AI research review
Human-validation queue
Questions the report would most like commissioned evaluators to check, rather than read as AI criticism. They are ordered by how likely an answer is to change how decision-makers use the model.
CriticalIs ε = 3 a defensible capital-supply elasticity for 2026–30 under rapid AI investment?
Why it matters: It changes GDP, the labour share and the sign of the average-wage effect.
What Claude checked: Table 5 values checked in the paper. The paper sets ε at half the long-run value implied by Moll, Rachel and Restrepo (2022), leaning high because AI capital is “disproportionately compute, financed in a world market and built in a year or two” (Section 3.1, footnote 1).
CriticalIs ξ = 0.5 a plausible wage rigidity for a rapid, permanent occupational shock?
Why it matters: Extreme-scenario cognitive unemployment ranges from 2.6% to 24.0% across the paper’s settings.
What Claude checked: Table 6 values checked in the paper. The authors say evidence on job stayers “suggests ξ near 0.9” while job changers’ wages are flexible, and choose 0.5 (a one-year half-life) because the shock is “large, visible, and permanent” (Section 3.4).
CriticalDoes the model need an endogenous AI-R&D module before it is used for AI-safety or governance questions?
Why it matters: Smooth, exogenous paths may understate transition speed and shorten policy lead time.
What Claude checked: The paper confirms the AI paths are inputs, calls RSI “hard-coded” into the productivity path, and its innovation effects “likely a lower bound” (Sections 1 and 2.2.1).
HighWhy does the public reconstruction of the observed-exposure anchor differ from the paper’s?
Why it matters: m₂₀₂₆ = 0.14 anchors all three capability paths.
What Claude checked: Econ-ARK recomputes 0.121 overall against the paper’s 0.14, which the authors call conservative (Section 3.3); open.
HighIs the eight-task survey ladder a valid estimate of economy-wide affected task spending?
Why it matters: The paper’s own comparison (27% own hours against 71% from the ladder) suggests a construct problem.
What Claude checked: Both figures checked in Appendix B.3. The tasks were chosen to represent increasing difficulty, not to sample the wage bill.
HighAre the 27% own-hours and 71% task-ladder figures like-for-like, and which better measures economy-wide capability?
Why it matters: The questions differ in scope (own job against eight listed tasks) and standard (“about as well as they can” against “as well as a skilled professional”). The authors read the gap as a common survey pattern; the report reads it as a measurement problem.
What Claude checked: Question wording checked in Appendix B.
HighHow sensitive is the full monthly model to σ, beyond the static production block?
Why it matters: Task complementarity affects output, factor prices and reallocation; the effect on unemployment is untested.
What Claude checked: The paper’s robustness checks cover ε and ξ only (Sections 4.5–4.6) and note estimates of σ “range from below one to above one”; the report’s own σ calculations are unavailable.
HighIs ρ an adequate summary of new-task creation?
Why it matters: Instant, uniform reinstatement may simplify real adaptation a great deal.
What Claude checked: In the paper new tasks enter at the same unit cost, are staffed at once and offset displacement one-for-one in the wage bill (Section 2.1.2).
HighShould chips, data centres, electricity and finance be modelled separately from generic capital?
Why it matters: Capital adjustment is already one of the most influential parts of the model.
What Claude checked: The paper says its single capital good “does not distinguish compute from other capital” (Section 1).
HighHow much does omitting physical automation bias the “other occupations” adjustment margin?
Why it matters: Determines whether displaced cognitive workers have a durable destination.
What Claude checked: The paper excludes robotics and stops at 2030 “largely for this reason” (Section 1). Robot cost and capability figures checked on Anthropic’s robotics page.
HighAre competitive factor payments informative enough for distributional policy?
Why it matters: AI rents, ownership concentration and market power could dominate who gains.
What Claude checked: The paper’s footnote 14 case already shows compensation could take 84% of the GDP gain.
Medium-highDoes changing the timing of AI progress, with 2030 endpoints fixed, change unemployment and adjustment costs?
Why it matters: Informs governance and social-policy lead time.
What Claude checked: The paper notes almost all divergence between scenarios comes after 2027 (Section 5).
Medium-highHow much of the scenario differences in unemployment comes from posting speed θH and the search discount μ?
Why it matters: These labour inputs move with the capability inputs, and θH moves in the cushioning direction.
What Claude checked: The paper calls θH “an assumption” and says μ and θH “set the pace of absorption” (Section 3.4). Can be run in the explorer or the Econ-ARK code.
VerificationAudit Econ-ARK’s twelve extra implementation details and its remaining discrepancies with public data
Why it matters: Separates “the published model is reproduced” from “the text alone specifies it”.
What Claude checked: Counts and items checked in the Econ-ARK repository.
No questions match these filters.
Expertise filters match any selected area. Items marked Claude were added while editing .
How to respond
A useful response does more than agree or disagree. For each item a reviewer could note whether the critique is correct, how much it bears on decisions, how large its quantitative effect is likely to be, and what model change they would recommend. That would make this AI audit itself evaluable. Use Hypothes.is on this page, or the rating and discussion panel on the paper brief.
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