AI research review
Claims and what they rest on
Each section states a claim from the report, what supports it and how strong that support is. Coloured tags say where each statement comes from; hover or tap a tag for details.
1. The model separates things that “exposure” measures merge
“AI can do the task”, “firms use AI on the task”, “AI makes the task much cheaper” and “the human job disappears” are different events. The model gives each its own parameter: , , , and . The labour block adds how well displaced workers search outside their old occupation group () and how fast the expanding occupations post vacancies () (Sections 2.1.2 and 2.3) .
Anthropic’s own labour-market work draws the same distinction. Its measure, which the paper uses to anchor m, finds that “AI is far from reaching its theoretical capability: actual coverage remains a fraction of what’s feasible” (Massenkoff and McCrory, March 2026) .
Accounting illustration: how the definitions nest
What the paper’s model adds to this arithmetic
The last bar equals what the paper calls “net displacement”, (1 − ρ)·ψ·m·d. To first order the paper adds the “labor released by the gains” on augmented tasks, (1 − σ)·(1 − ψ)·m·d·a, because a worker using AI needs less time and only part of that saving is spent on more of the same task (Equation 19; exact form in Proposition 1, Equation 15) . Wages, capital and unemployment then depend on prices, capital supply and job search, which this illustration ignores.
Presets use the 2030 inputs in the paper’s Table 1 . At those values the paper itself says AI performs about 4%, 12% and 30% of the economy’s task instances in the modest, substantial and extreme scenarios (Sections 4.1–4.3), the same as the second bar. The mid-2026 anchor sets only m = 0.14 and d = 0.10. The slider for m stops at 0.62 because m is capped at the cognitive occupations’ share, 0.624.
2. Simple formulas explain the mechanisms
To a first approximation, AI adds about sL·m·d·a to , where sL = 0.6 is the base labour share. The return to capital rises with that productivity gain and with automation’s extra demand for capital, damped by the . Each 1% rise in the rental rate lowers the wage by about two-thirds of a percent, so the more elastic capital supply is, the more of the gain reaches workers.
The paper’s own example uses the substantial scenario’s 2030 inputs: TFP rises by about 0.6 × 0.30 × 0.40 × 0.45 = 3.2% (2.9% exactly). With the rental rate fixed the wage would be 5.4% above the no-AI path; at ε = 3 the rental rate is 4.6% higher and the wage only 1.9% higher (Section 2.1.3) . The reported scenarios use the exact model, because first-order approximations may not be accurate for large gains.
3. The extreme scenario is a stress test, not a forecast
In the extreme scenario in 2030, GDP is 32.4% above the no-AI path and growing at 15.4% a year; cognitive employment is 21.5% below its mid-2026 level; cognitive unemployment is 17.9% (11.9% for all workers); and the labour share is 45.2%, against 60% without AI. The average wage is 9.7% above the no-AI path while the cognitive wage is 11.5% below it, and capital income is 81.4% higher. The authors call 17.9% “far above any postwar rate in the United States”, and compare the economy-wide 11.9% with annual peaks of slightly below 10% after the global financial crisis and about 8% in the Covid-19 recession (Section 4.3).
The useful point is that enormous productivity gains and serious losses for exposed workers occur together. Aggregate income is not exposed workers’ income, bargaining position or welfare. But the scenario bundles many assumptions (see what each scenario assumes), so it does not show the effect of more capable AI alone. The explorer’s independent controls allow combinations that the three labels hide.
4. Capital supply largely decides who gains
The says how readily machines, buildings and data centres are supplied when their return rises. It is “neither measured nor a dimension of the scenarios” (Section 4.5). The paper sets ε = 3 and reruns the substantial and extreme scenarios at 1, 6 and infinity. In the extreme scenario the GDP gain runs from 21.3% to 43.3%, and the average wage from 9.2% below the no-AI path to 30.1% above it. The sign of the average-wage effect depends on this one assumption.
Change ε and see the 2030 outcomes move
Source: the paper’s Table 5 ; also matches Econ-ARK’s transcription. No person has checked these values.
Table view
Substantial scenario, 2030
| Outcome | ε = 1 | ε = 3 (baseline) | ε = 6 | ε = ∞ |
|---|---|---|---|---|
| GDP vs no-AI path (%) | 6.4 | 8.3 | 9.1 | 10.0 |
| Average wage vs no-AI path (%) | −1.6 | 2.1 | 3.7 | 5.6 |
| Net return to capital (% a year) | 7.6 | 7.0 | 6.8 | 6.5 |
| Capital stock vs no-AI path (%) | 9.3 | 13.8 | 15.7 | 18.2 |
| Labour share (%) | 55.1 | 56.1 | 56.5 | 57.0 |
Extreme scenario, 2030
| Outcome | ε = 1 | ε = 3 (baseline) | ε = 6 | ε = ∞ |
|---|---|---|---|---|
| GDP vs no-AI path (%) | 21.3 | 32.4 | 37.2 | 43.3 |
| Average wage vs no-AI path (%) | −9.2 | 9.7 | 18.3 | 30.1 |
| Net return to capital (% a year) | 10.3 | 8.3 | 7.5 | 6.5 |
| Capital stock vs no-AI path (%) | 33.5 | 56.3 | 67.1 | 82.2 |
| Labour share (%) | 41.2 | 45.2 | 46.9 | 49.1 |
“Capital” here is one good with one supply curve. The paper says so plainly: “a single capital good supplied along an exogenous schedule with fixed elasticity, which does not distinguish compute from other capital and does not explicitly model saving decisions” (Section 1) . A transformative transition may instead be limited by chips, fabs, data centres, electricity, grid connections, construction and finance, each with its own lags and rents. The authors’ reason for a fairly high ε points the other way: “the capital that performs automated cognitive work is disproportionately compute, financed in a world market and built in a year or two” (Section 3.1). They also note that adding an Euler equation, linking the return to capital to growth, “could raise the return to capital substantially” (footnote 13) . For questions about AI ownership, taxation, compute investment or “abundance”, ε matters at least as much as the capability inputs.
5. The unemployment headline is much less robust than the worker-loss result
The extreme scenario’s 17.9% cognitive unemployment depends on . With fully flexible relative wages (ξ = 0) it is 2.6%; with very sticky wages (ξ = 0.9) it is 24.0%. Over the same range the cognitive wage moves from 42.2% below the no-AI path to 2.8% above it, and cognitive employment falls by 1.3% or by 28.5%. The shock to cognitive labour demand is large in every case; the settings decide whether it shows up in pay, jobs or unemployment. The authors put it the same way: “The effects of AI are felt in the cognitive labor market in either wages or unemployment.”
The baseline ξ = 0.5 is a judgement. The authors note that wages of workers who keep their jobs are rigid, which “suggests ξ near 0.9”, while wages of job changers are flexible; they choose 0.5 because the shock is “large, visible, and permanent” (Section 3.4) .
Where the same shock lands: pay, jobs or unemployment
Source: the paper’s Table 6 ; also matches Econ-ARK’s transcription. The paper reports ξ = 0 only for the extreme scenario. Each panel has its own scale.
Table view
Substantial scenario, 2030
| Outcome | ξ = 0.5 (baseline) | ξ = 0.75 | ξ = 0.9 |
|---|---|---|---|
| GDP vs no-AI path (%) | 8.3 | 7.9 | 7.7 |
| Average wage vs no-AI path (%) | 2.1 | 2.2 | 2.3 |
| Cognitive wage vs no-AI path (%) | −0.3 | 0.7 | 1.4 |
| Other-occupation wage vs no-AI path (%) | 5.9 | 4.5 | 3.7 |
| Cognitive employment since mid-2026 (%) | −3.9 | −4.6 | −5.0 |
| Cognitive unemployment rate (%) | 4.5 | 5.1 | 5.4 |
| All-worker unemployment rate (%) | 4.6 | 4.9 | 5.2 |
Extreme scenario, 2030
| Outcome | ξ = 0 | ξ = 0.5 (baseline) | ξ = 0.75 | ξ = 0.9 |
|---|---|---|---|---|
| GDP vs no-AI path (%) | 36.6 | 32.4 | 30.5 | 29.2 |
| Average wage vs no-AI path (%) | 1.6 | 9.7 | 11.1 | 11.9 |
| Cognitive wage vs no-AI path (%) | −42.2 | −11.5 | −2.9 | 2.8 |
| Other-occupation wage vs no-AI path (%) | 70.1 | 33.6 | 25.8 | 21.1 |
| Cognitive employment since mid-2026 (%) | −1.3 | −21.5 | −25.9 | −28.5 |
| Cognitive unemployment rate (%) | 2.6 | 17.9 | 21.7 | 24.0 |
| All-worker unemployment rate (%) | 3.1 | 11.9 | 13.9 | 15.2 |
This changes what should be monitored. Aggregate unemployment is a poor single warning signal; wages, hiring, entry-level openings, moves between occupations, unemployment duration and cumulative earnings losses all matter. Anthropic’s March 2026 evidence fits this: “no systematic increase in unemployment for highly exposed workers since late 2022”, but “suggestive evidence that hiring of younger workers has slowed in exposed occupations” (labour-market note) .
6. The survey is useful elicitation but weak calibration
The paper surveyed 10,980 US adults through Morning Consult, 11–23 August 2026, and mapped their answers into five model inputs (Table 2). The outcome table (Table 4) uses the 3,259 respondents who answered all five items and holds the other four inputs (ρ, θH, ξ and ε) at substantial-scenario values for everyone . The median answers land close to the substantial scenario: GDP 8.6% above the no-AI path and unemployment around 4.6%.
The capability question asks when AI will do each of eight tasks “as well as a skilled professional”, from routine business emails to a Nobel-level discovery. The tasks “were chosen to systematically represent increasing levels of difficulty from routine office work to Nobel-level frontier science and to be recognizable to a general audience” (Appendix B.3) ; they are not a sample of the wage bill. The mapping turns the number of tasks expected by 2030 into a share of knowledge work, n/8 × 0.95, then multiplies by the cognitive share for the economy-wide m (Table B.1). The median adult expects six of the eight, which gives m = 0.44, between the substantial and extreme scenarios.
The paper also asked employed respondents what share of their own working hours goes to tasks AI will be able to do about as well as they can by 2030. Among cognitive workers the median answer is 27%, “against 71 percent from the task grid”. The authors read this as respondents expecting “more of the listed tasks to be within AI’s reach than of their own work, a pattern common in surveys about new technology” (Appendix B.3) . The report treats the same gap as a construct-validity problem for the ladder, which drives m.
Two ways of asking how much work AI could do by 2030
Sources: mapping from the paper’s Table B.1 and the 27% from its Appendix B.3 ; the 28%, 33% and 75% averages from The AI expectations gap, a later Anthropic note with the same public sample . Six tasks is the median adult’s answer (Section 3.5).
The two questions differ in scope (one’s own job against eight listed tasks) and in standard (“about as well as they can” against “as well as a skilled professional”), so some gap is expected even with sound measurement . A gap this large still means the ladder is not a validated estimate of economy-wide capability. Treat the survey results as structured elicitation (“what follows if people’s answers pass through these mappings, with everything else fixed”), not as a probability distribution over 2030 outcomes. Views are also widely spread: about 30% of respondents expect AI to save no time on a suitable task, while 49% expect it to halve the time (Section 3.5) .
Expectations differ sharply by group. Anthropic’s October follow-up runs each group’s median answers through the model: implied 2030 GDP growth is 5.4% for US adults, 12.5% for self-selected explorer visitors and 17.1% for Anthropic staff (The AI expectations gap, its Figure 7) . The staff median is above the extreme scenario’s 15.4%. Such dispersion is informative about uncertainty, and makes assigning a probability to any named scenario harder.
7. The small ideas channel says nothing about AI-R&D feedback
The model includes a ideas block: more output funds more research, which raises the stock of ideas and so wages and TFP. Even in the extreme scenario the ideas stock is only 0.61% above the no-AI path by 2030 (0.07% and 0.20% in the other two; Table 3). Those ideas do not improve AI itself. The authors are explicit: “there is no feedback from ideas to the gains from AI”, so the gains from “can be thought of as being hard-coded into at”, adding that “this channel merits further exploration in the future” (Section 2.2.1 and footnote 6) . In the introduction they call the innovation effects “likely a lower bound”, since the model cannot generate the explosive acceleration of Davidson, Halperin, Houlden and Korinek (2026). The explorer notes that “more than one reviewer argued that we may be underestimating how much AI could accelerate technological progress itself” .
So the small ideas increment answers a different question. It is not evidence that feedback from AI research to AI progress will be small. See the first critique.
8. “The resources to compensate exist” depends on the case
In the extreme scenario total labour income ends up almost where it would have been without AI, and all of the GDP gain goes to capital income. The authors calculate that “a transfer of about 9 percent of GDP, roughly the size of Social Security and Medicare combined, would hold cognitive workers’ income at its no-AI level and leave the rest of the economy more than 20 percent ahead”, and conclude that “the resources to compensate them exist” (Sections 4.3 and 5). They add that transfers of that scale “have no precedent” and that how such resources reach people “is not something growth delivers by itself”.
Their footnote 14 shows the other side. Take the substantial scenario’s capability, adoption and productivity, but the extreme scenario’s automation share, re-employment and lack of new tasks, with inelastic capital (ε = 1): GDP is 7.2% above the no-AI path, total labour income is lower by 4.3% of no-AI GDP, and holding cognitive workers’ income at its no-AI level “would take a transfer equal to 84 percent of GDP gains”.
Values from the paper ; “about a third” follows from the paper’s statement that the gain is almost three times the loss . Econ-ARK reproduces the 84% when cognitive employment is counted from mid-2026 and gets 87% from a 2024 base, a definitional point the footnote leaves open (Econ-ARK notes) . In this “so-so automation” case, wide automation with modest productivity gains, compensation would use up most of the gain; the model produces it “with parameter values inside the ranges we consider” (Section 4.5).
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