Karger, Kuusela et al. 2026, "Forecasting the Economic Effects of AI" (working paper)

The paper’s principal disagreement result concerns dispersion conditional on its broad AI scenarios. In the stated fitted hierarchy, the within-scenario, between-forecaster component exceeds the between-scenario, between-forecaster component in every printed row. That conditional remainder generally differs from the variance of unconditional forecast means. In the paper’s two-forecaster example, these quantities are 0.835216 and 0.6724 respectively; an exact hypothetical example can have equal unconditional means and a positive conditional remainder.

The authors expressly acknowledge that the scenario descriptions leave different capability profiles and interpretations within each label. Consequently, substantial disagreement conditional on those descriptions does not uniquely allocate ordinary unconditional disagreement to economic mechanisms rather than capability beliefs. This limits the primary-source interpretation, capability/effects separation and economic-absorption explanation; it leaves the conditional statistic and the joint elicitation of forecasts and preferences intact. Jointly eliciting forecasts and policy support does not identify a causal link from expectations to preferences. The contribution of scenario ambiguity has no measured numerical bound.

Cunningham (2025) offers a hedged capability-progress explanation for a heterogeneous compilation of forecasts. The recruited survey and its conditional dispersion do not directly test that entire compilation or establish an exhaustive attribution. Most displayed within-group interquartile ranges exceed the range of group medians, with one exception: rapid-scenario GDP in 2050 has a 1.5-percentage-point economist interquartile range and a 1.8-point cross-group median range. That descriptive comparison supplies no general equivalence bound for the groups.

The slow-versus-rapid 2050 economist labor-force-participation medians are 59.2% and 55.0%. At the stipulated common population of 270 million, the 4.2-point contrast represents 11.34 million fewer labor-force participants, rounded in the paper to roughly ten million. Labor-force participation includes unemployed people actively seeking work; it is not a count of jobs.

Footnote 9 openly states the baseline and population assumptions and disclaims clean causal identification. The conditional comparison does not identify how many jobs AI eliminates or how much of the decline is caused by AI. The full rapid-scenario abstract sentence therefore needs that limit, while its central forecast description remains. Alternative baseline or population assumptions have no measured correction here, and the illustration supplies no conservative bound on AI-caused job loss.

The economist mean assigns 47.4% to moderate progress and 14.0% to rapid progress: 61.4% combined. A majority of mean probability mass is a different quantity from a majority of individuals predicting significant progress. Neither that mean nor a weighted median establishes the respondent headcount claimed in the headline.

The individual-majority assertion is not established here because the necessary count and linked respondent records are unavailable. This support limit does not imply that the opposite majority is true.

The published rationale examples contain attenuating, accelerating and catastrophic possibilities. Appendix I selects rationales for their quality, rather than representativeness or predictive accuracy, and explicitly uses answers from before coherence revision. The authors describe manual checking, adjustments and a second tagging agent.

The presence of these themes does not establish that a majority of experts agrees on net attenuation. The complete denominator, linked labels and final-probability linkage are unavailable here. The majority frequency is therefore not established here; no contrary majority or coding-error conclusion follows. The claimed comparison among economists with similar rapid-scenario probabilities likewise narrows to illustrated differences in views about diffusion, job creation, productivity lags and institutions.

The 61.4% economist mean describes a future expert panel’s modal scenario selection, the event specified in the questionnaire. Its equivalence to literal capability attainment is unmeasured; it is not an adoption probability or a calibrated event probability.

The displayed comparison pools industry and policy researchers, so it cannot separately substantiate the description of all five groups. The 14.0% rapid probability belongs to the economist mean, compared with 16.1% for the pooled AI-expert mean and 10.0% for the economist median. Wording about experts collectively needs those distinctions. The paper explicitly limits broader population interpretation of the recruited groups; the probabilities supply no measured action threshold.

The economist unconditional GDP medians are 2.5% at both horizons. Figure 4 instead gives 3.0% for AI experts in 2050 and 2.4% and 2.8% for the public in 2030 and 2050. Table 20’s public 2050 value of 3.0% uses the disclosed horizon-specific sample. The abstract’s exact 2.5% wording for every group and horizon needs an approximate, near-2.5% formulation.

These are reported belief summaries, rather than measured future effects. The discrepancy does not establish an incorrect native median, a demographic causal effect or reversal of the selected external-baseline ordering; independent reconstruction of the baseline vintages remains unavailable.

Many central rapid-scenario forecasts have selected historical parallels. The categorical claim that experts forecast no outcomes outside historical experience is broader than that evidence, including its adjoining near-term qualification. The authors disclose above-cap GDP views, a per-capita postwar exception, broad tails and possible missing extremes.

The historical and external GDP comparisons also use different outcome definitions and horizon proxies. They provide context rather than identical-parameter benchmarks. The GDP-level projection uses an expressly illustrative interpolated growth path; it does not establish every intervening year’s growth or an exhaustive historical bound on all forecasts.

Economists above the median rapid-scenario probability have higher unconditional GDP medians at both horizons. Their unconditional labor-force-participation medians fall from 61.4% to 60.4% in the 2030 comparison, but rise from 58.0% to 60.0% in 2050. Rapid-scenario participation is also higher in the above-median group at both horizons (Tables 50–53).

The lower-participation description therefore applies only to the unconditional 2030 comparison. Overlapping interquartile ranges describe heterogeneity; they establish neither equivalent medians nor statistical indistinguishability.

Weighting the same displayed economist sample moves rapid-scenario 2050 participation from 58% to 55%, unconditional participation from 60% to 58.3%, and rapid GDP growth from 4.0% to 3.5%. The authors disclose moderate-to-large participation changes. Blanket minimal-effects wording needs that qualification.

Raking adjusts experience and continent margins toward the invitation frame. Full representation of invitees’ beliefs additionally requires suitable response assumptions within those margins. Table 58 compares the unweighted sample with the frame; it does not show residual imbalance after weighting, and no bias direction or harmful weighting effect is established.

The displayed GDP and participation directions persist through combined revision and filtering, but experts alone had revision contact. Only unresolved Flag 2 excludes main outcome-year records; the other flags provide feedback. The combined comparisons do not isolate revision from exclusion, demonstrate accuracy or establish stability across every outcome family and decomposition. Balanced two-horizon plots and horizon-specific tables intentionally describe different samples.

Across questions and time horizons, Table 8 (pp. 69–70) gives ten economist fractions of incoherent fitted distributions ranging from 5.7% to 15.6%; the prose on p. 38 instead states an 11%–21% range. The table describes a smaller exclusion range. Public 2050 participation has a 41.8% exclusion fraction under the table’s convention, leaving 58.2%.

These negative-implied-variance exclusions are separate from the main coherence flag. They define a retained fitted-sample target and may reflect fit misspecification, as the authors acknowledge. The exact denominator and whether the reported fractions are weighted remain unavailable here. The fractions do not identify respondent counts, omitted weight mass or effects on component shares; their native effects remain unverified here, without an inferred selection direction.

Within-person components total 60.2% for economist 2050 labor-force participation, 37.6% for AI experts and 48.1% for all experts. In the latter two groups they do not dominate conditional between-forecaster variation. The statement that within-forecaster uncertainty dwarfs disagreement therefore needs an outcome and group qualification.

The distinct ordering of within-scenario versus between-scenario between-forecaster components concerns a different comparison. Raw quantiles and rationales describe substantial subjective uncertainty, but no error model separates substantive uncertainty from elicitation or fitting error, and no arbitrary component subtraction is warranted.

Economist 2050 forecasts put the labor compensation share at 50% unconditionally and 45% under rapid progress, while the top-decile household wealth share is 75% and 80%. Labor share is compensation as a fraction of nonfarm-business output. Its decline alone does not establish greater household inequality without information on capital ownership, labor-income dispersion and transfers.

The paper distinguishes income from wealth and allows rising median household income alongside greater concentration. The supported description concerns two different distributional indicators, rather than an identified household-welfare effect.

The current-study economist summaries differ between Tables 1, 11 and 18: the slow median is 39.9% versus 38.3%, slow first quartile 17.4% versus 16.8%, moderate third quartile 62.6% versus 64.5%, and rapid first quartile 7.3% versus 6.7%.

Unlabelled estimator, interpolation, weighting or snapshot differences could explain these cells. They constitute one unresolved reconciliation. The available summaries select neither a correct raw aggregate nor a coding defect; public rapid progress is 19.0% in both comparison locations.

The results and detailed job-guarantee description specify financing through a 0.5% VAT, but the printed survey transcription ends without that sentence. The general policy-package target remains, while the actual delivered wording and any respondent effect remain unverified here. The gap does not establish that participants received an unfunded policy or that a support or forecast figure is wrong.

The general methods name both horizons and the assistance table includes 2050, whereas the printed assistance item names only 2030. Intended 2050 elicitation is plausible; the exact deployed branching is unavailable. This is a documentation qualification, without a finding of missing elicitation or invented responses.

Equation 3 uses S where the local derivation defines R. The intended combined identity requires R in that position. This notation defect supplies no corrected component share or evidence that the authors implemented a different joint weighting law.