanisms · Validation discussion.
2. Optional checks of AI criticisms
If relevant to your assessment, the AI review offers specific checks .
Does the survey support the substantial scenario?
The paper says “the public’s expectations land close to our substantial change scenario” (p. 4). How well does the mapping from survey answers to model inputs support that statement?
How much work could AI do?
One survey measure asks when AI could perform eight listed tasks. Another asks employed people what share of their own working hours AI could handle. The paper reports quite different answers for cognitive workers: (Appendix B.3, p. 47).
The AI review finds that using the own-hours measure puts the modeled outcomes closer to the modest scenario. Are the two measures comparable, and which better supports a claim about what the public expects? The questionnaire and published tables provide a starting point.
The mapping uses respondents who answered all five input questions and fixes four other inputs for everyone. Could selection into that subset, or those fixed values, change the apparent agreement with the substantial scenario?
How long would finding new work take?
The survey describes someone who “needs to switch into a different occupation” and asks how long they would take to “find a new job” (p. 46). The answer in months is converted into a . The audit questions whether this conversion gives the waiting time the respondent described. A different occupation can also fall within the same broad group in the model.
Please assess the wording and Table B.1’s conversion. Fuller checks need respondent data, weights, and code from the authors; we can request them.
Appendix B and Tables 2 and 4. AI criticisms: issue 1, issue 2, issue 4.
Which jobs do new entrants to the labor force look for?
The model keeps the labor force fixed: new entrants offset people leaving, for example through retirement. The audit questions which occupation group these new workers try to enter. Its alternatives direct more of the entrants who replace departing cognitive workers toward the group where employment is growing.
Is the original assumption reasonable, and are the alternatives a fair test of how new entrants respond to changing opportunities?
Paper: Section 2.3.2, pp. 20–21. AI criticisms: issue 5 and related issue 6, issue 11, issue 12.
How much is available to compensate workers?
The paper compares the extra GDP with cognitive workers’ income losses. The AI review argues that allowing for the cost of the extra capital reduces the amount available. This depends on how that capital is financed and who owns it.
Does the AI use an accounting approach that fits the model? Does the paper’s claim that workers could be compensated still hold on that basis? It would help to separate a correction to the calculation from a qualification about what the calculation means for workers’ welfare.
Paper: pp. 35–39, footnote 14, and Table 3. AI criticisms: issue 7, issue 8, issue 17.
Are other AI criticisms worth pursuing?
The full digest covers innovation, comparisons with earlier work, validation, and smaller reporting points. Please use your own judgment about which deserve attention.
Innovation: issue 10 · Validation: issue 13 · Comparisons with earlier work: issue 14.
Writing your evaluation
Please submit through the evaluation form . The form asks for your report and a short summary, seven percentile ratings and two journal-tier ratings (each with a 90% interval), and a few short disclosures. Everything else on it is optional, though the claim identification and assessment section is particularly valuable for this paper (see below).
Please organize the report however you prefer, along the lines of a referee report. We can help with formatting.
How the questions on this page relate to the form’s ratings
- Methods. The questions about modeling choices, the evidence behind the inputs, and the timing of AI progress.
- Claims, strength and characterization of evidence. Whether the headline numbers are described in a way the model and its inputs support, for example whether they read as conditional scenarios, and how carefully the survey comparison is stated.
- Open, collaborative, replicable research. The authors’ public explorer code, which an independent group (Econ-ARK) reproduced, set against the survey data and calibration code that are not available.
- Relevance to global priorities. The decisions and open questions the model could inform, including the possible links to AI safety above.
- The claim section is optional on the form and was written mainly for empirical papers, but we’d particularly like you to use it here. Pick one or two of the paper’s conditional claims and say how far you believe them. A claim can take a few forms (), and they can come from your own list or the AI-drafted list above.
For disputed results, cite the paper passage and explain your reasoning or calculation. We’d particularly value your judgment about which conclusions hold up. The authors will be able to respond publicly.
About this guide
This guide was drafted by AI from the paper, the earlier working assessment, the isitcredible audit, and The Unjournal’s internal evaluation notes, and revised following The Unjournal’s comments. It is the one page we send to evaluators of this paper. The technical criticisms await human assessment.
Select a passage to annotate it, or open the Hypothes.is sidebar tab on the right edge. Public annotations can correct a number, challenge an inference or point to evidence.
How annotations may be used: Hypothes.is annotations on this page may be incorporated into The Unjournal’s knowledge base and may be quoted, summarized, or reposted on other Unjournal pages or public channels. Do not include private or confidential information.
Hypothes.is help