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.

Research agenda and what to monitor

The framework is not good enough to forecast 2030, but it is good enough to say what to measure now and which extensions would matter most.

Extensions ranked by decision value

Priority

Priority	Extension	Why it bears on decisions	Proposed by
Very high	Couple the economic model to endogenous AI-R&D and capability feedback	Determines transition speed, warning time and whether the 2030 path can accelerate sharply.	AI report
Very high	Replace the reduced-form capital supply with explicit compute, energy and data-centre investment	Capital supply changes even the sign of the average-wage effect in the extreme scenario (−9.2% at ε = 1, +9.7% at ε = 3). The authors plan to add an Euler equation (footnote 13).	AI report
Very high	Run a global uncertainty analysis instead of three bundled scenarios	Shows which uncertain inputs drive policy-relevant outcomes, and stops scenario labels acquiring implicit probabilities.	AI report
High	Hold the 2030 endpoints fixed and vary shock timing	Informs governance lead time and labour-market adjustment capacity.	AI report
High	Model richer labour transitions and earnings losses	The unemployment headline depends on wage rigidity; worker welfare does not. The paper gives movers the new group’s wage at once, with no loss for tenure or skills.	AI report
High	Measure tasks representatively and validate the survey mapping	Needed before survey beliefs can credibly calibrate macro capability.	AI report
High	Add a physical-automation (robotics) block	Tests whether non-cognitive occupations are a durable place for displaced workers.	AI report
High	Model ownership, taxes, transfers and market power	Turns output and factor-share results into welfare and distribution questions.	AI report
High	Feed expert forecasts (such as Karger et al.) through the model and compare with the survey-implied scenarios	Shows where expert and public expectations differ in their implications.	Earlier AI assessment
High	Decompose the scenario gaps input by input, including posting speed θ_H and the search discount μ	The named scenarios move seven inputs together; one of them, faster vacancy posting, cushions the disruptive cases.	Claude
Medium-high	Model international trade, adoption and capital ownership	Needed before treating a US transition model as a global-priorities model.	AI report
Medium-high	Check survey nonresponse: compare the 3,259 complete responders with the full sample, and impute	Complete-case results may differ from the representative sample.	Earlier AI assessment

Priorities are the AI report’s, except rows proposed by the earlier AI assessment or by Claude .

The paper’s own sensitivity tables already show why joint uncertainty matters more than one-at-a-time checks. A world with heavy automation, no new tasks, scarce capital, sticky wages and poor matching across occupations differs in kind from one with the same AI capability but abundant capital and flexible institutions. A useful next version of the explorer would replace “which of three stories are we in?” with an ensemble view: distributions over inputs, their correlations, the resulting spread of outcomes, and which inputs account for most of that spread. The named scenarios could stay as landmarks.

What to monitor now

Waiting for aggregate unemployment would be a mistake. A better monitoring system tracks these together.

Block

Observable	Model input it informs	Possible decision	Example current evidence
Representative task capability	m, a	Update economic-transition and safety scenarios	None recorded here
Actual workplace usage	d	Distinguish technical capability from the realised shock	Massenkoff & McCrory, Labor market impacts of AI: a new measure and early evidence
Autonomous versus assistive use	ψ	Assess displacement pressure	Appel et al., Anthropic Economic Index report: Uneven geographic and enterprise AI adoption (the paper’s source for ψ: about half automation on Claude.ai, three-quarters on the API); later: March 2026 report
New human tasks and occupations	ρ	Assess whether new tasks offset automation	None recorded here
Entry-level hiring and wage growth by exposure	ξ, μ, θ_H	Trigger a review of training, wage insurance or income support	Massenkoff & McCrory, Labor market impacts of AI: a new measure and early evidence
Time to move between occupations, and earnings lost	μ, θ_H	Size transition assistance and social insurance	None recorded here
Compute prices, construction pipelines and electricity constraints	ε	Update who captures the productivity gains, and how quickly	None recorded here
Robot capability against robot cost Claude	Outside the model (physical automation)	Judge whether other occupations stay a safe destination	Legate-Yang & Massenkoff, What work can robots do?
AI’s contribution to AI research	Outside the model (AI-R&D feedback)	Update governance lead time separately from labour-market signals	Chan et al., What If Automating AI R&D Triggers an Intelligence Explosion?
Expectations of the public, experts and lab staff Claude	Scenario choice	Show how far beliefs about the inputs diverge	Sacher, Cotter, Jones, Korinek & McCrory, The AI expectations gap

Rows from the AI report; the robot-cost and expectations rows and the evidence column were added by Claude . Evidence links show that a measurement exists, not that it settles the question.

AI’s contribution to AI research needs its own trigger system. The macro model cannot tell decision-makers when an acceleration has begun, because that mechanism sits outside it. Economic-policy monitoring and safety monitoring should be linked but kept distinct (Chan et al., 2026) .

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