AI research review
Economic Scenarios for Transformative AI
Korinek, Jones, Sacher, Cotter and McCrory map assumptions about AI capability, adoption and labour-market adjustment into paths for GDP, wages, the and unemployment to 2030, under three named scenarios: modest, substantial and extreme. Explore the results, test the assumptions, and identify which claims deserve closer evaluation.
What holds up, and what needs checking?
The published scenario outputs reproduce from the model code. The interpretation depends on how survey answers enter the model, where replacement workers search, and how capital costs enter compensation calculations. The isitcredible review examines these questions in detail: of 29 claims, it finds 12 hold, 15 hold in a narrower form, and 2 are not established.
How to read the scenarios
The paper is most useful as a transparent, conditional scenario translator. It separates quantities that discussions of “AI exposure” usually merge: how much work AI can do (), how widely it is used (), how much faster it makes a task (), whether it replaces or helps the worker (), how many new human tasks appear (), and how capital and labour markets adjust. Anthropic’s scenario explorer lets readers change those inputs themselves.
It does not estimate whether or when transformative AI arrives. AI progress is an input: the authors write that the gains from recursive self-improvement “can be thought of as being hard-coded” into the productivity path (Section 2.2.1) . “The scenarios are not predictions, and we attach no probabilities to them” (Section 1). The paper’s caveats leave out catastrophic risks, political economy, business cycles, financial disruption, demand feedback and advances in robotics (Section 1) .
The most defensible result. Under fast, automation-heavy AI diffusion, large aggregate gains can coexist with large losses for exposed workers and a shift of income to capital. Whether those losses appear as lower pay, fewer jobs or unemployment depends heavily on how fast wages adjust and how elastic capital supply is.
- Fragile headlines. The extreme scenario’s 17.9% cognitive unemployment ranges from 2.6% to 24.0% across the paper’s own wage-rigidity settings; its average-wage gain ranges from −9.2% to +30.1% across capital-supply settings (Tables 5 and 6) . See capital supply and wage rigidity.
- Well-checked arithmetic. An independent Econ-ARK matched the explorer’s monthly paths to within 3.5 × 10⁻¹⁴ in 25 configurations and reproduced all 226 checkable published numbers . That supports the calculations, not the assumptions.
- Weakest links. The mapping from survey answers to model inputs; a single capital supply; a two-group labour market whose “other occupations” sector cannot be automated; no feedback from AI research to AI progress; and no account of who owns the capital or captures AI rents.
The three scenarios in 2030
Choose an outcome. Most values are gaps from the same economy without AI (the ), not changes from today.
Source: the paper’s Table 3 ; the values also match Econ-ARK’s transcription and, for GDP, growth and the labour share, the explorer page. Reference lines are the paper’s no-AI column. The authors attach no probabilities to the scenarios.
Table view: all 2030 outcomes
| 2030 outcome | Modest | Substantial | Extreme |
|---|---|---|---|
| GDP vs no-AI path (%) | 1.6 | 8.3 | 32.4 |
| GDP growth in 2030 (% a year) | 2.4 | 5.4 | 15.4 |
| Labour share of income (%) | 59.4 | 56.1 | 45.2 |
| Average wage vs no-AI path (%) | 0.7 | 2.1 | 9.7 |
| Cognitive-occupation wage vs no-AI path (%) | 0.4 | −0.3 | −11.5 |
| Other-occupation wage vs no-AI path (%) | 1.1 | 5.9 | 33.6 |
| Cognitive employment, change since mid-2026 (%) | −0.5 | −3.9 | −21.5 |
| Unemployment rate, cognitive workers (%) | 2.9 | 4.5 | 17.9 |
| Unemployment rate, all workers (%) | 3.9 | 4.6 | 11.9 |
| Capital income vs no-AI path (%) | 3.1 | 18.9 | 81.4 |
| Cognitive wage bill vs no-AI path (%) | −0.3 | −4.6 | −31.0 |
| Capital stock vs no-AI path (%) | 2.3 | 13.8 | 56.3 |
| Measured TFP vs no-AI path (%) | 0.7 | 3.1 | 13.4 |
| Ideas stock vs no-AI path (%) | 0.07 | 0.20 | 0.61 |
The paper’s Table 3 . Cognitive employment is the change since mid-2026; unemployment, growth and shares are levels; the other rows are gaps from the no-AI path. Without AI the paper’s values are: growth 2.0% a year, labour share 60.0%, unemployment 2.9% (cognitive) and 3.8% (all workers).
What each scenario assumes
| Scenario input | Modest | Substantial | Extreme |
|---|---|---|---|
| · Affected task share by 2030 Anchor 0.14 in mid-2026 (observed exposure). 2030: low end, middle and near the high end of rated feasibility (Eloundou et al. 2024) | 0.20 | 0.30 | 0.50 |
| · Diffusion by 2030 Anchor 0.10 in mid-2026 (Census Business Trends and Outlook Survey). 2030: scenario assumptions | 0.20 | 0.40 | 0.60 |
| · Log productivity gain in 2030 Anchor values 0.30, 0.35, 0.45 from trials and frontier firms; multipliers eᵃ as stated in Sections 4.1–4.3 | 0.30 (×1.35) | 0.45 (×1.57) | 0.80 (×2.23) |
| · Automation share Held constant. 0.5 is roughly today's chat use; 0.9 a world where agentic use is the norm (Section 3.4) | 0.50 | 0.75 | 0.90 |
| · Reinstatement ratio Acemoglu and Restrepo (2019) estimate about 0.5 for 1987–2017 | 0.50 | 0.25 | 0 |
| · Cross-group search discount Normal times 0.17, from CPS occupational switching | 0.17 | 0.08 | 0.04 |
| · Vacancy-posting speed Assumption; faster in the more disruptive scenarios | 0.10 | 0.25 | 0.50 |
Common to all three: σ = 0.5, ε = 3, ξ = 0.5, base labour share 0.60, cognitive share of employment 0.624. Source: the paper’s Table 1 and Sections 3.3–3.4 ; also matches Econ-ARK’s encoding.
Moving from modest to extreme changes all seven inputs at once, so the gap between scenarios does not isolate the effect of more capable AI. One input moves the other way: vacancy posting () is faster in the more disruptive scenarios, which softens their labour-market frictions. The original report did not mention this.
Background, earlier reports, and AI ratings
This page retains explanatory material from the earlier deep-research report. The key questions include qualifications from the later audit. Methods and comparisons are under Sources & background.
Earlier explainer and supporting pages
The AI report’s ratings
The report proposed Unjournal-style for the paper. They are AI outputs and play no part in The Unjournal’s evaluation. Evaluators may prefer to form their own ratings before opening this fold.
Show the AI ratings (two AI passes, side by side)
| Criterion () | ChatGPT report, 27 Oct (90% interval) | Earlier AI assessment, 7 Oct |
|---|---|---|
| Overall assessment | 86 (72–94) | 84 (72–92) |
| Claims, strength and characterisation of evidence | 80 (65–91) | 80 (68–89) |
| Methods: justification, validity, robustness | 82 (68–92) | 87 (76–94) |
| Advancing knowledge and practice | 90 (76–97) | 89 (78–95) |
| Logic and communication | 94 (86–98) | 91 (82–97) |
| Open, collaborative, replicable science | 96 (88–99) | 96 (89–99) |
| Relevance to global priorities | 97 (90–99) | 91 (81–97) |
Intervals are each AI’s stated . The ChatGPT report left the journal-tier questions blank; the earlier assessment gave 4.3 (“should”) and 4.2 (“will”) out of 5. Reference group: applied and policy research aimed at a similar audience.
The two passes differ by up to six points on a criterion. Claude’s comments on these ratings are in the editorial notes.
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