AI research review
Technical critiques
These are the report’s critiques that would change how the model should be used, tightened and with sources added. The diagram shows where the paper is strong and what it leaves to other work.
From AI research to welfare: what the paper covers
Select a step to see how the paper treats it.
The paper’s strength is step 5. Steps 2 to 4 are scenario inputs, and steps 1, 7 and 8 are outside its scope.
Coverage judgements are the report’s, refined by Claude . Quotations are from the paper, with section numbers .
All steps as text
| Step | Coverage | How the paper treats it |
|---|---|---|
| 1. AI R&D dynamics | Outside scope | Not modelled. AI capability paths are inputs; in the paper’s words, the gains from recursive self-improvement “can be thought of as being hard-coded” into the productivity path (Section 2.2.1). The authors call their innovation effects “likely a lower bound” and point to Davidson, Halperin, Houlden & Korinek (2026) for explosive research-driven acceleration. |
| 2. Technical task capability (m, a) | Assumed input | Taken as given. Each scenario fixes how much of the economy’s work AI can affect (m) and how much faster it makes each task (a). The mid-2026 starting point for m (0.14) comes from Anthropic’s observed-exposure data, which the authors call conservative; Econ-ARK’s recomputation from public data comes out lower. |
| 3. Diffusion and adoption (d) | Assumed input | Taken as given: a smooth S-shaped path from 10% in mid-2026 to 20%, 40% or 60% by 2030. The timing of the path is a scenario choice. |
| 4. Automation vs augmentation (ψ, ρ) | Partly modelled | The shares are scenario inputs, but the model works out their consequences carefully: automated tasks shift income to capital, augmented tasks raise worker productivity, and reinstated tasks return work to the cognitive group. |
| 5. Production and labour transitions | Modelled in detail | The paper’s strength. A task-based production model with gross complements, two occupation groups, frictional job search across groups, sticky relative wages and an upward-sloping capital supply, solved month by month to 2030. Econ-ARK reproduces it to machine precision. |
| 6. Ownership and distribution | Partly modelled | Factor shares, capital income and group wages are modelled. Who owns the capital, how concentrated AI rents are, and taxes and transfers are not. On institutions for sharing the gains, the authors write that the question “is beyond our framework” (Section 4.3). |
| 7. Political and social response | Outside scope | Outside scope. The paper’s caveats list “catastrophic risks, political economy considerations, business cycles, and possible financial market disruptions”, and the model has no demand feedback beyond slow real wages (Section 1). |
| 8. Welfare and risk | Outside scope | Outside scope. GDP and factor payments are not welfare measures. Workers differ only by group and employment status; a displaced worker who moves earns the new group’s wage at once, “with no discount for lost tenure or occupation-specific skills” (Section 1). |
1. Feedback from AI research is the largest missing mechanism for AI-safety use
For ordinary macroeconomic questions, taking AI progress as given is a defensible simplification. For AI safety and governance it is the most consequential one, because the speed of the transition and the warning time depend on whether automating AI research makes AI progress accelerate. The authors acknowledge this: the model “omits some important forms of feedback between research and automation”, “cannot generate” explosive research-driven acceleration, and its innovation effects are “likely a lower bound” (Section 1) .
Recent work models exactly this link. Davidson, Halperin, Houlden and Korinek (NBER Working Paper 35155, 2026) derive conditions under which automating AI research produces explosive growth . Cunningham et al. (September 2026) estimate that current feedback loops are not yet self-sustaining but appear to be strengthening ; The Unjournal has an earlier AI-assisted assessment of that paper. A policy paper led by GovAI, What If Automating AI R&D Triggers an Intelligence Explosion? (28 September 2026, with Korinek among its authors), argues that years of AI progress could be compressed into much shorter periods and calls for visibility into AI-R&D automation inside frontier companies, ways to steer and constrain an intelligence explosion, and preparation for society to adapt . The original report dated this “October 26, 2026”; Claude found no item with that date .
The highest-value extension for safety use is to couple this paper’s macro and labour model to a capability model: AI-R&D automation → research throughput → capability → task reach, productivity and autonomy → more AI-R&D automation. The paper is well placed to be the of such a system. Epoch AI’s GATE model is one integrated attempt, combining compute scaling, task automation and semi-endogenous growth with a feedback from growth to AI investment ; the paper cites it, with Davidson et al., as a benchmark for its extreme scenario. These models complement Korinek et al. rather than replace it.
2. Shock timing needs its own sensitivity analysis
Capability and diffusion follow smooth S-shaped paths. The same 2030 endpoint reached gradually from 2026 gives firms and workers far more time to invest, retrain and rematch than one reached mostly in 2029 after a capability jump. The model already runs monthly, so a natural robustness check is to hold the 2030 values of m, d, a and ψ fixed and vary when and how steeply they arrive .
The paper’s own results show why timing matters: “almost all of the divergence comes after 2027”, a year and a half after the mid-2026 anchor (Section 5) . For governance, the relevant quantity may be less “GDP in 2030” than the interval between reliable warning signals and large labour or capital disruption. The model does not estimate that interval.
3. Task substitution and reinstatement need more scrutiny
The baseline is 0.5, so tasks are gross complements. The paper takes it from Acemoglu and Restrepo (2022), tracing to Humlum (2019), and notes that “estimates range from below one to above one” (Section 2.1.1) . σ governs how gains in AI-affected tasks spread, how much labour is released and how factor prices react; at σ = 0.5, “half of every gain is spent on more of the same task” (Section 2.1.5). The paper’s robustness checks vary capital supply and wage rigidity (Sections 4.5 and 4.6), not σ. The report says it ran a static check on σ and found output, wages and capital moved materially while the labour share was fairly stable; it did not rerun the monthly labour model . Those calculations are not available to The Unjournal, so treat that finding as unverified.
Reinstatement () creates new tasks for the cognitive group in proportion to automated tasks. They enter at the same unit cost, are staffed at once, and offset displacement “one-for-one in the affected wage bill” (Section 2.1.2) . That is a parsimonious device, but new tasks in a fast transition may differ in skills, productivity, learning time, eligibility, location and creation lag. It matters because the scenarios lower ρ as AI becomes more transformative: 0.5, 0.25, then 0. The modest value matches Acemoglu and Restrepo’s estimate for 1987–2017; the extreme value “assumes that AI does not create any new tasks for humans to perform” (Section 3.4).
4. The “other occupations” sector may be too safe
Workers leaving cognitive occupations move into a broad residual group that AI does not touch directly. The paper does “not allow for rapid advances in robotics and their effects on physical tasks”, and “largely for this reason” stops at 2030 (Section 1) . A worker who moves earns the new group’s wage at once, “with no discount for lost tenure or occupation-specific skills”. Excluding robots can understate displacement if physical work is later automated; but physical automation may also stay slow because of reliability and cost.
Anthropic’s What work can robots do? (30 September 2026) shows the gap: robots can perform about 74% of US physical job tasks in at least some settings, tasks that make up 34% of working hours, but “robots are cost-competitive for just 0.3% of job tasks” . The better extension is a separately parameterised physical-automation block with its own capability, diffusion, cost and capital constraints.
5. Competition and ownership matter for welfare
The model is competitive with constant returns. Transformative AI could create large rents from model ownership, scarce compute, proprietary data, platforms or temporary monopoly, precisely where the paper finds large transfers toward capital. Who owns that capital, how concentrated the rents are, whether governments can tax them, how much is owned abroad and how politics responds can decide whether a high-growth scenario is broadly good or destabilising. The paper lists “catastrophic risks, political economy considerations, business cycles, and possible financial market disruptions” among omitted forces, and says the question of “what institutions could lead to shared prosperity” is “beyond our framework” (Sections 1 and 4.3) . GDP and competitive factor payments are therefore not enough for global-priorities welfare analysis.
6. Reproducibility is excellent, with a few documentation gaps
Alan Lujan’s Econ-ARK rebuilt the model in Python from the paper’s appendix and calibration . With twelve items the paper does not state, it matches the explorer in every series and month to within 3.5 × 10⁻¹⁴ across 25 cases and reproduces all 226 checkable published numbers (43 inputs, 183 outputs). Run on Table 1 exactly as printed, with the other eleven items in place, 163 of the 183 outputs reproduce; one more printed digit recovers all 183 (Econ-ARK summary). A second implementation agrees to 6 × 10⁻¹⁴, and deliberately planted errors are caught.
The open items Econ-ARK reports
- Quit-rate ordering. The text does not say whether monthly quit fractions are converted to rates before or after being split into the part that responds to job prospects. The explorer splits first; converting first moves paths by up to 4 × 10⁻⁴ and pushes 6 published numbers out of their rounding (the twelve items) .
- Observed-exposure anchor. Recomputing the mid-2026 anchor from Anthropic’s published exposure data gives 0.189, 0.010 and 0.121 for cognitive, other and all work, against the paper’s 0.22, 0.01 and 0.14. Econ-ARK suspects how occupation scores are aggregated onto CPS occupation lines, and leaves it open (public data recheck) . The paper calls its 0.14 “conservative, since assistive use counts half, thinly used tasks count zero, and the data are from late 2025” (Section 3.3) , while the public recomputation comes out lower. Because m₂₀₂₆ = 0.14 anchors all three paths, a person should resolve this.
- Quits elasticity. The paper estimates 0.53 and rounds to 0.55 (Section 3.2); Econ-ARK’s standard specification gives 0.54, and the calibration record holds 6/11 ≈ 0.545. Also open .
- Survey microdata. They are confidential, so the survey results cannot be reproduced .
- Authors’ code. The model code is available only as the explorer’s browser JavaScript, which carries no licence; Econ-ARK records its outputs rather than redistributing it .
- The explorer has changed since the check. Econ-ARK compared against the explorer’s code as downloaded on 10 September 2026 and pinned by checksum. On 27 October 2026 that file’s address returned “Not Found”, and the model code sits in a differently named file that still cites the paper’s 0.624 calibration . So the explorer has been redeployed; whether its outputs changed has not been tested.
These are mostly documentation fixes. The observed-exposure anchor is the one with substantive consequences.
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