Korinek, Jones, Sacher, Cotter & McCrory 2026, "Economic Scenarios for Transformative AI" (Anthropic Institute Working Paper No. 2026-02)

The paper turns five survey items into five model inputs, fixes four other inputs at the substantial scenario's values (Table 4 note, p. 36), and computes the comparison two ways. The model at Table 2's item-by-item medians, the computation the Introduction describes, is not printed; it gives GDP 9.88 percent above the no-AI path, cognitive employment −4.49, cognitive unemployment 5.16 and overall unemployment 4.89. Table 4's medians of per-respondent runs, over the 3,259 respondents who answered all five items, with survey weights, give 8.6, −4.2, 4.6 and 4.6 (p. 36). Both are nearest the substantial scenario (GDP 8.3, cognitive employment −3.9).

Both computations use the task-grid capability answer (m = 0.44). The paper prints a second capability measure: employed cognitive-occupation respondents put the share of their own hours AI can do by 2030 at a median of 27 percent, against 71 from the task grid (p. 47), a gap the paper calls "a pattern common in surveys about new technology". Scaled the same way, the own-hours answer is m = 0.160 (0.168 without the 0.95 in-person factor). At that value every outcome is nearest the modest scenario: GDP 3.70, cognitive employment −1.15, cognitive unemployment 3.31 and overall unemployment 3.97. The nearest scenario switches between m = 0.2175 and 0.2575, inside the interval between the two printed measures. The paper does not run the own-hours alternative.

With the task-grid measure, GDP and cognitive employment stay nearest the substantial scenario under every reading of the re-employment answer, every setting of the fixed inputs and random item non-response. Under no variant does the survey land near the extreme scenario. The own-hours item covers employed cognitive workers only, and p. 49 discloses that "the median respondent" means item-by-item medians.

The paper converts the survey's re-employment answer into the model's search discount with the rule μ = 0.17 × 3/months, anchored on three months as the normal-times answer (Table B.1, p. 48; p. 49). The median respondent expects a displaced worker who must switch occupation to take about eight months to find a new job, which the rule turns into μ = 0.064, between the substantial scenario's 0.08 and the extreme's 0.04. On that basis p. 30 says the expected time is "slower than the substantial scenario assumes and faster than the extreme".

At μ = 0.17 the model's own job-finding rates give a cognitive-origin worker an expected wait of 27.7 months for a job in the other occupation group, so each Table B.1 code stands for 8 to 15 times its stated months. Read in the model's units, eight months is faster than the substantial scenario implies under every model-consistent reading: faster than all three scenarios if respondents meant time to a job in the other group, and between the modest scenario (3.2–3.5 survey months) and the substantial one (9.1–9.8) if they meant total time to a new job.

The reading also sets where the survey's unemployment lands. Read as time to a cross-group job, survey cognitive unemployment is 2.9–3.4, the modest scenario's level. Read as total re-employment time, the object question 5 names ("to find a new job", p. 46), it is 4.2–4.4 cognitive and 4.5 overall, nearest substantial. The survey data cannot say which reading respondents used. GDP and cognitive employment stay nearest substantial under every reading (GDP 10.3–10.9).

The deposit holds the scenario explorer's compiled model code, the authors' public implementation fetched on 27 Oct 2026, and a snapshot of the explorer page. It is not an author replication package. It lacks the calibration code that built the Table 1 and Table A.2 targets from CPS, IPUMS, JOLTS and AI-usage data; the survey microdata, weights and coding code; and the code that printed the tables and figures.

As published, the following cannot be checked:

• Table 4's per-respondent quantiles, and Table 2's medians beyond their printed cells;
• whether Table 2 is weighted (Table 4's note says survey weights; Table 2's does not say);
• whether Table 4 used the explorer's survey mappings;
• how the mid-2026 observed-exposure anchor of 0.14 was built from usage data;
• which code printed Tables 3, 5 and 6, which are the explorer's outputs.

An independent recompute of the exposure anchor (Econ-ARK 2026) gives 0.1212, which moves extreme GDP from 32.41 to 32.20 percent above the no-AI path and extreme cognitive unemployment from 17.95 to 18.19.

Table 4 uses the 3,259 respondents who answered all five items, 29.7 percent of the 10,980 surveyed, and the paper states no assumption about item non-response. At the corners Table 2's printed quartiles allow, the median economy runs from GDP 1.8 to 31.7 percent above the no-AI path, a worst-case bound rather than an estimate. Under random item non-response, synthetic respondents matched to every Table 2 quartile give median GDP of 6.5–10.2, and all 29 runs, across 20 dependence structures, are nearest the substantial scenario on all four outcomes. None of them gives Table 4's GDP quartiles of [3, 19], so the print does not reconcile its two survey tables. That gap depends on how the synthetic sample is built and is not evidence of selection.

Two coding choices are fragile or unstated:

• Adoption. The share of capable tasks people "will regularly use it for" is coded directly as the share of task instances done with AI. The median run moves nearer the modest scenario as that share falls: the first outcome switches at 0.2475 and the last, GDP, at 0.1975, an instance share of under about half among regularly used tasks.
• The explorer's quiz code. It maps "less than 1 month" to μ = 0.51 (0.17 × 3/max(1, months)), where Table B.1 prints "0.17×3/months, capped at 1", which gives 1 (p. 48). It also runs capability and adoption answers below today's levels at the mid-2026 anchors (0.07 or 0 become 0.14; 0.05 becomes 0.10), which the print does not say, and a respondent who marks every automation task "workers do it alone" gives 0/0, with no stated handling. Whether Table 4 used any of these rules cannot be checked without the survey code.

Three other choices move the level but leave the survey nearest substantial: fixing reinstatement and posting speed (GDP 8.0–11.7 across their scenario values), the productivity answer's code edge that p. 49 notes (6.2–12.2), and reading the productivity question's present tense as today's gain (11.0). Fixing reinstatement and posting speed also compresses the upper tail: a respondent whose five answers equal the extreme scenario's reaches GDP 27.39 and cognitive unemployment 13.42, against the extreme scenario's 32.41 and 17.95. Bundling the two inputs with each respondent's capability answer, as Table 1's scenarios do, leaves the synthetic median unchanged (GDP 6.5 against 6.3, and 5.5 against 5.6, with the inputs fixed) and raises the upper GDP quartile from 13.5 to 14.8 and from 17.7 to 19.7 at two dependence structures, so Table 4's upper quartiles understate what the same answers imply under the scenarios' own bundling.

In the model all separations pass through the unemployment pool (p. 21), and exits from the labor force are 0.05 of the 0.11 annual quit rate (Table A.2, p. 44). Every quit from a cognitive job enters the pool as a cognitive-origin searcher, who crosses to the other occupation group only at the search discount μ. In the modest and substantial scenarios no worker is laid off in any month, so cognitive employment falls only through quits that are not replaced, and the substantial scenario's whole rise in cognitive unemployment is the search of quitters, including the people who replace retirees.

Routing those replacements differently, while keeping the calibrated steady state, gives these 2030 values:

• "Nearly one in five" (abstract; p. 38) stands as specified and under proportional entry (17.62), not once a quarter or more of the replacements enter the growing group (16.99 with a quarter, 13.85–16.02 with half or more).
• The extreme wage split, −11.54 and +33.61 against the no-AI path as specified, is −8.50 and +29.66 with a quarter redirected and −2.71 and +22.23 with all.
• The substantial cognitive wage ends 1.03 to 1.45 percent above its no-AI path once a quarter or more are redirected, against the Introduction's clause that cognitive wages "grow more slowly than they would without AI" (p. 4).

The paper's entrant rule (p. 20) points to proportional entry. The human-capital rationale for the search discount (pp. 19–20) points the other way for first-time entrants, as does p. 21's remark that past adjustment ran through entry and exit. The paper and its deposit contain no measure of the share, and sending all replacements to the growing group is the strong end of the range, not a central figure. With AI switched off, the routing rules leave 2030 cognitive unemployment at 2.88–2.90 against the no-AI path's 2.85, an offset of 0.03–0.05 points in the routed levels.

Under every routing, extreme cognitive unemployment stays in the teens, "far above any postwar rate in the United States" (p. 35); the relative cognitive wage falls steeply and the extreme cognitive wage stays below its no-AI path in 2030; extreme GDP is unchanged or higher (36.41 with all redirected) and the labor share stays at 44.96–45.21. The mechanics sentence on p. 38 says unemployment "depends both on" the size of the shock and the ease of switching, not only on them, and it stands.

p. 34 says the cognitive unemployment rate "rises from 2.9 percent in mid-2026 to 4.5 percent in 2030, more than a 50 percent increase". As specified, the substantial scenario's rate is 3.06 in mid-2026 and 4.53 in Jan 2030, a rise of 48 percent. The 4.53 is 59 percent above the no-AI level of 2.85, which is the 2.9 that Figure 4's note gives as the normal level; the sentence dates that no-AI level to mid-2026.

The size of the rise also depends on where the replacements of cognitive labor-force exits search: the 2030 rate is 3.06–3.77 with a quarter to all of them sent to the growing group, and 2.02, below normal, under the employment-overhang cap. The rise holds under the paper's search discounts and posting speeds one at a time (3.8–5.3) and under entry in proportion to target employment (4.49).

p. 39 says "The aggregate gain is nearly three times what cognitive workers lose in wages and employment, so the resources to compensate them exist", and p. 35 that a transfer of about 9 percent of GDP would hold cognitive workers' income at its no-AI level and leave the rest of the economy "more than 20 percent ahead". On the GDP basis the extreme scenario's 2030 gain is 32.41 percent of no-AI GDP against a cognitive wage-bill loss of 11.61, a multiple of 2.79, and the transfer is 8.8 percent of AI-economy GDP.

In 2030 the extreme scenario's capital stock is 56.3 percent above its no-AI path (Table 3, p. 31); at the no-AI capital-output ratio of 3.48, that is extra capital of 1.96 times no-AI GDP. The model has no consumption-investment split (footnote 5, p. 16) and justifies ε = 3 by compute "financed in a world market" (p. 26). At the scenario's own parameters:

Paying the transfer leaves the rest of the economy ahead by exactly the net gain, on whichever basis the gain is measured, so the resources exist as potential compensation on every closure consistent with the model's capital supply. The rest stays ahead with the capital foreign-owned in all 32 cells of the Table 5 × Table 6 grid and in all 1,296 extreme-capability combinations of the paper's own parameter values (+1.4 to +15.5). The multiple measures the transfer's size relative to the gain; compensation needs only a positive net gain.

"Nearly three times" (p. 39) and "almost three times" (p. 35) do not hold as measures of the resources once the capital is charged for; on the foreign-owned basis, the one p. 26's world-market financing implies, the gain is 0.54 times the loss. "More than 20 percent ahead" does not hold on the replacement, supply-price and foreign-owned bases, and on the foreign-owned basis it holds in none of the 1,296 combinations. The resources do not exist as consumable 2030 output if the build-up were financed at home, the closure p. 26 sets aside. p. 39 omits the qualifier "in the textbook sense" that p. 35 carries. p. 35 does not print the base of "more than 20 percent ahead"; 32.4 − 11.6 = 20.8 is the figure that rounds to it.

The transfer's size at the measured wage bill is open. The cognitive occupations' share of the May 2025 OEWS wage bill is 0.6933 on footnote 10's grouping (U.S. Bureau of Labor Statistics 2026c), against the employment share of 0.624 the model uses for both (p. 25; Table A.2, p. 44). The loss to be replaced is 9.2 to 12.9 percent of no-AI GDP depending on the construction, against 11.6; settling it needs a version of the model with heterogeneous pay. The measured share puts the multiple at 3.51–3.68 when run through the model and at 2.51 when the loss is scaled with the share at a fixed percentage loss, the construction that gives the 12.9. The GDP-basis multiple also moves with the across-task elasticity σ, which the paper does not vary: 1.58 at σ = 0.25 and 5.38 at 0.95.

p. 37 says that with so-so technologies "compensating workers would consume most of the gains", and footnote 14 that holding cognitive occupations' income at its no-AI level "would take a transfer equal to 84 percent of GDP gains". At the footnote's parameters GDP is 7.21 percent of no-AI GDP above its no-AI path and labor income 4.31 below. The transfer is 84.2 percent of the GDP gain when the cognitive wage bill is measured on employment since mid-2026, the decomposition p. 35's prose uses, and 87.0 percent on Table 3's no-AI-path employment. The footnote does not say which baseline it uses.

Paying the transfer leaves the gainers the whole net gain of 7.21 percent of no-AI GDP. On Table 3's no-AI-path employment the transfer is 6.27 percent of no-AI GDP (87.0 percent of 7.21), 46 percent of their gross gain of 13.5 (7.21 + 6.27).

The example keeps the substantial scenario's per-task gain (a = 0.45, 57 percent per instance), so it is not "only marginally more productive" in p. 37's sense; p. 14 defines so-so automation in relative terms, as displacement outweighing a modest productivity gain. A setting that is so-so in that sense and inside the paper's ranges (m 0.5, d 0.6, a 0.30, ψ 0.9, ρ 0, μ 0.04) gives a transfer of 173 percent of the gain at ε = 1 and 69 percent at ε = 3.

p. 38 says "almost all of the divergence comes after 2027, a year and a half from the mid-2026 anchor", "because the scenarios share today's readings". The gloss puts the date at Jan 2028. For GDP and cognitive-origin unemployment, 16.5–29.5 percent of the Jan-2030 gaps between scenarios is in place by Jan 2028, so 70–84 percent comes after it, about four-fifths at most. The small substantial-against-modest cognitive-wage gap is 34 percent in place at mid-2026 and 54 percent by Jan 2028. Read as after Jan 2027, 85–96 percent of the divergence is still to come.

The scenarios share the capability anchor (0.14) and the adoption anchor (0.10). They do not share the productivity anchor (0.30 / 0.35 / 0.45), the automation share (0.5 / 0.75 / 0.9) or reinstatement (0.5 / 0.25 / 0), which sit at scenario values from 2024, although the paper reads today's use as about half automation on Claude.ai and three-quarters on the API (pp. 27–28). The back-loading is set by the assumed logistic paths, not by data, and each capability path bridges a change of construct, from a usage anchor to endpoints on rated feasibility (p. 27).

p. 4 presents as "One surprising finding of our model" that the growth speedup from AI in innovation is relatively minor. Research input is a fixed share of GDP on both paths (Equation 22), so with a nonnegative ideas gap and λ ≤ 1 the 2030 ideas gap is at most g times the cumulated GDP gap: 0.639–0.661 against the simulated 0.605 in the extreme scenario, and 0.211–0.216 against 0.198 in the substantial. The bound holds whatever the ideas-block parameters.

The paper derives this smallness itself: pp. 16–17 say the gains "take decades to cumulate", and "bottlenecked by physical tasks" (p. 4) describes the same structure, in which research inherits the goods bottleneck. The printed effects (the 2030 ideas stock 0.61, 0.20 and 0.07 percent above its no-AI path in the extreme, substantial and modest scenarios, Table 3, p. 31) and the reading of them as "likely a lower bound" stand.

p. 4: "In each scenario, employment shifts from the cognitive occupations toward all other occupations, and cognitive wages grow more slowly than they would without AI while wages elsewhere grow faster." As specified, the modest cognitive wage is above its no-AI path in every month from Jan 2024 and 0.41 percent above it in 2030. Table 3 prints +0.4 (p. 31), Figure 3's cognitive-wage panel marks the modest scenario at 0.4 above the no-AI line (p. 33), and p. 32 says "the cognitive wage up 0.4 percent". The gap over the no-AI path rises in 59 of 61 twelve-month windows, the two exceptions starting in Jan and Feb 2024 at the paths' disclosed start (p. 40); from mid-2026 the modest cognitive wage grows 1.75 percent a year against 1.67 without AI.

The wage that would clear the cognitive group is also above the no-AI path from Jan 2025, so a fall in cognitive wages against the no-AI path, the comparison the overview on p. 7 states ("relative to the paths they would have taken without AI"), does not occur in the modest scenario. The conclusion states the narrower claim: "the cognitive wage falls below its no-AI path in the extreme scenario" (p. 38).

As specified, the substantial and extreme cognitive wages are below their no-AI paths in every month (−0.34 and −11.54 in 2030), and the other occupations' wage is above its no-AI path in every month of every scenario (+1.14 / +5.86 / +33.61). The substantial clause rests on where the replacements of labor-force exits search: once a quarter or more enter the growing group, the substantial cognitive wage ends 1.03 to 1.45 percent above its no-AI path. The sign also depends on wage rigidity: at flexible wages (ξ = 0) the modest cognitive wage is 0.24 percent below its no-AI path, and Table 6 prints positive cognitive wages in three of its seven columns. Read as the displacement channel, p. 3's statement that AI is "lowering demand for labor and wages in cognitive occupations" holds in every scenario, since the cognitive wage is below the common wage in every month.

pp. 34–35 say "Cognitive wages are 11.5 percent below the no-AI path" and, because no-AI wage growth is around 2 percent a year, that "cognitive wages are slightly lower in 2030 than in mid-2026". In the extreme scenario as specified, the fall from mid-2026 to Jan 2030 is 5.96 percent at the model's own no-AI wage growth of 1/60 a year, about 1.67 percent, and 4.85 percent at the sentence's 2 percent, which is the no-AI GDP growth rate (Table 1, p. 23).

The sign depends on where the replacements of cognitive labor-force exits search: the change is −2.95 percent with a quarter of them sent to the growing group, −0.53 with half, +3.19 with all, +4.13 under the employment-overhang cap and −5.58 under entry in proportion to target employment.

p. 39: "The scenarios therefore make statements about data we will observe over the coming years, and those data will tell us which scenario we are in." The paper runs no exercise showing which observables discriminate the scenarios, or by when. The four objects it names are measurable in principle, and capability and adoption are anchored on mid-2026 usage data. But each scenario bundles parameters with no anchor (reinstatement, the search discount on the path, posting speed), and the capability anchor (usage) and the 2030 endpoints (rated feasibility) measure different things.

The steady state matches its labor-market targets by construction, so no moment tests the fit. Inside the data window, the mid-2026 values the scenarios imply (GDP 0.55 / 0.83 / 1.21 percent above a no-AI path; cognitive unemployment 2.89 / 3.06 / 3.33) have no observed counterpart that isolates AI. By Jan 2028 the extreme scenario's gaps from the others are large (cognitive unemployment 6.07 against 2.91 and 3.38; GDP gaps of 4.0–5.6 points), while the modest–substantial gaps are small (GDP 1.6 points, cognitive unemployment 0.47). Telling the scenarios apart would require attributing those gaps to AI.

The same paragraph's "Today, none of the three can be ruled out" stands. "Potentially", in the preceding sentence, and the closing paragraph's "may become clearer within a year or two" hedge the neighbouring text, while this sentence says "will tell us".

p. 3 calls the modest scenario's 2030 GDP gain of 1.6 percent "only slightly higher than the estimate of Acemoglu (2025)" and "of the same order as the OECD figures", and says the substantial scenario's 8.3 percent "corresponds to the forecasts that circulated from financial and consulting institutions in 2023". p. 33 compares the substantial scenario's labor-share fall with "the four decades after 1980" in Karabarbounis and Neiman (2014).

• Acemoglu (2025). The modest gap of 1.61 is slightly above his upper, investment-boom range (1.4–1.56 percent over ten years), and its TFP gap of 0.71 slightly above his 0.66 bound. Against his baseline range (0.93–1.16) it is 1.4 to 1.7 times as large. His figures accrue over ten years and the scenario's over six from 2024, so GDP accrues 1.7 to 2.3 times, and TFP 1.8 times, as fast.
• The OECD figures. Every figure in Filippucci, Gal and Schief (2024), Filippucci, Gal, Laengle and Schief (2025) and the "up to 0.9 percentage points" of annual TFP growth in the abstract of Filippucci, Gal and Schief (2026) lies above the modest scenario, 1.3 to 7.6 times its pace. The scenario sits between Acemoglu and the lowest OECD figure: the same order of magnitude, outside the OECD range.
• Briggs and Kodnani (2023). The substantial scenario adds 1.45 points a year to output per worker from 2024, matching their figure of just under 1.5 points a year. Their ten-year window opens once about half of businesses have adopted, at a date they call hard to predict, so they do not forecast 8.3 percent by 2030. From the mid-2026 anchor the scenario's pace is 2.23 points a year, about 1.5 times theirs.
• The works said to give "comparable magnitudes". Aghion and Bunel (2024) are 0.55 to 0.90 times the substantial pace on labor productivity and 1.33 times on TFP. Baily, Brynjolfsson and Korinek's (2023) illustrative 18 percent over ten years is 1.15 times; the forecast they endorse, 7 percent of global GDP over ten years, is 0.47 times. McKinsey Global Institute's (2023) generative-AI contribution is 0.07 to 0.48 times, while its all-automation range brackets the scenario (0.34 to 2.34 times), and the paper does not say which it means. The Penn Wharton Budget Model (2025) brief cumulates to a 2030 GDP gain of 0.60 percent, 0.073 times the substantial scenario's 8.29 and below the modest scenario's 1.61.
• The extreme scenario. It "lies well beyond these" at 32.4 percent of GDP by 2030, at least four times every AI-specific figure cited on a 2030 horizon, but only 1.29 to 1.46 times the top of McKinsey's all-automation range.
• The labor share. "Four points in four years" rounds 3.43 points in 3.5 years (3.83 from the 2024 base, which is p. 33's "60 percent of income in 2026"). That is slightly less than the fall of roughly 4 to 5 points to 2011 in the US corporate labor share, read from the chart in Karabarbounis and Neiman's Figure 2. Their data end in 2012, about three decades after 1980, not four; they measure the corporate sector, not the aggregate share the model reports; and their text's only figure is global, 5 points over 35 years.

Footnote 3 states that AI raises the average wage "if and only if" ε exceeds the threshold of Equation 12, and that labor keeps the fraction (ε − ε*)/(ε + σ/s_L) of the gain a pegged rental rate would give it (p. 14). Equation 12 marks its threshold with "≈", so the "if and only if" holds for the exact threshold. The first-order threshold is 0.96 / 1.30 / 0.81 (extreme / substantial / modest, 2030); the exact threshold for the Proposition 1 wage is 2.10 / 1.76 / 0.97.

At the extreme parameters the first-order threshold places Table 5's ε = 1 column on the wrong side: the first-order wage rises 0.55 percent, the exact wage falls 14.3, and Table 5 prints −9.2. The share clause carries no "≈" and is off by more than half: 53 percent first order against 24 exact. The paper's sign statements (p. 14 at ε = 3; p. 37 at ε = 1) use exact solutions, and no printed table rests on the approximation.

• Capital supply. ε = 3 rests on half a long-run wealth elasticity applied over four years. It is the largest driver of extreme GDP, which spans 15.7–52.9 percent above the no-AI path over the paper's own parameter values (main-effect shares: ε 0.41, reinstatement 0.28). Table 5 discloses its effect, and "transformative" holds in every cell.
• Search parameters. Over the paper's own values with calibrated ε and ξ, extreme cognitive unemployment spans 4.7–19.5 (median 9.4); it reaches 17.5 in 3.7 percent of cells, and never when reinstatement is 0.5. One at a time, a search discount of 0.08 gives 14.3 and 0.17 gives 10.3, reinstatement of 0.5 gives 8.8, and flexible wages give 2.6 (Table 6); a slower posting speed raises it (18.5 and 19.5 at 0.25 and 0.10). The robustness sections vary only ε and ξ. The paper does not attribute 17.9 to capability alone, and a scenario statement defines its bundle.
• Within-task elasticity. An AI-performed and a human-performed instance of the same task substitute at σ = 0.5 by assumption (p. 8); the cited evidence concerns substitution across tasks or occupation groups. In a nested variant (extreme 2030, ε = 3, target economy) the labor share is 44.6 at a within-task elasticity of 0.5, 42.4 at 1 and 37.4 at 3, so a higher elasticity deepens the fall. Its effect on unemployment is not established.
• Worker-side responses. With flexible wages the cognitive wage is 42.2 percent below its no-AI path and the other wage 70.1 percent above, yet cognitive employment falls only 1.3 percent, because nothing in the model moves workers toward the higher wage. With illustrative responses (quits responding to the wage gap; directed search), the cognitive wages at ξ = 0 / 0.5 / 0.9 are −23.8 / −2.5 / +5.0 against the printed −42.2 / −11.5 / +2.8. The direction of the rigidity trade-off holds; the response strengths are not calibrated.

In Table 3's extreme column, the changes in labor income (+0.5 percent) and capital income (+81.4 percent), each weighted by its no-AI share of income, add to more than the change in GDP (+32.4 percent) by the cognitive firms' loss (p. 31). Labor income is measured at wages paid, while the labor-share row values cognitive labor at its marginal product; the share and GDP rows imply labor income of −0.2. The table does not say so. The cognitive-firm term that reconciles the rows is −0.31 percent of GDP in 2030 and negative in 901 of 982 non-zero months, with a largest size of 0.47, inside Table A.1's "at most half a percent of GDP" (p. 43), a bound the table states without a sign. No row is wrong by its own definition, and p. 35's "almost exactly what it would have been without AI" holds under either concept (+0.45 at wages paid, −0.23 at the marginal product).

• The quits elasticity. p. 27 reports an elasticity of 0.53 over 2001–19, "which we round to 0.55"; the regression specification is not printed. The code uses 6/11 = 0.5454, and the code's own note derives that value from the 0.06/0.11 occupational-move share rather than from the elasticity; the print does not say which rationale is meant. Set at 0.55, four extreme-column cells move by at most 0.06. Equation 27 applies this levels elasticity to last month's job-finding rate, while month to month the response in the cited JOLTS and CPS series (U.S. Bureau of Labor Statistics 2026a, 2026b) is 0.03–0.27, a mapping the print does not discuss; across the whole range 0 to 1, extreme cognitive unemployment stays within 17.6–18.2.
• No-AI ideas growth. The printed tables correspond to the code's no-AI ideas growth of 1/60 a year; at the printed 1.67 percent, three cells differ by up to 0.06.
• The ψ logistic. ψ is "held constant" (p. 11) and constant in the code; Table A.2's logistic for ψ is unused.
• The solved logistic slope. The code caps it at 3 a year and floors it at 0, which the print does not state; the cap does not bind in the scenarios.
• Table 4's net-return interquartile range. It prints as [7, 7]: at integer rounding no dispersion shows (the capital stock's is [6, 27]).
• The cognitive employment share. OEWS May 2025 gives 0.560 on footnote 10's grouping (U.S. Bureau of Labor Statistics 2026c); the model uses 0.6235 from CPS 2025 and carries an OEWS May 2021 option of 0.614. The effect of the gap on the scenarios is not established.