The DOI matches its Crossref record; 23 of the 23 references it cites were found in a registry or an index; 3 of the 38 sources its claims cite were checked for support.
AI, Output, and Employment
Open accessAbstract
Does artificial intelligence (AI) increase productivity - and does it displace workers? We examine aggregate effects using administrative data covering essentially all U.S. employers in a difference-in-differences design exploiting occupational AI exposure across industries and states. A one standard deviation increase in exposure raises output by 7%, with effects emerging in 2021 when enterprise AI tools entered the market. Employment effects follow the same timing but diverge by exposure type: where AI likely requires human collaboration, employment rises 4%; where AI can perform tasks independently, we find no significant employment effect. Results are robust to state-by-year and industry-by-year fixed effects and suggest AI has caused a decrease in the labor share of income.
Overview
CoChat Summary
TL;DR
Using administrative data covering essentially all U.S. employers and a difference-in-differences design, the paper finds that a one standard deviation increase in occupational AI exposure raised sector output, employment, and wage bills beginning in 2021, with augmentative exposure driving positive employment effects and automating exposure showing no significant employment effect. The output gains are capital-biased: only about 29% of AI-induced output gains accrue to labor, and AI reduces the labor share by roughly 5 percentage points per standard deviation of exposure.
Key findings
- A one standard deviation increase in AI exposure raises sector output by roughly 10% by 2023–2024 (event-study coefficients rise to 0.095–0.10); the abstract reports a 7% output increase. Effects emerge in 2021, coinciding with enterprise AI tool deployment.
- By 2024, a one standard deviation increase in exposure is associated with a 4.8% higher wage bill (p < 0.01), 3.9% higher employment (p < 0.01), and 2.6% more establishments (p < 0.05); the abstract summarizes employment as rising approximately 4% and wage bills by 5%.
- Decomposing exposure: augmentative exposure raises employment by 3.6% by 2024 (p < 0.01) and wage bills by 4.2% (p < 0.01); automating exposure has no significant employment effect (small positive). The abstract states augmentative employment rises 4%.
- Individual-level ACS data show hourly wages rise 0.9–1.1% per standard deviation by 2023–2024, larger for ages 22–30 (1.4% by 2023) than ages 31–65 (0.8%), and concentrated among college graduates (1.4% by 2023–2024) versus non-college (0.3–0.5%). Only about 29% of AI-induced output gains accrue to labor, and AI reduces the labor share by roughly 5 percentage points per standard deviation of exposure.
Methods
- State × 3-digit NAICS industry panel covering 2017–2024. Output from Bureau of Economic Analysis GDP-by-State; employment, wage bill, and establishments from the Quarterly Census of Employment and Wages (QCEW), covering over 95% of U.S. wage and salary jobs; individual-level wages from the American Community Survey (ACS).
- AI exposure measured using Eloundou et al. (2024) occupation-level LLM exposure scores aggregated to state-industry cells with pre-2020 ACS occupation shares; decomposed into automating (E1) and augmentative (E2) exposure.
- Difference-in-differences event study with continuous treatment: log outcomes regressed on exposure × year indicators (omitting 2020) with state×industry, state×year, and industry×year fixed effects; standard errors clustered at the state level. Robustness includes controls for teleworkability, interest rate exposure, excluding information and finance sectors, state-by-industry linear trends, and placebo tests.
Limitations
- Stated: the exposure measure is static, constructed from pre-2020 occupational shares and task assessments that predate current AI tools, so estimates likely understate true effects; the analysis is partial equilibrium within sectors, with cross-sector reallocation, factor price adjustment, and general equilibrium effects beyond the design; results are specific to a low-collective-bargaining environment; residual confounding cannot be fully ruled out.
- Stated/evident: QCEW excludes self-employed and some agricultural workers, so shifts to self-employment would appear as employment declines; 2024 ACS data were unavailable; automating employment coefficients are detrended by subtracting a linear pre-trend because of a significant positive pre-trend in raw estimates (table note); mismeasurement in the time-invariant treatment variable attenuates estimates.
- Evident: the abstract’s 7% output effect differs from the results’ roughly 10% by end of sample; the paper notes the labor-share reduction estimate is imprecise; positive effects are concentrated in high labor-market-fluidity states, limiting external validity.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
2 Valid · 8 Not valid · 32 Not checked · 147 Own results and interpretations
At least one cited source doesn’t support this claim or points to a different paper, and none confirmed it — see the verdicts below.
In the paper
“Their within-occupation design detects employ-ment declines in these roles at large firms.”
§ AI, Output, and Employment > 6.2 Comparison with Humlum and Vestergaard (2025)
What the sources say
- Disputedcites Humlum and Vestergaard (2025) sourceExperimental — accuracy not yet measured
The within-occupation design (survey-occupation × time fixed effects, Section 3.1.2, p. 23) yields nulls, not detected declines: Section 3.4 (p. 33) states 'we find no statistically significant effects in any occupation, and point estimates are generally near zero' for workplace employment and early-career jobs, including software developers and customer support, and Appendix F.4 (p. 102) shows overall workplace employment remains flat. The employment declines in software development and customer support that the paper does document are aggregate early-career trends in Appendix D.2 (pp. 84-85), a replication of Brynjolfsson, Chandar and Chen (2025), which the paper explicitly says 'are not driven by firms adopting generative AI chatbots'. No passage attributes any decline to large firms — firm employment appears only as a predictor of adoption in F.1 (p. 96), and there is no firm-size heterogeneity result for employment. (LLM, conf=0.82)
Checked by a reasoning model — may be wrong.
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Disputed
Every source the paper cites for this claim was checked and supports it.
In the paper
“Humlum and Vestergaard (2025) study individual-level outcomes using linked survey and administrative data in Denmark, comparing earnings and hours for workers who self-report AI chatbot use, and find no significant effects.”
§ AI, Output, and Employment > 6.2 Comparison with Humlum and Vestergaard (2025)
What the sources say
- Supportedcites Humlum and Vestergaard (2025) source
The Abstract (p. 2) states the design exactly as the citing sentence describes it: linking large-scale adoption surveys to administrative labor market records in Denmark, with difference-in-differences estimating 'precise null effects on earnings and recorded hours at both the worker and workplace levels.' The Abstract (p. 5) adds 'We find precisely null effects of AI chatbot use on earnings and hours,' and Section 3.2 Discussion (p. 27) confirms that 'users of AI chatbots have not experienced differential changes in earnings or hours,' including among daily users. Self-reported chatbot use is the measure used to define adopters (pp. 25-26). (LLM, conf=0.93)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Their design exploits within-occupation differences in adoption: among workers in the same occupation at similar firms, those who report using AI chatbots are compared to those who do not.”
§ AI, Output, and Employment > 6.2 Comparison with Humlum and Vestergaard (2025)
What the sources say
- Supportedcites Humlum and Vestergaard (2025) source
The identification strategy (3.1.1, p. 20) states the paper compares adopters and non-adopters before and after the arrival of AI chatbots at worker and workplace levels, and the Figure 5 notes (p. 45) specify that chatbot adopters are compared to non-adopters 'controlling for survey occupation fixed effects and predetermined worker characteristics' — i.e. a within-occupation users-versus-non-users contrast. The 'similar firms' element is the paper's own framing in 2.1 (p. 12): firm characteristics explain only 1.3% of the within-occupation variation in encouragement, so 'many otherwise comparable firms have adopted markedly different strategies toward AI chatbots, even for similar workers within the same occupation.' The only nuance is that the citing sentence describes the comparison without the before/after (difference-in-differences) dimension, but nothing attributed to this source is unsupported. (LLM, conf=0.85)
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“A one standard deviation increase in AI exposure raises sector output by roughly 10 percent by the end of the sample-an economically large effect that accumulates gradually over the post-2021 period rather than appearing as a one-time jump, consistent with the diffusion of enterprise tools across firms within exposed sectors (Bresnahan et al., 2002;Bloom et al., 2012).”
§ AI, Output, and Employment > 5.1 Output Effects
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“By 2024, a one standard deviation increase in exposure is associated with a 4.8 percent increase in the wage bill (p < 0.01), a 3.9 percent increase in employment (p < 0.01), and a 2.6 percent increase in establishments (p < 0.05).”
§ AI, Output, and Employment > 5.2 Labor Market Effects
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
87 more claims
Create a free account to see all 189 claims and to verify the 12 references still open.
- It may state two separate claims as one. 76
- Probably not a substantial claim. 38
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.
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ AI, Output, and Employment > 5.3 Robustness
“This alternative measure does not predict post-2020 growth, supporting the interpretation that our results capture AI-specific effects rather than spurious correlates.”
§ AI, Output, and Employment > Panel B: Augmentation Exposure
“Drafting Intensity × 2024 .011.012* * .012* * [.007] [.005] [.005]”
Rated major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ AI, Output, and Employment > 5.5 Individual-Level Wages
“Positive effects on hourly wages emerge in 2022 (0.3-0.5 percent) and strengthen through 2023-2024 (0.9-1.1 percent per standard deviation of exposure)”
§ AI, Output, and Employment > Panel B: log(Employment)
“2024 ACS data were not available at this writing.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ AI, Output, and Employment > Panel B: log(Employment)
“We also report results using 2022 as the base year, corresponding to ChatGPT's consumer launch, in the Online Appendix.”
§ AI, Output, and Employment > References
“Figure A.14: Robustness of Dynamic Effects Using Alternative Base Year Normalization -.02 0 .02 .04 .06 .08 Effect on log total wages (per 1 s.d. exposure) 2017 2018 2019 2022 2023 2024 Year [reference: 2019]”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ AI, Output, and Employment > E2 LLM-assisted exposure:
“At the occupation level, human-derived and GPT-4-derived exposure scores correlate at r ≈ 0.65.”
§ AI, Output, and Employment > Panel B: log(Employment)
“The two sets of scores correlate at approximately r = 0.40 at the occupation level.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ AI, Output, and Employment > Panel B: Augmentation Exposure
“interacted with year dummies (normalized to 2019), conditional on state × industry and year fixed effects.”
§ AI, Output, and Employment > Panel B: Augmentation Exposure
“Exposure × 2019 .000.013.013[.017] [.017] [.014]”
Citations
Cites
Robots and Jobs: Evidence from US Labor Markets
Acemoglu, D., & Restrepo, P. (2020) · Journal of Political Economy · doi:10.1086/705716 · Cited by 4,088
Bresnahan, T. F., Brynjolfsson, E., & Hitt, L. M. (2002) · The Quarterly Journal of Economics · doi:10.1162/003355302753399526 · Cited by 3,211
GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models
Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023) · arXiv · doi:10.48550/arxiv.2303.10130 · Cited by 2,330
Experimental evidence on the productivity effects of generative artificial intelligence
Noy, S., & Zhang, W. (2023) · Science · doi:10.1126/science.adh2586 · Cited by 2,150
How many jobs can be done at home?
Dingel, J. I., & Neiman, B. (2020) · Journal of Public Economics · doi:10.1016/j.jpubeco.2020.104235 · Cited by 2,055
Brynjolfsson, E., Li, D., & Raymond, L. (2025) · The Quarterly Journal of Economics · doi:10.1093/qje/qjae044 · Cited by 1,083
Americans Do IT Better: US Multinationals and the Productivity Miracle
Bloom, N., Sadun, R., & Reenen, J. V. (2012) · American Economic Review · doi:10.1257/aer.102.1.167 · Cited by 1,028
What Can Machines Learn and What Does It Mean for Occupations and the Economy?
Brynjolfsson, E., Mitchell, T., & Rock, D. (2018) · AEA Papers and Proceedings · doi:10.1257/pandp.20181019 · Cited by 695
A Quarterly, Utilization-Adjusted Series on Total Factor Productivity
Francisco, F. R. B. O. S., & Fernald, J. G. (2012) · Federal Reserve Bank of San Francisco, Working Paper Series · doi:10.24148/wp2012-19 · Cited by 551
The Rapid Adoption of Generative AI
Bick, A., Blandin, A., & Deming, D. (2024) · NBER Working Paper · doi:10.3386/w32966 · Cited by 200
Research works printed but not linked: 9
- Kreiner, C. T. and Svarer, M. (2022). Danish flexicurity: Lessons for Comparative labor market institutions. Journal of Economic Perspectives, 36(4):81-102.
- Lichtinger, G. and Massoum, S. (2025). Generative ai as seniority-biased technological change: Evidence from u.s. résumé and job posting data. SSRN working paper.
- Pitchbook (2026). Venture monitor: The definitive review of the US venture capital ecosystem, Q4 2025. Technical report, PitchBook Data, Inc. and National Venture Capital Association.
- Westby, S. and Modestino, A. (2025). The impact of generative ai on job opportunities for junior software developers. SSRN working paper. -.04 -.02 0 .02 .04 .06 Effect on log total wages (per 1 s.d. exposure) 2017 2018 2019 2020 2021 2022 2023 2024 Year [reference: 2020]
- FE: state×industry, state×year, industry×year; SEs clustered by state. Exposure z-scored. Panel B: log(Employment) Notes.-Sources: Eloundou et al. (2024), Quarterly Census of Employment and Wages (QCEW) (2017-2024), and the American Community Survey (ACS) (2011-2015, 2016-2019). The figure reports the coefficients associated with log wage bill (deflated using the 2017 personal consumption expenditure index) and log employment on a state-by-industry (NAICS3) weighted exposure index of exposure index (LLM-coded) from Eloundou et al. (2024) using the ACS occupation-by-state-by-industry shares pre-sample, interacted with year dummies (normalized to 2020), conditional on state × industry and year fixed effects. Standard errors are clustered at the state-level and observations are weighted by the number of employees in a state-industry cell. Figure A.12: Robustness of Dynamic Effects Dropping NAICS 51 and 54 -.05 0 .05 .1 Effect on log total wages (per 1 s.d. exposure) 2017 2018 2019 2020 2021 2022 2023 2024 Year [reference: 2020]
- FE: state×industry, state×year, industry×year; SEs clustered by state. Exposure z-scored. Panel B: log(Employment) Notes.-Sources: Eloundou et al. (2024), Quarterly Census of Employment and Wages (QCEW) (2017-2024), and the American Community Survey (ACS) (2000, 2005-2009, 2011-2015, 2016-2019). The figure reports the coefficients associated with log wage bill (deflated using the 2017 personal consumption expenditure index) and log employment on a state-by-industry (NAICS3) weighted exposure index of exposure index (LLM-coded) from Eloundou et al. (2024) using the ACS occupation-by-state-by- industry shares pre-sample, interacted with year dummies (normalized to 2020), conditional on state × industry and year fixed effects. We drop NAICS 51 and 54. Standard errors are clustered at the state-level and observations are weighted by the number of employees in a state-industry cell. Figure A.13: Robustness of Dynamic Effects Using Historical Industry-State × Year Employment Trends -.02 0 .02 .04 .06 Effect on log total wages (per 1 s.d. exposure) 2017 2018 2019 2020 2021 2022 2023 2024 Year [reference: 2020]
- Panel B: log(Employment) Notes.-Sources: Eloundou et al. (2024), Quarterly Census of Employment and Wages (QCEW) (2016-2019, 2017-2024), and the American Community Survey (ACS) (2000, 2005-2009, 2011-2015, 2016-2019). The figure reports the coefficients associated with log wage bill (deflated using the 2017 personal consumption expenditure index) and log employment on a state-by-industry (NAICS3) weighted exposure index of exposure index (LLM-coded) from Eloundou et al. (2024) using the ACS occupation-by- state-by-industry shares pre-sample, interacted with year dummies (normalized to 2020), conditional on state × industry and year fixed effects. We also control for historical 2016-2019 employment growth within the state-industry cell interacted with year fixed effects. Standard errors are clustered at the state-level and observations are weighted by the number of employees in a state-industry cell. Figure A.14: Robustness of Dynamic Effects Using Alternative Base Year Normalization -.02 0 .02 .04 .06 .08 Effect on log total wages (per 1 s.d. exposure) 2017 2018 2019 2022 2023 2024 Year [reference: 2019]
- FE: state×industry, state×year, industry×year; SEs clustered by state. Exposure z-scored. Panel B: log(Employment) Notes.-Sources: Eloundou et al. (2024), Quarterly Census of Employment and Wages (QCEW) (2017-2024), and the American Community Survey (ACS) (2000, 2005-2009, 2011-2015, 2016-2019). The figure reports the coefficients associated with log wage bill (deflated using the 2017 personal consumption expenditure index) and log employment on a state-by-industry (NAICS3) weighted exposure index of exposure index (LLM-coded) from Eloundou et al. (2024) using the ACS occupation-by-state-by- industry shares pre-sample, interacted with year dummies (normalized to 2019), conditional on state × industry and year fixed effects. We drop 2020 in this robustness exercise. Standard errors are clustered at the state-level and observations are weighted by the number of employees in a state-industry cell. Figure A.15: Heterogeneous Effects of Generative AI Exposure on Workers, by Turnover Rate -.05 0 .05 .1 Effect on log total wages (per 1 s.d. exposure) 2017 2018 2019 2020 2021 2022 2023 2024 Year [reference: 2020]
- High turnover Low turnover FE: state×industry, state×year, industry×year; SEs clustered by state. Exposure z-scored. Panel B: log(Employment) by High/Low Turnover Rate Notes.-Sources: Eloundou et al. (2024), Quarterly Census of Employment and Wages (QCEW) (2017-2024), Longitudinal Employer Household Dynamics (LEHD, 2016-2019), and the American Community Survey (ACS) (2000, 2005-2009, 2011-2015, 2016-2019). The figure reports the coefficients associated with log wage bill (deflated using the 2017 personal consumption expenditure index) and log employment on a state-by-industry (NAICS3) weighted exposure index of β displacement risk of at least 10% of tasks (LLM-coded) from Eloundou et al. (2024) using the ACS occupation-by-state-by-industry shares pre-sample, interacted with year dummies (normalized to 2020), conditional on state × industry and year fixed effects. The coefficients are estimated on states with job turnover rates above versus below the median based on 2016-19 averages. Standard errors are clustered at the state-level and observations are weighted by the number of employees in a state-industry cell.
Cited by
Displacement or Augmentation? The Effects of AI Innovation on Workforce Dynamics and Firm ValuePreprint
Chen, M. A., & Wang, X. (2025) · SSRN Electronic Journal · doi:10.2139/ssrn.5246388 · Cited by 6
AI Utilization and Economic Performance
Makridis, C. A., Highfill, T., & Samuels, J. D. (2026) · doi:10.66137/hmxc8574 · Cited by 3
Self-citationMakridis, C. A., & Hickman, L. (2023) · SSRN Electronic Journal · doi:10.2139/ssrn.4365176 · Cited by 1
Self-citationAI-exposed jobs deteriorated before ChatGPT
Frank, M. R., Sabet, A. J., Simon, L., Bana, S. H., & Yu, R. (2026) · arXiv · doi:10.48550/arxiv.2601.02554
The LLM Productivity Cliff: Threshold Productivity and AI-Native InequalityPreprint
Bisardi, F. (2026) · Qeios · doi:10.32388/5vaolk.2
Metadata
- Authors
- Johnston, A., & Makridis, C. A.
- Affiliations
- University of California, Merced (US)
- Ifo Institute for Economic Research (DE)
- Corresponding author
- Johnston, Andrew
- Venue
- CESifo Working Papers
- Publisher
- CESifo
- Published
- 2026-03-25
- ISSN
- 1617-9595 (ISSN-L)
- 2364-1428
- OpenAlex
- W7167807929
- Cited by
- 8
- References
- 20
- Citations by year
- 2026: 6
- Type
- report
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- other-oa
- Locations
- CESifo (publishedVersion, other-oa)
- SSRN Electronic Journal (submittedVersion)
- Topics
- Social Sciences › Economics, Econometrics and Finance › Economics and Econometrics › Labor market dynamics and wage inequality
- Social Sciences › Decision Sciences › Management Science and Operations Research › Impact of AI and Big Data on Business and Society
- Social Sciences › Sociology and Political Science › Digital Economy and Work Transformation
- Keywords
- artificial intelligence, labor productivity, employment effects, job displacement, wage share, human-AI collaboration, aggregate production, United States
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.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-11-17, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-11-17 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (qwen3.8-omni-flash).
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
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: AI, Output, and Employment. https://app.cochat.ai/papers/10.65864/ylohmsrt5g
- The paper
- Johnston, A., & Makridis, C. A. (2026). AI, Output, and Employment. CESifo Working Papers. https://doi.org/10.65864/ylohmsrt5g
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.