CoChat
Sign Up
Cited by 1,193Labor market dynamics and wage inequalityDigital Economy and Work TransformationImpact of AI and Big Data on Business and Society

Artificial Intelligence and Jobs: Evidence from Online Vacancies

Open access
Acemoglu, Daron, Autor, David, Hazell, Jonathon, Restrepo, Pascual·2022·Journal of Labor Economics
DOI 10.1086/718327
Open this paper in CoChat — free account

Abstract

We study the impact of artificial intelligence (AI) on labor markets using establishment-level data on the near universe of online vacancies in the United States from 2010 onward. There is rapid growth in AI-related vacancies over 2010–18 that is driven by establishments whose workers engage in tasks compatible with AI’s current capabilities. As these AI-exposed establishments adopt AI, they simultaneously reduce hiring in non-AI positions and change the skill requirements of remaining postings. While visible at the establishment level, the aggregate impacts of AI-labor substitution on employment and wage growth in more exposed occupations and industries is currently too small to be detectable.

AI snapshot

Written from the paper’s full text.

TL;DR

Using near-universe US online vacancies from Burning Glass, the paper measures establishment AI exposure from baseline occupational task structure and finds a post-2015 surge in AI vacancies, skill churn, and reduced non-AI hiring at AI-exposed establishments. It finds no detectable relationship between AI exposure and employment or wage growth at the industry or occupation level.

Key findings

  • AI vacancy postings grew rapidly after 2015, concentrated in AI-exposed establishments. A one standard deviation increase in Felten et al. AI exposure is associated with 15.96 (SE = 1.73) in the baseline AI-vacancy-growth specification, or 9.19 (SE = 1.21) with three-digit industry fixed effects; the paper describes this as roughly 15–16% more AI vacancy posting.
  • AI exposure predicts both disappearance of previously posted skills and emergence of new skills. With Felten et al., one standard deviation in AI exposure is associated with a 0.83 absolute decline in per-vacancy frequency of previously demanded skills (SE = 0.09; sample mean negative skill change = 4.70) and a 0.95 absolute increase in new or increasingly demanded skills (SE = 0.08; sample mean positive skill change = 6.30).
  • AI-exposed establishments reduce non-AI hiring. With Felten et al., the baseline estimate is −13.80 (SE = 4.22), about a 14% decline in non-AI vacancies; with firm fixed effects it is −4.81 (SE = 1.44), and for 2014–2018 it is −11.94 (SE = 3.80). The implied 2010–2018 non-AI employment decline is 7.2% (coefficient −7.24, SE = 4.66). Webb results are broadly similar but less stable; SML shows no consistent negative relationship.
  • No detectable aggregate relationship: industry-by-commuting-zone employment estimate for 2010–2016 is −0.05 (SE = 0.08), and occupation-level employment and wage regressions also show no differential behavior for more AI-exposed occupations.

Methods

  • Data: Burning Glass online vacancy postings covering the near-universe of US online vacancies from 2010 onward, with 2007 data also used; establishment defined as firm-by-commuting-zone; occupations and skills coded from postings; establishments in NAICS 51 (information) and 54 (professional/business services) excluded.
  • AI exposure: baseline 2010–2012 occupational structure weighted by three occupation-level AI suitability indices—Felten, Raj, and Seamans; Webb; and Brynjolfsson, Mitchell, and Rock’s SML—then standardized.
  • Design: establishment-level regressions of changes from 2010–2012 to 2016–2018 on baseline AI exposure, weighted by baseline vacancies, with controls for firm size decile, commuting zone, three-digit industry, and sometimes firm fixed effects; outcomes include AI vacancies, negative/positive skill change, non-AI vacancies, and industry/occupation employment and wages.

Limitations

  • The paper states that it leaves open the juxtaposition of negative establishment-level impacts and zero aggregate effects; that its task-structure focus may exclude AI impacts from new products or services; that estimates are not informative about AI applications missed by its exposure measures; and that future AI technologies may be more complementary or substitutive.
  • The text makes evident that results are much less robust with the SML measure and less stable with Webb than with Felten et al.; that the AI exposure measure is a shift-share design and standard errors may not be the most conservative; that County Business Patterns analysis ends in 2016, excluding later rapid AI expansion; and that online vacancy postings overrepresent technical and professional jobs relative to blue-collar and personal service jobs.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1086/718327

Cites

29 of 29 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 38 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    There is a surge in AI activity, particularly after 2015, proxied by vacancies seeking workers with AI skills.

    “We see a surge in AI activity, particularly after 2015, proxied by vacancies seeking workers with AI skills”
    Quote in the paperSubstantial claim
  • Findingown data

    Adding three-digit NAICS industry fixed effects (column 3) reduces the Felten et al. AI exposure coefficient for AI vacancies to 9.19 (SE = 1.21).

    “Column 3 additionally adds three-digit (NAICS) industry fixed effects.Reflecting the sizable variation in AI exposure across industries shown in Figure 3 above, these controls are more important for our regressions, and they reduce the magnitude of our estimate by about a third, to 9.19, but the standard error of the estimate also declines (to 1.21).”
    Quote in the paperNot a substantial claim
  • Findingown data

    The relationship between AI exposure and non-AI vacancies remains comparable when including baseline shares of sales and administration occupations in columns 5 and 6: -12.42 (SE = 4.01) and -4.04 (SE = 1.47).

    “The relationship between AI exposure and non-AI vacancies remains comparable when we include the baseline shares of sales and administration occupations in columns 5 and 6: -12.42 (se = 4.01) and -4.04 (SE = 1.47), respectively.”
    Quote in the paperNot a substantial claim
  • Findingnone given

    The relationship between AI exposure and AI hiring is much less robust and stable with the SML measure.

    “the relationship between AI exposure and AI hiring is much less robust and stable with SML as well.”
    Quote in the paperSubstantial claim
  • Findingreasoning

    A one standard deviation increase in Felten et al. AI exposure is associated with approximately a 16% increase in AI vacancies.

    “This estimate implies that a one standard deviation increase in AI exposure-which corresponds to the difference between finance and mining and all extraction-is associated with approximately a 16% increase in AI vacancies.”
    Quote in the paperSubstantial claim
  • May combine two claims. 49
  • Classified as probably not a substantial claim. 20
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingown data

    Establishments with task structures suitable for AI cease to post vacancies listing a range of previously-sought skills and start posting additional skill requirements.

    “establishments with task structures suitable for AI cease to post vacancies that list a range of previously-sought skills and start posting additional skill requirements.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    There is no evidence for major human-AI complementarities in AI-exposed establishments or for the expectation that AI will increase hiring because of large productivity effects.

    “We find no evidence for either the view that there are major human-AI complementarities in these establishments or the expectation that AI will increase hiring because of its large productivity effects-though we cannot rule out that other applications of AI that are not captured here could have such effects.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The SML measure is high for office and administrative occupations and for sales occupations, above average for personal services, and low for professional occupations and most blue-collar and service occupations.

    “In contrast, the SML measure is high for office and administrative occupations, and for sales occupations, and (perhaps surprisingly) above average for personal services, but is low for professional occupations and most bluecollar and service occupations.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The estimate in column 3 of Table 8 is -0.05 (SE = 0.08), implying a one standard deviation increase in industry AI predicts an economically small and statistically insignificant 0.049% decline in industry employment.

    “For example, the estimate in column 3 is -0.05 (se = 0.08).The point estimate implies very small effects associated with industry AI exposure: a one standard deviation increase in industry AI predicts an economically small and statistically insignificant 0.049% decline in industry employment.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The partial R2 associated with the Webb measure is 0.0074, roughly one quarter of the corresponding R2 for the Felten et al. measure.

    “However, from Appendix Table A3, the partial R 2 associated with the Webb measure is 0.0074, roughly one quarter of the corresponding R 2 for the Felten et al measure.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Estimates using the Webb measure are similar to Felten et al. in the first four specifications, though not fully robust to controls for baseline shares of sales and administration vacancies in columns 5 and 6.

    “Estimates using the Webb measure, reported in Panel B, are similar to those in Panel A in the first four specifications, though they are not fully robust to controls for the baseline shares of sales and administration vacancies in columns 5 and 6.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Limiting the sample to establishments that posted in 2016-2018, the estimates are somewhat smaller but still negative and statistically significant at 5%: -8.38 (SE = 3.46) in column 7 and -3.56 (SE = 1.86) in column 8.

    “The estimates are now somewhat smaller, but still negative and statistically significant at 5%: -8.38 (se = 3.46) in column 7, with three-digit industry fixed effects, and -3.56 (SE = 1.86) in column 8, with firm fixed effects.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The Webb measure is not particularly high in sales occupations and shows a strong positive relationship with occupational wage percentiles.

    “The Webb measure is not particularly high in sales occupations and shows a strong positive relationship with occupational wage percentiles.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    With two of the three measures, AI-exposed establishments reduce their non-AI and overall hiring.

    “With two of our three measures, we find that AI-exposed establishments reduce their non-AI and overall hiring.”
    • Also stated in: § 8 Conclusion · hedged
      “AI-exposed establishments appear to be reducing their non-AI hiring”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Results with the SML index are broadly similar to Felten et al. but significantly weaker.

    “Results with the SML index in Panel C are broadly similar, but significantly weaker.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In column 4 with firm fixed effects, the AI exposure point estimate for non-AI vacancies is -4.81, about half the magnitude in the preceding column, and the standard error falls from 4.08 to 1.44.

    “In this specification, the point estimate for AI exposure is -4.81, which is about half of the magnitude in the preceding column.Simultaneously, the estimates become more precise as the standard error falls from 4.08 to 1.44.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Firm-level estimates are generally imprecise and inconsistently signed, but models for firm-level mean of establishment AI vacancy growth or share of AI vacancies are very similar to main results in Table 1.

    “As shown in columns 1 and 2 of Table A7, these estimates are generally imprecise and inconsistently signed.However, when we estimate models for the firm-level mean of establishment AI vacancy growth (columns 3 and 4) or models with share of AI vacancies (columns 5 and 6), the estimates are very similar to our main results in Table 1.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    There is a positive association between the SML-based measure of AI exposure and AI vacancy growth without covariates, but when three-digit industry fixed effects are included, this relationship becomes negative.

    “There is a positive association between the SML-based measure of AI exposure and AI vacancy growth without any covariates, but when three-digit industry fixed effects are included, this relationship becomes negative.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Replacing narrow AI vacancy measures with broad AI vacancy indices, and using change in share of AI vacancies as dependent variable, corroborates main findings and is if anything stronger and more stable, especially with the Webb measure.

    “The results in both tables corroborate our main findings in Table 1 and, if anything, are stronger and more stable, especially with the Webb measure.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The top two quartiles of establishment AI exposure post significantly more AI vacancies and drive the surge in AI vacancies after 2015.

    “It shows that the top two quartiles post significantly more AI vacancies and drive the surge in AI vacancies after 2015.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The Felten et al. measure has the most stable and largest quantitative relationship to skill changes, followed by Webb and SML.

    “Paralleling our findings at many points in the paper, the Felten et al. measure proves to have the most stable and largest quantitative relationship to the outcome variable, followed by Webb and SML.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The 2014-18 result is robust to firm fixed effects added in column 8, though their inclusion reduces the magnitude of the relationship.

    “This result is robust to firm fixed effects, added in column 8, though, as before, their inclusions reduces the magnitude of the relationship.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    The Felten et al. measure is particularly high for managers, professionals, and office and administrative staff, and very low for service, production, and construction workers.

    “The Felten et al. measure, for example, is particularly high for managers, professionals and office and administrative staff and is very low for service, production and construction workers”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Controlling for baseline shares of sales and administration vacancies (columns 5 and 6) does not substantially change the Felten et al. AI exposure coefficient for AI vacancies, which remain 9.75 (SE = 1.20) and 16.87 (SE = 1.86).

    “These controls do not substantially change the estimate in either column, which remain, respectively, at 9.75 (se = 1.20) and 16.87 (se = 1.86).”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Findingown data

    In the most loaded specification (column 6) with software exposure control, the Felten et al. AI exposure point estimate for AI vacancies is 17.47 (SE = 1.90), compared to 16.87 (SE = 1.86) in Table 1.

    “For example, in the most loaded specification (column 6), the point estimate is now 17.47 with a standard error of 1.90, as compared to 16.87 and a standard error of 1.86 in Table 1.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    AI-related vacancy postings grew rapidly over 2010-2018, with acceleration around 2015-2016.

    “Our first result is that there is a rapid takeoff in AI vacancy postings starting in 2010 and significantly accelerating around 2015-2016.”
    • Also stated in: § 3.1 Burning Glass Data · boosted
      “Figure 2 shows the evolution of postings of narrow and broad AI vacancies in the BG data, highlighting the rapid takeoff of AI vacancies after 2015, as noted in the Introduction.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    There is no economically or statistically significant relationship between AI exposure in 2010 and non-AI hiring during 2010-14, while there is a strong negative association for 2014-18.

    “Focusing on Felten et al.'s AI exposure measure in Panel A, we see no economically or statistically significant relationship between AI exposure in 2010 and non-AI hiring during the years 2010-14, while there is a strong negative association for 2014-18, during the period of rapid AI takeoff.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The relationship between AI exposure and AI vacancy postings is fairly close to linear across the distribution and does not appear to be driven by outliers.

    “The relationship between AI exposure and AI vacancy postings is fairly close to linear across the distribution and does not appear to be driven by outliers.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Industry AI exposure in 2010 does not predict differential employment behavior before 2010; thus, three-digit industries with different AI exposure were on roughly parallel trends before the pickup in AI activity in the late 2010s.

    “These columns show that industry AI exposure in 2010 does not predict differential employment behavior before 2010; thus, threedigit industries with different AI exposure were on roughly parallel trends before the pickup in AI activity in the late 2010s.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Controlling for Webb's software exposure, the software exposure measure itself has no consistent association with non-AI hiring, while the effects of AI exposure remain very similar to baseline estimates in Table 5.

    “The results reported in Table 6 document that the software exposure measure itself has no consistent association with non-AI hiring, while the effects of AI exposure remain very similar to our baseline estimates in Table 5.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    In Table 6 column 1, the estimate using Felten et al.'s measure is -14.62, compared to -13.80 for the same specification in Table 5.

    “For example, the estimate using Felten et al.'s measure in column 1 is -14.62, compared to -13.80 for the same specification in Table 5.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The 2014-18 estimate remains stable and becomes more precise when controlling for firm size deciles, commuting zone controls, and three-digit industry fixed effects in column 6.

    “This estimate remains stable and becomes more precise when we control for firm size deciles, commuting zone controls, and three-digit industry fixed effects in column 6.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). The quote only partly states this claim.

    Findingown data

    Panels B and C of Table 8 show similar null results using Webb and SML measures.

    “Panels B and C of the table show similar results using the Webb and SML measures in place of the Felten et al. index.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The strong association between AI exposure and subsequent AI activity is robust to numerous controls and specification checks when using the Felten et al. and Webb measures, but is less apparent with the SML index.

    “The strong association between AI exposure and subsequent AI activity is robust to numerous controls and specification checks when using the Felten et al. and the Webb measures, but this is less apparent with the SML index.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Firm AI exposure in 2010 is not significantly associated with firm AI vacancy growth between 2010 and 2018, using the Felten et al. (2019) measure (coefficient -3.77, SE 4.81, column 1 of Table A7).

    “Firm AI Exposure, 2010 -3.77 -0.4913.60 8.01 0.29 0.27 (4.81) (4.13) (1.85) (1.49) (0.03) (0.05)”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Felten et al. (2019)

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Findingown data

    The relationship between mean wage of an occupation and the three AI exposure measures is distinct, which is the basis for the claim that each index captures a different aspect of AI exposure.

    “Figure4below shows that the relationship between mean wage of an occupation and the three AI exposure measures is distinct, which is the basis of our claim that each one of these indices captures a different aspect of AI exposure.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Assuming that 1.6% of the increase in AI vacancies were driven by task-level substitution of algorithms for labor, this type of AI adoption led to a decline of 1% in non-AI hiring in finance and 0.92% in manufacturing.

    “Assuming that 1.6% of these increases-the partial R 2 of our AI exposure measure in Table 1-were driven by task-level substitution of algorithms for labor, one may infer that this type of AI adoption led to a decline of 1% in non-AI hiring in finance and 0.92% in manufacturing.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: 1.6% of the increase in AI vacancies were driven by task-level substitution of algorithms for labor.

  • May combine two claims.

    Interpretationown data

    An important component of AI activity is linked to the types of tasks performed in an establishment, though AI activity may have other drivers such as development of new products or business models.

    “This evidence suggests that an important component of AI activity is linked to the types of tasks performed in an establishment-though it does not preclude the possibility that AI activity has other drivers, such as the development of new products or business models.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The paper's evidence finds negative effects of AI exposure on non-AI hiring, suggesting the current generation of AI technologies is predominantly task-replacing and generates only modest productivity gains.

    “Our evidence below finds negative effects of AI exposure on (non-AI) hiring, and thus suggests that the current generation of AI technologies is predominantly task-replacing and generates only modest productivity gains.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductown data

    All three AI exposure measures are robust to inclusion of firm size deciles, commuting zone dummies, and controls for initial establishment vacancy structures in sales and administrative occupations for skill changes.

    “In particular, all measures prove robust to the inclusion of firm size deciles, commuting zone dummies, and controls for initial establishment vacancy structures in sales and administrative occupations.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductown data

    The results on lower hiring are robust in all specifications using the Felten et al. measure and in most specifications with the Webb measure, but not with SML.

    “These results are robust in all of our specifications using the Felten et al. measure and in most specifications with the Webb measure, but, as anticipated, not with SML.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    In all columns, the analysis uses baseline employment as weights.

    “and use baseline employment as weights”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Evidence of negative or zero effects of AI exposure on hiring weighs against both the human-complementary view and the possibility of large productivity gains from AI.

    “Conversely, evidence of negative, or even zero, effects would weigh against both the human-complementary view and the possibility of large productivity gains from AI-since both AI-human complementarity and large productivity effects boosting employment in non-automated tasks could generate a positive relationship between AI exposure and establishments hiring.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductown data

    The negative hiring results are statistically significant, economically sizable, and robust with the Felten et al. measure, and robust in most specifications with the Webb measure.

    “These results are statistically significant, economically sizable, and robust with the Felten et al. measure, and robust in most specifications with the Webb measure.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Felten et al.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Webb

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Interpretationown data

    AI is currently substituting for humans in a subset of tasks but is not yet having detectable aggregate labor market consequences.

    “implying that AI is currently substituting for humans in a subset of tasks but it is not yet having detectable aggregate labor market consequences.”
    • Also stated in: § 1 Introduction · plain
      “We conclude that despite the notable surge in AI adoption, the impact of this new technology is still too small relative to the scale of the US labor market to have had first-order impacts on employment patterns outside of AI hiring itself.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    In each regression, observations are weighted by total establishment vacancies in 2010-2012.

    “In each regression, observations are weighted by total establishment vacancies in 2010-2012.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    While AI technologies appear to be changing task and skill composition at exposed establishments and firms, any aggregate effects of AI are too small to detect.

    “Our overall interpretation is that while AI technologies appear to be changing task and skill composition at exposed establishments and firms, any aggregate effects of AI are too small to detect.28”
    • Also stated in: § 8 Conclusion · hedged
      “However, while AI technologies appear to be changing task and skill composition at exposed establishments, any aggregate effects of AI, if present, are not yet detectable-plausibly because AI technologies are still in their infancy and have spread only to a limited part of the US economy.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductown data

    The pattern of surge driven by high-exposure establishments is highly robust with the Felten et al. and Webb AI-exposure measures, and still present but less robust with the SML measure.

    “This pattern is highly robust with two of our three AI-exposure measures-those based on the indices constructed by Felten et al. and Webb-and still present but less robust with our third measure, SML, based on Brynjolfsson, Mitchell and Rock's work.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites Felten et al.

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Webb

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Brynjolfsson, Mitchell and Rock

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Interpretationreasoning

    The AI exposure variable identifies meaningful variation in the suitability of establishment task structure for AI, and this variation is distinct from exposure to traditional software and digital technologies.

    “This set of estimates bolsters our confidence that the AI exposure variable identifies meaningful variation in the suitability of establishment task structure for AI, and that this variation is distinct from exposure to traditional software and digital technologies.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    AI adoption has real effects on the task content of non-AI jobs-enabling firms to replace some tasks previously performed by workers, making certain skills redundant, while simultaneously generating demand for new skills.

    “The finding that AI exposure is associated with significant changes in the skills listed in vacancies bolsters our confidence that AI adoption has real effects on the task content of non-AI jobs-enabling firms to replace some of the tasks previously performed by workers, making certain skills redundant, while simultaneously generating demand for new skills.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: AI is intended to supplement, replicate, and in some cases, exceed human-level intelligence in a variety of tasks.

  • May combine two claims.

    Interpretationreasoning

    If the productivity effect (ε_e•ρ_e-1)•π_e exceeds -1, AI technologies increase hiring; otherwise AI advances reduce non-AI hiring in exposed establishments.

    “If the second term on the right-hand side of ( 5), (ε e • ρ e -1) • π e , exceeds -1, the productivity effect dominates and AI technologies increase hiring.7 Otherwise, AI advances will reduce (non-AI) hiring in exposed establishments.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The elasticity of substitution between labor and algorithms is infinite (σ=∞) in Propositions 1 and 2.
    • The authors state this as an assumption: The initial cost share of AI is small.
    • The authors state this as an assumption: AI services are produced by combining AI capital and in-house AI workers in fixed proportions.
    • The authors state this as an assumption: All establishments are price takers for production workers, AI workers, and AI capital.
    • The authors state this as an assumption: Each establishment faces a downward-sloping demand curve for its product and sets its price to maximize profits.
    • The authors state this as an assumption: Each establishment is small in the labor market and takes other prices and aggregate output as given.
    • The authors state this as an assumption: The importance or quality of each task in production, α(x), is common across establishments.
    • The authors state this as an assumption: The task-specific labor and AI productivity parameters γℓ(x) and γa(x) are common across establishments.
    • The authors state this as an assumption: All establishments have constant returns to scale.

  • May combine two claims.

    Study design and conductown data

    Over 2010-2018, the occupational and industry composition in BG is closely aligned with overall occupation employment shares from OES and industry vacancy shares from JOLTS.

    “Appendix Figure A1 shows that over the 2010-2018 period, the occupational and industry composition in BG is closely aligned with both overall occupation employment shares from the OES and with industry vacancy shares from JOLTS.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductreasoning

    The estimates in footnote 24 should be interpreted with caution because they ignore general equilibrium effects and spillovers, and the partial R2 may over or understate the role of task substitution in AI activity.

    “These estimates should be interpreted with caution, since they ignore general equilibrium effects and spillovers, and the partial R 2 may over or understate the role of task substitution in AI activity.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    AI adoption will induce churn of tasks performed by workers as some tasks previously performed by humans are taken over by algorithms.

    “These results are also consonant with our theoretical model in Section 2, which suggests that AI adoption will induce churn of tasks performed by workers as some tasks previously performed by humans are taken over by algorithms.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: AI is intended to supplement, replicate, and in some cases, exceed human-level intelligence in a variety of tasks.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The standard errors are robust against heteroscedasticity.

    “The standard errors are robust against heteroscedasticity.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Standard errors are clustered by firm.

    “Standard errors are clustered by firm.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The evidence bolsters the case that AI is altering the task structure of jobs, replacing some human-performed tasks while simultaneously generating new tasks accompanied by new skill demands.

    “This evidence bolsters the case that AI is altering the task structure of jobs, replacing some human-performed tasks while simultaneously generating new tasks accompanied by new skill demands.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    In all columns, the analysis controls for three-digit occupation fixed effects.

    “In all columns, we control for three-digit occupation fixed effects”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The coefficient estimates relative change in non-AI hiring at more versus less AI-exposed establishments, and do not imply an aggregate reduction in total hiring.

    “We also note that because this coefficient estimates the relative change in non-AI hiring at more versus less AI-exposed establishments, it does not imply an aggregate reduction in total hiring.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductown data

    The establishment-level regressions include fixed effects for firm size decile, commuting zone, 3-digit NAICS industry, and firm, to adjust for confounders.

    “Columns 2-3 and 5 include fixed effects for the decile of firm size (defined as the total vacancies posted by an establishment's firm in 2010-2012).Columns 2-6 include commuting zone fixed effects.Columns 3 and 5 include 3-digit NAICS industry fixed effects.Columns 4 and 6 include firm fixed effects.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Allowing establishments to substitute outsourced AI workers for in-house services would imply that the proxy for AI activity at the establishment level may understate the extent of AI, potentially leading to attenuation of the estimates.

    “the second would imply that our proxy for AI activity at the establishment level may understate the extent of AI, potentially leading to attenuation of our estimates.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    When σ = ∞ and the share of AI algorithms in cost is initially small, exposure to AI equals the share of tasks that can now be completed with AI in total costs.

    “When σ = ∞ as in Propositions 1 and 2 below and the share of AI algorithms in cost is initially small, exposure to AI is x∈Te∩T A α(x)dx, which gives the share of tasks that can now be completed with AI in total costs.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The share of AI algorithms in cost is initially small.

  • May combine two claims.

    Premisereasoning

    BG data represent vacancy flows, while the OES reports employment stocks, so the two data sources are not expected to align perfectly.

    “We note that BG data represent vacancy flows, while the OES reports employment stocks, and thus we do not expect the two data sources to align perfectly.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductcitation

    Standard errors in the establishment AI vacancy growth regressions are from Borusyak et al. (2021), estimated from a weighted regression at the 6-digit occupation level with weights depending on baseline establishment size.

    “Standard errors are from Borusyak et al (2021) and come from a weighted regression estimated at the 6 digit occupation level, with weights that depend on baseline establishment size.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Borusyak et al (2021)

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    Productivity gains, if present, have little effect on demand, potentially due to low passthrough rates.

    “Or that productivity gains, if present, have little effect on demand, potentially due to low passthrough rates.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    Allowing the technology to be more general than Leontief would not have any major effect on the results.

    “The first modification would not have any major effect on our results”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    When the passthrough rate is less than 1/ε_e, the establishment's price increases sufficiently that output does not expand and thus employment always declines.

    “This expression also clarifies that when the passthrough rate is less than 1/ε e , the establishment's price increases sufficiently that output does not expand and thus employment always declines.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The passthrough rate is less than 1/ε_e.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    The adjusted R2 is similar in the 10% sample and the full data when the regressors are the measures of AI exposure.

    “We have verified that the adjusted R 2 is similar in the sample and the full data when the regressors are our measures of AI exposure.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    The inverse hyperbolic sine transformation is well defined when x = 0, which is a frequent occurrence in the sample of establishments.

    “For small values of x, this approximates a proportional change, but is well defined when x = 0, which is a frequent occurrence in our sample of establishments.”
    Quote in the paperNot a substantial claim
64 of 64

  • The authors state this as an assumption.

    The elasticity of substitution between labor and algorithms is infinite (σ=∞) in Propositions 1 and 2.

    “To illustrate how the task structure determines AI adoption, we follow Acemoglu and Autor (2011) and Acemoglu andRestrepo (2018, 2019a) and assume that σ = ∞, so that algorithms and labor are perfectly substitutable within a task.”

  • The authors state this as an assumption.

    The initial cost share of AI is small.

    “We also focus on the realistic case in which the initial cost share of AI, denoted by s A e (= (Rk e (x) + w AI ℓ AI e (x))/total costs), is small.”

  • The authors state this as an assumption.

    AI services are produced by combining AI capital and in-house AI workers in fixed proportions.

    “We assume that AI services are provided by combining AI capital (machinery or algorithms) purchased from the outside, k e (x), and in-house workers operating, programming, or maintaining this capital, ℓ AI e (x), with the following technology:”

  • The authors state this as an assumption.

    All establishments are price takers for production workers, AI workers, and AI capital.

    “We assume throughout that all establishments are price takers for production workers, AI workers, and AI capital, whose respective prices are: w, w AI , and R.”

  • The authors state this as an assumption.

    The set of tasks in the economy is partitioned into tasks performed by distinct occupations.

    “Formally, we assume that the set of tasks in the economy, T , is partitioned into tasks performed by a set of distinct occupations and denote the set of tasks performed in occupation o ∈ O by T o .”

  • The paper relies on this without support.

    AI adoption can be partially identified from the footprints it leaves at adopting establishments as they hire workers specializing in AI-related activities.

    “Our starting point is that AI adoption can be partially identified from the footprints it leaves at adopting establishments as they hire workers specializing in AI-related activities, such as supervised and unsupervised learning, natural language processing, machine translation, or image recognition.”

  • The paper relies on this without support.

    Baseline occupational structure measures AI exposure throughout the analysis.

    “We construct an establishment's AI exposure from its baseline (2010)(2011)(2012) occupational structure according to each one of these indices, and use these baseline measures as proxies for AI exposure throughout our analysis.”

  • The authors state this as an assumption.

    Each establishment faces a downward-sloping demand curve for its product and sets its price to maximize profits.

    “We also assume that each establishment faces a downwardsloping demand curve for its product and will set its price p e to maximize profits (and its problem is separable from the profit-maximization problem of the firm's other establishments in case of multi-establishment firms).”

  • The authors state this as an assumption.

    Each establishment is small in the labor market and takes other prices and aggregate output as given.

    “In this profit-maximization problem, we assume that each establishment is small in the labor market and takes other prices and aggregate output as given.”

  • The authors state this as an assumption.

    The importance or quality of each task in production, α(x), is common across establishments.

    “where T is the set of feasible tasks, a subset T e of which is used in the production process of establishment e, and α(x) ≥ 0 designates the importance or quality of task x in the production process and are common across establishments.”

  • The authors state this as an assumption.

    The task-specific labor and AI productivity parameters γℓ(x) and γa(x) are common across establishments.

    “where σ is the elasticity of substitution between labor and algorithms, and γ ℓ (x) and γ a (x) are assumed to be common across establishments.”

  • The authors state this as an assumption.

    All establishments have constant returns to scale.

    “We impose Te α(x)dx = 1 for all feasible T e , which ensures that all establishments have constant returns to scale.”

  • The authors state this as an assumption.

    1.6% of the increase in AI vacancies were driven by task-level substitution of algorithms for labor.

    “Assuming that 1.6% of these increases-the partial R 2 of our AI exposure measure in Table 1-were driven by task-level substitution of algorithms for labor”

  • The paper relies on this without support.

    AI is intended to supplement, replicate, and in some cases, exceed human-level intelligence in a variety of tasks.

    “AI is intended to supplement, replicate, and in some cases, exceed human-level intelligence in a variety of tasks.”

  • The paper relies on this without support.

    The US BLS Occupational Employment Statistics (OES) data provides more accurate estimates of employment and wages in occupations than is available from household surveys.

    “This establishment-based data series provides more accurate estimates of employment and wages in occupations than is available from household surveys.”

  • The authors state this as an assumption.

    The common technology assumption in (2) and (3) holds.

    “The common technology assumption in (2) and (3) can also be relaxed but is useful for simplifying the exposition by ensuring that differences in factor demands across establishments are driven entirely by task structures, making the link between the model and our empirical approach more transparent.”

  • The authors state this as an assumption.

    The share of AI algorithms in cost is initially small.

    “When σ = ∞ as in Propositions 1 and 2 below and the share of AI algorithms in cost is initially small”

  • The authors state this as an assumption.

    The passthrough rate is less than 1/ε_e.

    “when the passthrough rate is less than 1/ε e”

  • The authors state this as an assumption.

    Establishment employment follows the law of motion ℓ_{e,t+1} = f v_{e,t} + (1 - s) ℓ_{e,t}.

    “we assume that establishment employment, ℓ e,t , follows the law of motion ℓ e,t+1 = f v e,t + (1 -s) ℓ e,t , where v e,t denotes the establishment's vacancies, f is the vacancy fill rate and s is the separation rate.”

  • The authors state this as an assumption.

    The establishment is in steady state initially in 2010.

    “We calculate employment in 2010 by assuming that the establishment is in steady state initially”
20 of 20

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • The literature often measures firm-level AI adoption using labor market data because firms rarely disclose AI-specific capital expenditures.

    “Because firms rarely disclose their AI-specific capital expenditures, the literature often measures firm-level AI adoption using labor market data.”

    From The Growth and Performance of Artificial Intelligence in Asset Management (2026) · cites this paper as Acemoglu, Autor, Hazell, and Restrepo (2022)

    The Introduction (page not provided) says the paper identifies AI adoption from AI-related hiring and online vacancy postings, supporting the use of labor-market data as a measure. It does not say that firms rarely disclose AI-specific capital expenditures or give that as the reason for using those data; nor does it establish that the literature often uses this approach. (LLM, conf=0.86)

    Experimental — accuracy not yet measured

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • AI-vacancy-share effect is described as 19 percentage points instead of about 0.18

    Major

    The stated column-6 estimate for the change in AI-vacancy share is 0.18, and the table defines that outcome as the change in share multiplied by 100. It therefore represents about 0.18 percentage points, not 19 percentage points.

  • Claim of no occupational wage effect conflicts with Table 8

    Major

    The paper says it detects no wage effect in more AI-exposed occupations after 2010. Table 8 reports a post-2010 occupational wage-growth estimate of -0.17 with a standard error of 0.06 for the Felten measure, an estimate approximately 2.8 standard errors from zero.

  • Figure 6 caption interchanges the Webb and SML panels

    Minor

    The plotted panels are labeled Felten, Webb, then SML, but the caption identifies the second as SML and the third as Webb.

  • Hiring-increase condition has the wrong threshold

    Minor

    Equation (5) makes the employment effect proportional to -1 plus the productivity term. Hiring increases only when that term exceeds 1, whereas the text says it need only exceed -1.

  • Table 4 specifies a sample posting in 2019 despite an end date of 2018

    Minor

    The table labels its outcome and sample period as ending in 2018, and the described source data run only through October 2018, but its caption says the establishments posted in both 2010 and 2019.

  • Skill-family descriptions reverse the signs of the plotted outcomes

    Minor

    Figure 7 is identified as negative skill change, yet the text describes its estimates as increasing demand. Figure 8 is identified as positive skill change, yet the text describes its estimates as lower demand.

6 of 6

Cites

Flagged first, then by citations

Research works printed but not linked: 12

  1. Babina Tania, 2020, Alex Xi He, and James Hodson
  2. Bessen James, Automatic reaction: What happens to workers at firms that automate? Unpublished manuscript
  3. Bessen James, 2018, Lydia Reichensperger, and Robert Seamans.
  4. Bonfiglioli Alessandra, Robot imports and firm level outcomes. Unpublished manuscript
  5. Brynjolfsson Erik, Machine learning and occupational change. Unpublished manuscript
  6. Carnevale Anthony, Understanding job ads data: A technical report. Unpublished manuscript
  7. Felten Edward, 2019, Academy of Management Proceedings 1, https://doi.org/10.5465/AMBPP., 2019, 140 · DOI
  8. Ford Martin, The rise of the robots
  9. Grennan Jillian, Artificial intelligence and high-skilled work: Evidence from analysts. Unpublished manuscript
  10. Humlum Anders, Robot adoption and labor market dynamics. Unpublished manuscript
  11. McKinsey Global Institute, Artificial intelligence: The next digital frontier? Unpublished manuscript
  12. West Darrell, AI, and automation.

Cited by

Sorted by citations

Related

Authors
Acemoglu, D., Autor, D., Hazell, J., & Restrepo, P.
Affiliations
  • Massachusetts Institute of Technology (US)
  • London School of Economics and Political Science (GB)
  • Boston University (US)
ORCID
Autor, David
Venue
Journal of Labor Economics
Publisher
University of Chicago Press
Published
2022-04-01
Volume
40
Issue
S1
Pages
S293-S340
ISSN
  • 0734-306X (ISSN-L)
  • 1537-5307
DOI
10.1086/718327
OpenAlex
W4224990248
Cited by
1,193
References
29
Citations by year
2026: 395 · 2025: 324 · 2024: 317 · 2023: 91 · 2022: 41 · 2021: 12 · 2020: 7 · 2019: 2 · 2018: 1 · 2017: 1 · 2013: 1
FWCI
31.58
Type
journal article
Language
en
Open access
Locations
  • Journal of Labor Economics (publishedVersion)
  • SSRN Electronic Journal (publishedVersion)
  • London School of Economics and Political Science Theses Online (London School of Economics and Political Science) (acceptedVersion)
  • London School of Economics and Political Science Research Online (London School of Economics and Political Science) (acceptedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • SSRN Electronic Journal (submittedVersion)
  • DSpace@MIT (Massachusetts Institute of Technology) (submittedVersion, cc-by-nc)
Funders
  • National Science Foundation
  • Alfred P. Sloan Foundation
  • Smith Richardson Foundation
  • Sid W. Richardson Foundation
  • Washington Center for Equitable Growth
  • Economic and Social Research Council (ES/T014431/1)
Topics
  • Social Sciences › Economics, Econometrics and Finance › Economics and Econometrics › Labor market dynamics and wage inequality
  • Social Sciences › Sociology and Political Science › Digital Economy and Work Transformation
  • Social Sciences › Decision Sciences › Management Science and Operations Research › Impact of AI and Big Data on Business and Society
Keywords
artificial intelligence, labor market outcomes, online job postings, labor substitution, skill requirements, personnel selection, wage growth, occupation, industry

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-11-06.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-11-06.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review