AI research review

AI-generated research notes, not an Unjournal evaluation. No person has checked this content or its numbers. Claude is made by Anthropic, which publishes the paper (conflict note).

Sources and review history

These explanatory pages were prepared from a ChatGPT deep-research report, edited and source-checked by Claude on 27 October 2026. The isitcredible review and key questions and evidence provide later checks and qualifications. The paper was selected for evaluation on 9 October; commissioned human evaluations will follow. Sources and method.

Sources and method

How these pages were made, what was checked and how, and what changed from the original AI report.

Three sources, one review workspace

What went into it, and how much work was involved?

The isitcredible review suite v0.9.2 (Joe Francis; The Catalogue of Errors Ltd, MIT licence) ran through paper review, code-and-data checks, a claim ledger, challenges, and a verdict, followed by audit-review and review-digest. This was a staged review with fresh-context challenges and audit checks, rather than a single report-generation prompt.

The run covered 28–29 October. The retained job history records roughly five further hours from the completed initial verdict to the finalized audit and digest on 29 October (09:12–14:20 UTC). That is elapsed workflow time, including source-access waits; it does not measure active model reasoning time, tokens, or financial cost. More processing is not evidence that a conclusion is correct.

The earlier AI reports were available as leads, so the run was not blind. Findings rest on the paper, its deposit, and cited works. Claude is made by Anthropic, which publishes the paper. No person has checked these findings or their numbers. Full provenance and conflict disclosure.

The evidence guide is a new AI-generated crosswalk, not a fresh audit. It links the preserved working assessment (4–7 October), the ChatGPT deep-research explainer (27 October), and the isitcredible review and audit (28–29 October). The later audit used the earlier reports as leads; their agreement is not independent confirmation.

The finalized digest retains the suite’s wording, numbers, severities, and Markdown download. Its 18 issues are navigation targets, not Unjournal ratings. Review records: synthesis and claim ledger, issue ledger, and source receipts and five open synthesis errata. The issue ledger and digest already carry the corrected wording; the synthesis remains preserved as received.

The source links inside those records sometimes point to job-relative files that are not redistributed. Survey microdata, calibration construction, and some cited published versions remain unavailable to this review. Read a reproduced model output, an unchecked empirical input, and an AI interpretation as different kinds of evidence. No human verification is claimed.

How these pages were made

The report. ChatGPT deep research produced a report titled “Technical Evaluation of Economic Scenarios for Transformative AI and the Anthropic Scenario Explorer”, supplied to The Unjournal on 27 October 2026. Its citation markers (such as citeturn21view0) and its links to sandbox:/mnt/data are artefacts of that tool and resolve to nothing. The verbatim text is kept in The Unjournal’s repository at docs/ai-governance/deep-research/economic-scenarios-chatgpt-deep-research-2026-10-27.md for provenance.
Checking. Claude (Anthropic’s Claude Opus 5.5) checked each factual statement on 27 October 2026. It checked every number and quotation about the paper against the paper’s own text (Anthropic Institute Working Paper 2026-02, September 2026, 57 pages including appendices), citing tables and sections. It also read Anthropic’s research pages directly, read the Econ-ARK reproduction’s repository (commit 87170fa, 25 September 2026), whose code independently transcribes the paper’s tables, and used web-search results for pages it could not open, such as GovAI, NBER, arXiv and The Unjournal’s guidance pages.
Editing. Claude shortened the text, corrected or labelled statements, added context where a check turned something up, and wrote the editorial notes.
Illustrations. The charts plot published numbers only. The one calculation, the accounting illustration, multiplies the model’s definitions and is not the paper’s model.

The earlier AI-assisted assessment on this site (4–7 October 2026) is a separate document; where these pages use it, they say so.

What the tags mean

Checked by Claude against the paper’s text; the table or section is cited nearby.
Read by Claude on that page or in that repository.
Confirmed through web-search results; the page itself could not be opened.
Taken from the AI report or the earlier assessment and not checked.
Claude’s arithmetic or additions.
Corrected, or no source found.

A check by Claude is not a check by a person. No number on these pages has been checked by a person.

Verification log

Status

Statement in the report	Status	Source	How it was checked
Paper identity: title, five authors, Anthropic Institute Working Paper 2026-02, September 2026	Verified	Economic Scenarios for Transformative AI	Title page of the paper. CEPR Discussion Paper 21939 is from the collection’s brief; the CEPR page could not be opened.
The scenarios are not forecasts and carry no probabilities	Verified	Economic Scenarios for Transformative AI	Section 1: “The scenarios are not predictions, and we attach no probabilities to them”.
Table 3 outcomes (GDP +1.6/+8.3/+32.4%, growth 15.4%, labour share 45.2%, cognitive unemployment 17.9%, and the rest)	Verified	Economic Scenarios for Transformative AI	Table 3; also matches Econ-ARK’s transcription and, for GDP, growth and labour share, the explorer page.
Capital-supply sensitivity: GDP 21.3→43.3%, average wage −9.2→+30.1%, labour share 41.2→49.1%, capital +33.5→+82.2%	Verified	Economic Scenarios for Transformative AI	Table 5 (Section 4.5).
Wage-rigidity sensitivity: cognitive unemployment 2.6→24.0%, cognitive wage −42.2→+2.8%, cognitive employment −1.3→−28.5%	Verified	Economic Scenarios for Transformative AI	Table 6 (Section 4.6).
Scenario inputs and common parameters (σ = 0.5, ε = 3, ξ = 0.5)	Verified	Economic Scenarios for Transformative AI	Table 1; also matches Econ-ARK’s encoding.
First-order formulas: TFP ≈ s_L·m·d·a; wage and rental rate depend on ε	Verified	Economic Scenarios for Transformative AI	Equations 9–11 (Section 2.1); worked example in Section 2.1.3.
Small ideas increment (+0.61% in the extreme scenario)	Verified	Economic Scenarios for Transformative AI	Table 3 and Section 2.2.1.
No feedback from ideas to AI progress; recursive self-improvement “hard-coded” into the AI path	Verified	Economic Scenarios for Transformative AI	Section 2.2.1 and footnote 6.
Omitted channels: catastrophic risk, political economy, financial disruption, aggregate demand, physical automation	Verified	Economic Scenarios for Transformative AI	Section 1, Caveats; also on the explorer page.
Single capital good	Verified	Economic Scenarios for Transformative AI	Section 1: “a single capital good … which does not distinguish compute from other capital”.
Survey of 10,980 US adults	Verified	Economic Scenarios for Transformative AI	Section 3.5 and Appendix B.1 (Morning Consult, 11–23 August 2026).
3,259 complete responders; four structural parameters fixed at substantial values	Verified	Economic Scenarios for Transformative AI	Table 4 note and Appendix B (ρ, θ_H, ξ, ε).
Eight tasks chosen as a progression of difficulty	Verified	Economic Scenarios for Transformative AI	Appendix B.3.
Cognitive workers’ own-hours median of 27% against 71% from the task grid	Verified	Economic Scenarios for Transformative AI	Appendix B.3, where the authors read it as a common survey pattern; the report omits that reading.
“The paper’s own sensitivity analysis” covers capital supply and wage rigidity but not σ	Verified	Economic Scenarios for Transformative AI	Sections 4.5 and 4.6 are the only robustness sections.
Econ-ARK: 226 numbers; 25 cases to 3.5 × 10⁻¹⁴; 163 of 183 with printed Table 1; twelve extra items	Verified	Econ-ARK article: summary of results	Read in the Econ-ARK repository, commit 87170fa. The 163 of 183 holds with the other eleven items in place.
Econ-ARK: unresolved observed-exposure anchor discrepancy	Verified	Econ-ARK article: public data behind the calibration	Recomputed 0.1888/0.0097/0.1212 against the paper’s 0.22/0.01/0.14 (Section 3.3, which calls the figure conservative).
Econ-ARK: survey microdata unavailable	Verified	Econ-ARK article: summary of results	“its microdata are confidential”.
Econ-ARK matches “the Scenario Explorer” (the report does not say which version)	Partly verified	Anthropic Econ Scenario Explorer	The explorer file Econ-ARK pinned on 10 September returned “Not Found” on 27 October; the code now sits in a renamed file. Outputs not retested.
Anthropic’s March 2026 observed-exposure labour-market research	Verified	Massenkoff & McCrory, Labor market impacts of AI: a new measure and early evidence	Page read: no systematic rise in unemployment for highly exposed workers; suggestive slower hiring of younger workers.
Anthropic Economic Index on automation against augmentation	Verified	Appel et al., Anthropic Economic Index report: Uneven geographic and enterprise AI adoption	September 2025 report (the paper’s source for ψ) and March 2026 report read: automation and augmentation shares differ between Claude.ai and the API and across tasks.
Anthropic’s September 2026 robotics work	Verified	Legate-Yang & Massenkoff, What work can robots do?	Page read: about 74% of physical tasks in some settings; 34% of hours; cost-competitive for 0.3% of tasks.
October 2026 follow-up comparing Anthropic staff, the US public and explorer visitors	Verified	Sacher, Cotter, Jones, Korinek & McCrory, The AI expectations gap	Page read. Visitors appear only in the economic-scenario part (Figure 7); median implied 2030 growth 5.4%, 12.5% and 17.1%.
“October 26, 2026 assessment involving Korinek” on compressing years of AI R&D into months	Corrected	Chan et al., What If Automating AI R&D Triggers an Intelligence Explosion?	No item with that date found. The matching paper is the GovAI-led policy paper dated 28 September 2026, with Korinek among its authors (search results; page not opened).
GATE (Epoch AI)	Partly verified	Epoch AI, GATE: an integrated assessment model for AI automation	Confirmed by search and cited in the paper’s references (Erdil et al., 2025): an integrated assessment model combining scaling and semi-endogenous growth, with feedback from growth to AI investment. The report’s mention of adjustment costs was not confirmed and has been dropped.
“Recent work modelling AI-R&D feedback”	Verified	Davidson, Halperin, Houlden & Korinek, When Does Automating AI Research Produce Explosive Growth?	Davidson, Halperin, Houlden and Korinek (2026), which the paper itself cites (Section 1); also Cunningham et al. (arXiv 2609.15802), found by search.
Unjournal applied-stream guidance: reference group of applied and policy work with a similar audience; less weight on originality	Partly verified	The Unjournal: Applied and Policy track	Confirmed by search results from The Unjournal’s guidance pages; pages not opened.
Unjournal applied stream makes journal-tier ratings optional	Not located	The Unjournal: guidelines for evaluators	Not located. The guidelines describe journal-tier ratings for the applied stream without calling them optional.
Unjournal “April 2026” AI policy: AI evaluations as a separate input; humans should not delegate ratings to AI	Partly verified	The Unjournal: AI/LLM policy for evaluators	The policy page exists and, per search results, encourages evaluators not to use AI for overall evaluations or ratings, requires independent verification, and plans to supply LLM evaluations after evaluators’ initial assessments. The April 2026 date was not located.
The report’s own static σ check, sandbox site, package and skill file	Not located	—	Not available to The Unjournal. Removed or labelled unverified.

Corrections and refinements

The AI-R&D policy paper the report dated “October 26, 2026” is dated 28 September 2026.
“Roughly a dozen” implementation details is exactly twelve; “163 of 183” holds with the other eleven in place.
The observed-exposure discrepancy is now quantified (0.121 against 0.14 overall), next to the authors’ view that 0.14 is conservative.
The report’s critiques of ε, the survey ladder and the compensation result are now set beside the authors’ own reasoning (Sections 3.1, 3.4, Appendix B.3, footnote 14).
Econ-ARK’s explorer comparison refers to the 10 September explorer; the explorer has since been redeployed.
Explorer visitors appear only in the economic-scenario part of the October follow-up, which reports their median implied growth.
Two statements about Unjournal guidance (optional journal-tier ratings; an “April 2026” policy date) could not be located and are labelled.
The description of GATE was reduced to what could be confirmed.
The list of what changes between scenarios now includes faster vacancy posting.

What was removed

The report’s section “Dynamic page, reproducibility, and limitations” described a sandbox website, a downloadable Netlify package, a skill file and a draft of The Unjournal’s form, none of which are available. It has been replaced by this page. Its epistemic summary (high confidence in the computation; moderate confidence in the classes of mechanism; low confidence in unconditional predictions; nothing established about the probability of transformative AI, intelligence-explosion speed, catastrophic risk or welfare) is reflected on the overview.

Names

These pages name only the paper’s authors and the authors of public work they cite.

Sources

Glossary

ε, capital-supply elasticity: How strongly the supply of capital (machines, buildings, data centres) rises when its return rises. All three scenarios use ε = 3, half the long-run value implied by Moll, Rachel and Restrepo (2022); the authors lean high because AI capital is largely compute, financed globally and built quickly. A high ε lets capital flow in, which keeps the return down and passes more of AI’s gain to wages; ε = ∞ fixes the return; ε = 1 makes capital scarce. The schedule stands in for household saving, which the model does not include.
σ, elasticity of substitution across tasks: How easily production can shift between tasks. The paper uses σ = 0.5. Below 1, tasks are gross complements: output needs every task, so a task that stays slow limits the gain from speeding up the others.
labour share: The fraction of national income paid as wages. The rest is the capital share. It is 60% in the model’s base year and on the no-AI path.
ρ, reinstatement ratio: New human tasks created per unit of automated tasks: 0.5 (modest), 0.25 (substantial), 0 (extreme). In the model the new tasks go to cognitive workers and immediately offset part of the displaced work.
d, diffusion: Of the tasks AI can perform, the fraction of task instances actually done with AI. 0.10 in mid-2026; 0.2, 0.4 or 0.6 by 2030.
m, affected task share: The fraction of the economy’s tasks that AI affects, each task weighted by its share of the base-year wage bill. Anchored at 0.14 in mid-2026 using Anthropic’s observed-exposure data; 0.2, 0.3 or 0.5 by 2030. Its ceiling is the cognitive occupations’ share, 0.624.
ψ, automation share: Of the task instances done with AI, the fraction AI performs outright (automation) rather than alongside a worker (augmentation): 0.5, 0.75 or 0.9.
ξ, wage rigidity: How slowly cognitive workers’ pay, relative to the economy-wide wage, moves toward the level that would clear their labour market: the gap shrinks by the factor ξ a year. ξ = 0.5 (the baseline) is a half-life of one year; ξ = 0.9 is a half-life of about seven years; ξ = 0 clears the market every month.
a, productivity gain: The log gain on each task instance done with AI: such an instance is eᵃ times as productive. In 2030, a = 0.30, 0.45 or 0.80 (about 1.35, 1.57 or 2.23 times).
μ, cross-group search discount: How effectively unemployed workers search for jobs outside their old occupation group, relative to inside it. μ = 1 would make both equally effective. Normal times: 0.17; the scenarios use 0.17, 0.08 and 0.04.
θ_H, vacancy-posting speed: The fraction of the other occupations’ hiring shortfall that firms post as openings each month: 0.10, 0.25 or 0.50 across the three scenarios. It stands for the recruiting and training capacity of the expanding occupations. Faster posting in the more disruptive scenarios speeds re-employment.
semi-endogenous growth: Growth models in which research produces new ideas, but ideas get harder to find as the stock grows, so sustained growth needs ever more research. Here more GDP funds more research and raises the ideas stock slightly. The ideas stock does not speed up AI itself.
TFP: Total factor productivity: output growth not explained by more capital or labour.
no-AI path: The model’s path for the same economy without the AI shock. Most results are percentage gaps from this path in 2030, not changes from today.
cognitive occupations: The model’s knowledge-work group (management, professional, technical, office and sales occupations), about 62% of employment. AI directly affects only this group; all other occupations form the second group.
observed exposure: Anthropic’s measure of AI exposure by occupation: the share of an occupation’s task time spent on tasks with significant work-related Claude use, with assistive use counted at one half. The paper uses it to anchor m in mid-2026.
Econ-ARK REMARK: A replication format from the Econ-ARK project: a public repository whose code regenerates a paper’s results and fails when a number stops reproducing.
clean-room implementation: A second implementation written from the paper alone, without reading the authors’ code.
first-order approximation: A linear approximation that is accurate for small shocks. The paper reports results from its exact model; first-order formulas help explain the mechanisms.
percentile rating: Unjournal ratings place a paper within a reference group of comparable work: 80 means better than 80% of that group on the criterion.
90% credible interval: A range the rater believes has a 90% chance of containing the value they are rating.
construct validity: Whether a measure captures the concept it is supposed to measure.
recursive self-improvement: AI systems speeding up the research that produces better AI systems, which could make AI progress accelerate.
wage bill: Total wages paid to a group of workers: the wage times the number employed.
automation and augmentation: Automation: AI performs a task instance with no worker. Augmentation: a worker performs it with AI’s help, so the worker stays employed but is more productive.
sticky wages: Wages that adjust slowly to a fall in demand, so the adjustment shows up more in jobs and unemployment than in pay.
reduced form: A summary relationship that stands in for a more detailed mechanism. Here one supply curve stands in for saving, investment, construction and financing decisions.
consequences module: A component that takes AI capability and adoption paths as given and works out their economic effects, as opposed to a module that predicts the paths themselves.

Annotate this page with Hypothes.is

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