# Structural sensitivity of Economic Scenarios for Transformative AI: draft menu Draft of 29 October 2026. Requested by David Reinstein (The Unjournal); drafted by Claude (Anthropic), not checked by a person. Not an Unjournal evaluation. Claude is made by Anthropic, which publishes the paper. Web page: https://uj-prioritization-dashboard.netlify.app/ai-governance/scenarios-review/structural.html ## Claims to check (fixed before any run) - K1: Extreme scenario: by 2030 GDP is 32 percent above its no-AI path, growing 15 percent a year (abstract; p. 38) - K2: Extreme scenario: the labor share falls from 60 to 45 percent (abstract) - K3: Extreme scenario: nearly one in five cognitive workers is unemployed (abstract; p. 38) - K4: Extreme scenario: the cognitive wage is 11.5 percent below its no-AI path; other wages are 34 percent above (p. 4) - K5: The aggregate gain is almost three times the cognitive loss; a transfer of about 9 percent of GDP would compensate (pp. 35, 39) - K6: Substantial scenario: cognitive unemployment rises from 2.9 to 4.5 percent; GDP is 8.3 percent higher (p. 34; Table 3) - K7: Modest scenario: AI adds less than half a point to growth and a tenth of a point to unemployment (abstract) - K8: The median survey respondent's answers are consistent with the substantial scenario (abstract; p. 36) - K9: The mechanics are the same in every scenario: output rises, cognitive employment shrinks, other employment grows (p. 38) ## Choice points ### 1. Capital supply: how elastic, and how fast (Production and capital) The paper's choice: Capital is supplied instantly along a constant-elasticity schedule, Δln K = ε Δln r (Equation 6, p. 9), the reduced form of Moll, Rachel and Restrepo’s (2022) long-run wealth schedule. ε = 3 is “higher than five years of saving at historical rates would imply”, because the capital doing automated cognitive work is mostly compute, “financed in a world market and built in a year or two” (p. 26). Table 5 varies ε (p. 37). Alternatives: - Partial adjustment: the stock closes part of its gap to the long-run schedule each month, with the long-run ε of about 6 that the paper derives from Moll et al. (p. 9, footnote). **Piloted: results on the web page.** - Time to build: investment is committed months before the capital can be used (Kydland and Prescott 1982). - Accumulation from saving, at a fixed saving rate or by forward-looking households (an Euler equation), instead of a supply schedule. - Separate compute from other capital, each with its own supply. Claims it could change: K1, K2, K4, K5 Already tested: Table 5 varies ε. The isitcredible review charged for the extra capital under several financing assumptions. No dynamic adjustment before this pilot. Effort: Pilot done. Time to build: hours. Saving-based accumulation: days, and a different solution method. Two capital types: a day or more. Your view (explore first / worth exploring / low priority / not reasonable), values or sources: ### 2. Wage setting (Labor market) The paper's choice: The cognitive wage closes part of its gap to the market-clearing wage each month (Equation 30, p. 19, after Blanchard and Galí 2007). “What is rigid is the cognitive discount”: the rigidity is relative to the economy-wide wage. Other occupations’ wages clear. Layoffs are job rationing (Michaillat 2012; p. 19). Table 6 varies the rigidity ξ. Alternatives: - Rigidity in the level of the cognitive wage (against its own no-AI path) rather than in its discount. - Rigidity downward only. - A wage norm or bargained wage with a slow-moving outside option (Hall 2005). - Rigid wages in both groups. Claims it could change: K3, K4, K6 Already tested: ξ one at a time (Table 6) and jointly with the paper’s other values: the isitcredible review found wage rigidity accounts for 0.56 of the variance of extreme cognitive unemployment over the paper’s own values. The rule’s form was not varied. The pilot below shows it matters for how other choices play out. Effort: Hours for each. Your view (explore first / worth exploring / low priority / not reasonable), values or sources: ### 3. Matching function (Labor market) The paper's choice: Hiring follows the den Haan, Ramey and Watson (2000) form with curvature ι = 1.27 (p. 20; Table 1, p. 24); matching efficiency is set to hit the normal-times job-filling rate of 0.65. Alternatives: - Cobb–Douglas matching with an elasticity near 0.5 (Petrongolo and Pissarides 2001), recalibrated to the same normal-times targets. Claims it could change: K3, K6 Already tested: Not varied; the curvature is held at 1.27. Effort: Hours. Probably small at these targets, but unchecked. Your view (explore first / worth exploring / low priority / not reasonable), values or sources: ### 4. Moving between occupation groups (Labor market) The paper's choice: A single discount μ on job search across the two groups (p. 19, citing Kambourov and Manovskii 2009): 0.17 in normal times, fitted to switching data, and 0.08 and 0.04 in the substantial and extreme scenarios, which have no outside anchor (p. 28). Alternatives: - Search directed toward the group whose wages are rising. - Mobility that differs across cognitive occupations. - A one-off retraining cost or delay instead of a permanent discount. Claims it could change: K3, K6, K8 Already tested: μ jointly with the paper’s other values; illustrative, uncalibrated worker responses to wage gaps (isitcredible review). Effort: Hours for directed search (partly coded in the review); a day for the others. Your view (explore first / worth exploring / low priority / not reasonable), values or sources: ### 5. Who enters and leaves the labor force (Labor market) The paper's choice: All separations, including exits from the labor force, pass through the unemployment pool; entrants join the groups and the pool in proportion to their sizes (pp. 20–21). Alternatives: - Exits leave the labor force directly. - Entrants who replace cognitive workers join the growing group instead. Claims it could change: K3, K4, K6 Already tested: The isitcredible review built a family of routings: 2030 cognitive unemployment of 2.0–4.5 percent in the substantial scenario and 13.9–17.6 in the extreme, one change at a time. Effort: Hours, to combine with the other choices. Your view (explore first / worth exploring / low priority / not reasonable), values or sources: ### 6. Worker heterogeneity and number of groups (Labor market) The paper's choice: Two groups of occupations; “workers are homogeneous in production” (p. 10). The cognitive group’s share of the wage bill is set equal to its employment share, 0.624. Alternatives: - Different pay levels: measured, the cognitive share of the wage bill is about 0.69. - Three or more groups that differ in exposure and mobility. - Displacement costs that differ by age or tenure. Claims it could change: K3, K5 Already tested: The isitcredible review varied both shares together; a version with different pay was not built. Effort: A day for different pay; days for more groups. Your view (explore first / worth exploring / low priority / not reasonable), values or sources: ### 7. Tasks within occupations (Production and capital) The paper's choice: Workers within an occupation are treated alike across its tasks. The authors note a literature in which they differ (Autor and Thompson 2025; Freund and Mann 2026; Althoff and Reichardt 2026) and write: “We do not currently incorporate this richness” (p. 6). Alternatives: - Expertise tiers: as the paper summarizes Autor and Thompson (2025), automating an occupation’s inexpert tasks raises the expertise and wage of the job that remains, and automating its expert tasks lowers both. - Standardization that lowers wages without lowering employment, as the paper summarizes Fukui, Nakamura and Steinsson (2026). Claims it could change: K2, K3, K4 Already tested: Not tested. Effort: A research prototype: days or more, and the result would be stylized. It could change the sign of the cognitive wage result. Your view (explore first / worth exploring / low priority / not reasonable), values or sources: ### 8. Substitution within a task (Production and capital) The paper's choice: An AI-performed and a human-performed instance of the same task substitute at the same elasticity, σ = 0.5, as tasks do across the economy (p. 8). Alternatives: - A nested production function with a separate within-task elasticity. Claims it could change: K2, K4, K5 Already tested: The isitcredible review built a nested version of the full-employment economy: the extreme 2030 labor share is 44.6 percent at a within-task elasticity of 0.5, 42.4 at 1 and 37.4 at 3. Unemployment effects were not simulated. Effort: A day, to run it through the full monthly model. Your view (explore first / worth exploring / low priority / not reasonable), values or sources: ### 9. New tasks for people (Production and capital) The paper's choice: A constant reinstatement ratio ρ (0.5, 0.25 and 0 across the scenarios): new tasks belong to cognitive occupations and enter at the same unit cost (p. 11). The modest value follows Acemoglu and Restrepo (2019); the other two have no outside anchor (p. 28). Alternatives: - New tasks that arrive with a lag. - New tasks in other occupations. - A reinstatement rate that rises or falls with AI capability. Claims it could change: K1, K2, K3 Already tested: ρ over the paper’s values (the second-largest driver of extreme cognitive unemployment, isitcredible review). Effort: Hours for a lag; a day for the others. Your view (explore first / worth exploring / low priority / not reasonable), values or sources: ### 10. Innovation (Production and capital) The paper's choice: A semi-endogenous ideas stock (Romer 1990; Jones 1995), with parameters from Bloom, Jones, Van Reenen and Webb (2020) (p. 16; Appendix C). Alternatives: - AI raises research productivity directly. - Ideas raise the productivity of capital as well as labor. Claims it could change: K1 Already tested: Across the block’s parameters, extreme 2030 GDP moves only between 32.1 and 32.4 percent; letting ideas raise capital productivity has small effects by 2030 (isitcredible review). Effort: Hours. Low priority for the 2030 claims; more relevant beyond 2030. Your view (explore first / worth exploring / low priority / not reasonable), values or sources: ### 11. Shape of the capability path (Scenario inputs) The paper's choice: Capability and adoption follow logistic paths from mid-2026 anchors to each scenario’s 2030 values; the automation share and reinstatement are held constant (p. 11). Alternatives: - The same 2030 endpoints, reached gradually or in a late jump. Claims it could change: K1, K3 Already tested: The isitcredible review added one steeper path for its compensation figures. Effort: Hours. This changes the scenario inputs rather than the model; included because an earlier AI report suggested it. Your view (explore first / worth exploring / low priority / not reasonable), values or sources: