Pith. sign in

REVIEW 2 major objections 4 minor 60 references

The Potential Impact of Disruptive AI Innovations on U.S. Occupations

T0 review · 2 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Splitting U.S. AI patents by citation-based disruption reveals that disruptive innovations target mental and unpredictable tasks while consolidating innovations target physical, routine, and solo tasks.

desk verdict Useful new disruptive/consolidating split of AI patents by task exposure, but the fragile argmax matching and internal contradictions mean the unpredictable and collaborative claims are oversold. read the letter →

arxiv 2507.11403 v1 pith:7V4QTCHP submitted 2025-07-15 cs.CY cs.SI

Munjung Kim , Marios Constantinides , Sanja Šćepanović , Yong-Yeol Ahn , Daniele Quercia This is my paper · ORCID
classification cs.CYcs.SI
keywords disruptionindexAIpatentsoccupationaltaskslabormarketimpacttaskcharacteristicspatentcitationsshortagesregionalanalysis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that AI is not a single force: when U.S. AI patents are split by how much they break from prior technology, the two kinds point at opposite kinds of work. Consolidating innovations, which build on existing ideas, are tied to physical, predictable, solo tasks in manufacturing and construction, mostly in the Midwest and central states. Disruptive innovations, which redirect later work, are tied to mental and unpredictable tasks in coastal science and technology sectors, but notably not to collaborative tasks. If true, standard exposure measures that treat all AI alike will miss a systematic difference in which jobs are touched first. The paper also claims that disruptive AI concentrates in sectors already short of skilled workers, suggesting it may fill gaps rather than flood surplus labor markets.

What carries the argument

The load-bearing device is the disruption index for patents, a citation-based score of whether a patent redirects or reinforces the direction of later inventions, defined as $D_f = \frac{n_i - n_j}{n_i + n_j + n_k}$, where $n_i$ counts later patents that cite only the focal patent, $n_j$ counts later patents that cite both it and its references, and $n_k$ counts later patents that cite only its references. Patents in the top quartile are labeled disruptive and those in the bottom quartile consolidating. The index is paired with a matching step: each job task is assigned the patent with the highest cosine similarity between abstract and task text, and tasks above the 90th percentile of similarity are counted as impacted by that patent type. Together, the two mechanisms turn citation structure into a claim about which kinds of work are touched.

What would settle it

Compare the 397 disruptive-AI-impacted tasks and 536 consolidating-AI-impacted tasks against actual occupation-level outcomes from 2015 to 2022: if workers in disruptive-impacted occupations show no systematic difference in employment, wage, or task-content changes relative to consolidating-impacted occupations, the claimed split would be falsified.

Watch

Extended reading notes

Core claim

The central discovery is that a single patent corpus, split by the disruption index, yields opposite occupational task profiles. Disruptive AI patents, defined as the top quartile of the index, disproportionately match tasks labeled mental and unpredictable, found in information technology, professional services, education, and transportation; consolidating AI patents, the bottom quartile, disproportionately match physical, predictable, and solo tasks in manufacturing, construction, mining, agriculture, and retail. The match is made by embedding patent abstracts and task descriptions and taking the highest cosine similarity, with a threshold validated at the 90th percentile. A null-model significance test with 500 random shuffles shows the associations are unlikely to be chance. The paper additionally finds that the share of tasks touched by disruptive AI within an industry correlates with that industry's 2022 vacancy rate, with a correlation of $r=0.73$ after removing one outlier, while the consolidating correlation is weaker and not statistically significant.

Load-bearing premise

The load-bearing premise is that when a patent's abstract and a task's description read very similarly, the patent actually changes that job task in the real world, and the paper does not check this against actual job losses, wage changes, or task reassignments.

Editorial extensions

If this is right

  • Exposure measures that treat all AI as one will systematically understate the distinct risks: consolidating AI extends existing automation in production regions, while disruptive AI reaches into mental, unpredictable work in coastal service and tech sectors.
  • Jobs built around collaborative tasks appear least touched by either type, implying that teamwork and social interaction remain a relative safe zone for now.
  • High-vacancy industries show a stronger association with disruptive AI impact, so AI adoption may be accelerating where workers are already scarce rather than where workers are abundant.
  • The geographic split, with coastal states producing more disruptive patents and central states more consolidating ones, suggests regional policy responses will differ: reskilling for automation in the center, support for innovation and lifelong learning on the coasts.
  • The framework gives firms a task-audit tool: identify which tasks in their workflow are likely to be redirected by disruptive patents rather than merely refined by consolidating ones.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the pattern holds, existing AI exposure scores computed without the disruption split may be positively misleading for occupations whose tasks are mental but unpredictable: those tasks would appear low-exposure under routine-automation logic while actually being first-order targets.
  • The collaborative-task exception suggests a testable boundary: human-AI teaming tools, such as copilots and shared decision support, may be the next place disruptive patents appear, since solo cognitive tasks are already being hit.
  • The vacancy-rate correlation could be read either way; a direct test with occupation-level wage and employment changes, which the paper flags as future work, would tell whether disruptive AI is filling shortages or creating them.
  • The same patent-to-task matching pipeline could be re-run with newer patents and with the threshold validated against actual job transitions; the divergence between the 2015-2019 patent cohort and current AI would reveal whether the coastal, mental-task profile is stable or drifting.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper proposes a computational pipeline to distinguish 'consolidating' from 'disruptive' AI innovations and to link each type to occupational task characteristics. The authors compute the Funk and Owen-Smith disruption index for 3,237 U.S. AI patents, match O*NET task descriptions to patent abstracts via SBERT cosine similarity, and label tasks as impacted by disruptive or consolidating AI depending on the quartile of their single best-matching patent. They then compare the shares of mental/physical, predictable/unpredictable, and individual/collaborative tasks across these groups using a permutation null model, and extend the analysis to industries, states, and vacancy rates. The headline claims are that consolidating AI targets physical, predictable, individual tasks (Midwest manufacturing/construction), while disruptive AI targets mental and unpredictable tasks (coastal science/technology), with collaborative tasks remaining largely untouched; they also report a positive correlation between disruptive AI exposure and industry vacancy rates.

Significance. If the central contrast were robust, the paper would provide a useful refinement of the routine-vs-non-routine automation literature by separating AI innovations according to their citation-based disruptiveness. The strengths include a transparent and reproducible pipeline (code and data are linked), a null model with explicit z-scores, a scoping review to motivate the three task dimensions, and a human validation component with documented agreement rates. The paper also clearly states its limitations, including the lack of direct testing against labor outcomes. However, the significance is currently limited by fragile components of the measurement pipeline and by internal inconsistencies between the main text and the supplementary material.

major comments (2)
  1. [Results, first paragraph; Figure 2] The classification of each task as impacted by disruptive or consolidating AI hinges on a single argmax: the task is assigned the label of the patent with the highest SBERT cosine similarity, and the 90th-percentile threshold is applied to that same top score. This rule is fragile to near-ties: a task whose top patent is just above the disruptive-quartile cutoff and whose second-best patent is just below the consolidating cutoff receives full weight in the disruptive category. The paper provides no robustness check, such as re-running the analysis with the second-best patent, using a margin threshold between the top-1 and top-2 similarities, or weighting all patents above the impact threshold. Because the central contrast between 'disruptive AI targets mental/unpredictable tasks' and 'consolidating AI targets physical/predictable tasks' is computed entirely from these labels, the headline z-scores in Figure 2 may reflect the arbitrary winner of the argmax rather than a stable property of tasks. I request a sensitivity analysis of the assignment rule before the central claim can be considered supported.
  2. [Results, first paragraph; Figure 2] The classification of each task as impacted by disruptive or consolidating AI hinges on a single argmax: the task is assigned the label of the patent with the highest SBERT cosine similarity, and the 90th-percentile threshold is applied to that same top score. This rule is fragile to near-ties: a task whose top patent is just above the disruptive-quartile cutoff and whose second-best patent is just below the consolidating cutoff receives full weight in the disruptive category. The paper provides no robustness check, such as re-running the analysis with the second-best patent, using a margin threshold between the top-1 and top-2 similarities, or weighting all patents above the impact threshold. Because the central contrast between 'disruptive AI targets mental/unpredictable tasks' and 'consolidating AI targets physical/predictable tasks' is computed entirely from these labels, the headline z-scores in Figure 2 may reflect the arbitrary winner of the argmax rather than a stable property of tasks. I request a sensitivity analysis of the assignment rule before the central claim can be considered supported.
minor comments (4)
  1. [Abstract vs Methods] The abstract says the patent data cover 2015–2022, but the Methods state that patents published between 2020 and 2022 are excluded because the disruption index needs citation accumulation time. The actual analysis period is 2015–2019. The abstract should be corrected to avoid misleading readers about the data range.
  2. [Supplementary 'Comparison of Task Classification Results'] The z-scores in Supplementary Figure S3 are reported separately for human and GPT annotations, but the main-text z-scores in Figure 2 are presumably based on the final (adjudicated) labels. The relationship between these three sets should be stated explicitly, especially because the supplementary reports that disruptive-AI tasks are more likely to be collaborative in both human and GPT annotations, which conflicts with the main text.
  3. [Methods, 'Classifying Tasks Based on Common Job Characteristics'] The paper samples 933 of the 17,548 tasks not impacted by AI using a network-centrality heuristic, then uses this sample in the null model and in the comparisons. The selection procedure is described, but the potential bias from oversampling central or representative tasks is not discussed. A brief analysis showing the robustness of the 'not impacted' category to alternative sampling would strengthen the comparison.
  4. [Supplementary Tables S4 and S5] The industry-level p-values in Tables S4 and S5 are not corrected for multiple comparisons across 19 industries. The paper highlights the Information sector (p = 0.033) and Manufacturing (p = 0.027) as significant, but these would not survive a simple Bonferroni correction. Reporting adjusted p-values or explicitly noting the exploratory nature of these tests would be appropriate.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the disruptive/consolidating split comes from external citation data and the task-type labels from O*NET/GPT-4o with human validation; the self-cited matching threshold is a methodological input, not an output.

full rationale

The central claim is not derived from its inputs by construction. Patent disruption labels are computed from an external bibliometric index (Eq. 1, Funk and Owen-Smith), and task characteristics come from O*NET descriptions classified by GPT-4o and validated by human annotators. The null model in Eq. (2) shuffles task characteristics while holding the patent-impact assignments fixed, so the reported z-scores are a genuine permutation test rather than a restatement of a fitted parameter. The only inherited component is the task–patent matching threshold, adopted via the authors' prior work: 'Following Septiandri et al., tasks with a cosine similarity above the 90th percentile were classified as impacted by AI.' That threshold is a methodological input and does not by itself determine whether disruptive patents preferentially match mental and unpredictable tasks; the empirical contrast between quartile-based patent types is not entailed by the threshold. Concerns about the argmax assignment rule and the human/GPT disagreement on collaboration are robustness and validity issues, not circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new entities. Its results rest on several domain assumptions: patent text similarity is a proxy for impact, the disruption index correctly labels patent types, GPT-4o labels are accurate despite partial human disagreement, and industry vacancy rates represent labor shortages. The free parameters are threshold choices and the outlier exclusion, all of which materially shape the reported findings.

free parameters (4)
  • Disruptive/consolidating quartile thresholds = top 25% / bottom 25% of disruption index
    Chosen by hand to define clear examples; directly determines which patents are labeled disruptive vs consolidating and hence all downstream task classifications.
  • Cosine similarity impact threshold = 90th percentile
    Tasks whose highest patent similarity exceeds the 90th percentile are classified as AI-impacted; this determines the impacted vs not-impacted split. The stated result (933 impacted of 17,879 tasks) does not match a literal 10% cutoff, suggesting the threshold is applied in a non-transparent way.
  • Network edge threshold for representative sampling = top 95th percentile of similarity edges
    Used to construct task networks and sample representative 'not impacted' tasks; affects the composition of the 933-task comparison set used in the main analysis.
  • Outlier exclusion in vacancy correlation = accommodation and food services removed
    The headline r=0.73 (p<0.05) for disruptive AI and vacancy rates requires removing this sector; with all sectors included r=0.40 (p=0.12) and not significant.
assumptions (5)
  • domain assumption Patent abstract text similarity to task description is a valid proxy for AI impact on that task
    Used to match AI patents to O*NET tasks (Methods: Matching tasks to AI patents). No validation against actual job outcomes; the authors list this as a limitation.
  • domain assumption Funk-Owen-Smith disruption index computed from USPTO citation data correctly identifies disruptive vs consolidating innovations
    The paper uses quartiles of this index as ground truth for patent type, relying on citation accumulation and the index's construct validity (refs 11, 12, 38).
  • domain assumption GPT-4o task labels are accurate for all 1,866 tasks in the analysis
    Only 77 tasks were human-validated; human annotators disagreed with GPT on the predictability of 'not impacted by AI' tasks (humans: predictable z=1.965; GPT: unpredictable z=1.178), yet GPT labels are used for the remaining tasks (Supplementary Information).
  • domain assumption O*NET task descriptions (version 26.3) are representative of actual job content
    All task-level results depend on this public database as the source of tasks.
  • domain assumption Industry vacancy rates in 2022 measure skilled labor shortages
    Used in the correlation that supports the claim about labor shortages; no direct measure of skill shortage is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Potential Impact of Disruptive AI Innovations on U.S. Occupations." pith.science (2026). https://pith.science/paper/7V4QTCHP

@misc{pith2026250711403,
  author       = {Pith},
  title        = {Pith review of: The Potential Impact of Disruptive AI Innovations on U.S. Occupations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7V4QTCHP}},
  note         = {Machine review of arXiv:2507.11403}
}
read the original abstract

The rapid rise of AI is poised to disrupt the labor market. However, AI is not a monolith; its impact depends on both the nature of the innovation and the jobs it affects. While computational approaches are emerging, there is no consensus on how to systematically measure an innovation's disruptive potential. Here, we calculate the disruption index of 3,237 U.S. AI patents (2015-2022) and link them to job tasks to distinguish between "consolidating" AI innovations that reinforce existing structures and "disruptive" AI innovations that alter them. Our analysis reveals that consolidating AI primarily targets physical, routine, and solo tasks, common in manufacturing and construction in the Midwest and central states. By contrast, disruptive AI affects unpredictable and mental tasks, particularly in coastal science and technology sectors. Surprisingly, we also find that disruptive AI disproportionately affects areas already facing skilled labor shortages, suggesting disruptive AI technologies may accelerate change where workers are scarce rather than replacing a surplus. Ultimately, consolidating AI appears to extend current automation trends, while disruptive AI is set to transform complex mental work, with a notable exception for collaborative tasks.

Figures

Figures reproduced from arXiv: 2507.11403 by the authors.

Figure 1
Figure 1. Three steps to link AI patents to job tasks. (Left) Step 1: We analyzed 3,237 AI patents from the U.S. Patent and Trademark Office (USPTO) and labeled each as disruptive or consolidating using the Disruption Index. Patents in the top 25% were disruptive; those in the bottom 25% were consolidating. (Middle) Step 2: We grouped 1,866 job tasks from the O*NET database using three dimensions: mental or physical, collabor… view at source ↗
Figure 2
Figure 2. Task characteristics by AI patent type. Each arrow shows how far the observed share of tasks with a certain characteristic deviates from what would be expected under random assignment. Black triangles show the actual z-scores. Positive values indicate a higher-than-expected presence of a trait (e.g., predictable, individual, or physical), while negative values suggest the opposite. Tasks not impacted by AI tend to b… view at source ↗
Figure 3
Figure 3. Top five industry sectors most affected by each type of AI. Bars show the difference in task impact between disruptive or consolidating AI and all AI patents. Positive values mean a sector is more affected by that type of AI than the overall average. in production-heavy industries. These include manufacturing (4.03% above average), construction (2.09%), mining (0.94%), agriculture (0.58%), and retail trade (0.37%). … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Regional patterns in AI development. Red indicates states where disruptive AI patents are more prevalent; blue indicates a focus on consolidating AI. Color intensity reflects the difference in patent share compared to general AI. A value of 0.01 means that, in that sta…
Figure 5
Figure 5. Figure 5: Industry-level relationship between vacancy rates and AI impact of disruptive vs. consolidating AI. A stronger positive correlation is observed between vacancy rates and impact of disruptive AI compared to consolidating AI. This suggests that disruptive AI patents may …
Figure 6
Figure 6. Figure 6: Overview of the Classification Process for Patents and Their Impacted Tasks. (A) AI-related patents were classified based on their Funk and Owen-Smith’s disruption index11. (B) Task descriptions were matched to patent abstracts based on the highest cosine similarity of…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 47 canonical work pages

  1. [1]

    & Zhu, X

    Demirci, O., Hannane, J. & Zhu, X. Who is ai replacing? the impact of generative ai on online freelancing platforms. Working Paper 11276, CESifo (2023). CESifo Working Paper No. 11276

    work page 2023

  2. [2]

    "Generate" the Future of Work through AI: Empirical Evidence from Online Labor Markets

    Liu, J., Xu, X., Li, Y . & Tan, Y . “generate" the future of work through ai: Empirical evidence from online labor markets. arXiv preprint arXiv:2308.05201 (2023)

    work page Pith review arXiv 2023

  3. [3]

    & Zhang, W

    Noy, S. & Zhang, W. Experimental evidence on the productivity effects of generative artificial intelligence. Science 381, 187–192, DOI: 10.1126/science.adh2586 (2023)

    doi:10.1126/science.adh2586 2023

  4. [4]

    & Goldfarb, A

    Agrawal, A., Gans, J. & Goldfarb, A. ChatGPT and How AI Disrupts Industries. Harv. Bus. Rev. (2022)

    work page 2022

  5. [5]

    & Bescond, D

    Gmyrek, P., Berg, J. & Bescond, D. Generative AI and jobs: a global analysis of potential effects on job quantity and quality (ILO, 2023)

    work page 2023

  6. [6]

    Cappelli, P., Tambe, P. S. & Yakubovich, V . Will large language models really change how work is done? MIT Sloan Manag. Rev. 19–24 (2024). Included in the Companion Article Pack

    work page 2024

  7. [7]

    Morandini, S. et al. The impact of artificial intelligence on workers’ skills: Upskilling and reskilling in organisations. Informing Sci. 26, 39–68, DOI: 10.28945/5078 (2023)

    work page doi:10.28945/5078 2023

  8. [8]

    Tushman, M. L. & Anderson, P. Technological discontinuities and organizational environments. Adm. Sci. Q. 31, 439–465 (1986)

    work page 1986

Show all 60 references
  1. [9]

    The technology–organization–environment framework

    Baker, J. The technology–organization–environment framework. In Dwivedi, Y . K., Wade, M. R. & Schneberger, S. L. (eds.) Information Systems Theory: Explaining and Predicting Our Digital Society, V ol. 1 , 231–245, DOI: 10.1007/ 978-1-4419-6108-2_12 (Springer, New York, NY , 2012)

  2. [10]

    Chen, Z. & Li, Y . The sharply decreasing disruptiveness of hci. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI ’25, DOI: 10.1145/3706598.3713917 (Association for Computing Machinery, New York, NY , USA, 2025)

    2025

  3. [11]

    Funk, R. J. & Owen-Smith, J. A dynamic network measure of technological change. Manag. science 63, 791–817, DOI: 10.1287/mnsc.2015.2366 (2017)

    2017

  4. [12]

    & Evans, J

    Wu, L., Wang, D. & Evans, J. A. Large teams develop and small teams disrupt science and technology. Nature 566, 378–382, DOI: 10.1038/s41586-019-0941-9 (2019)

    2019 doi

  5. [13]

    & Boudreau, J

    Jesuthasan, R. & Boudreau, J. Work without jobs. MIT Sloan Manag. Rev. 62, 1–5 (2021)

    2021

  6. [14]

    Autor, D. H. & Dorn, D. The growth of low-skill service jobs and the polarization of the us labor market. Am. economic review 103, 1553–1597 (2013)

    2013

  7. [15]

    H., Dorn, D

    Autor, D. H., Dorn, D. & Hanson, G. H. Untangling trade and technology: Evidence from local labour markets. The Econ. J. 125, 621–646, DOI: 10.1111/ecoj.12245 (2015). https://academic.oup.com/ej/article-pdf/125/584/621/26438955/ej0621. pdf

    2015 doi

  8. [16]

    Deming, D. J. The growing importance of social skills in the labor market*. The Q. J. Econ. 132, 1593–1640, DOI: 10.1093/qje/qjx022 (2017). https://academic.oup.com/qje/article-pdf/132/4/1593/30637898/qjx022.pdf

    2017 doi

  9. [17]

    A., Constantinides, M

    Septiandri, A. A., Constantinides, M. & Quercia, D. The impact of ai innovations on us occupations. arXiv preprint arXiv:2312.04714 (2023)

    2023 arXiv

  10. [18]

    The impact of artificial intelligence on the labor market

    Webb, M. The impact of artificial intelligence on the labor market. Available at SSRN 3482150 (2019)

    2019

  11. [19]

    Applying ai to rebuild middle class jobs

    Autor, D. Applying ai to rebuild middle class jobs. Tech. Rep., National Bureau of Economic Research (2024)

    2024

  12. [20]

    & Quercia, D

    Constantinides, M. & Quercia, D. AI, Jobs, and the Automation Trap: Where Is HCI? In Proceedings of the 4th ACM Annual Symposium on Human-Computer Interaction for Work (CHIWORK ’25) , 8 (2025)

    2025

  13. [21]

    & Michaels, G

    Graetz, G. & Michaels, G. Robots at work. Rev. economics statistics 100, 753–768 (2018)

    2018

  14. [22]

    & Restrepo, P

    Acemoglu, D. & Restrepo, P. Robots and jobs: Evidence from us labor markets. J. political economy 128, 2188–2244 (2020)

    2020

  15. [23]

    M., Quercia, D

    Bonaventura, M., Aiello, L. M., Quercia, D. & Latora, V . Predicting urban innovation from the us workforce mobility network. Humanit. Soc. Sci. Commun. 8, 1–9 (2021)

    2021

  16. [24]

    J., Anselin, L

    Acs, Z. J., Anselin, L. & Varga, A. Patents and innovation counts as measures of regional production of new knowledge. Res. policy 31, 1069–1085, DOI: 10.1016/S0048-7333(01)00184-6 (2002). 10/13

    2002 doi

  17. [25]

    Mikhaylov, A. S. Coastal agglomerations and the transformation of national innovation spaces. Balt. Reg. 11, 29–42, DOI: 10.5922/2079-8555-2019-1-3 (2019)

    2019 doi

  18. [26]

    Autor, D. H. & Handel, M. J. Putting tasks to the test: Human capital, job tasks, and wages. J. Labor Econ. 31, S59–S96, DOI: 10.1086/669332 (2013)

    2013 doi

  19. [27]

    H., Levy, F

    Autor, D. H., Levy, F. & Murnane, R. J. The skill content of recent technological change: An empirical exploration*. The Q. J. Econ. 118, 1279–1333, DOI: 10.1162/003355303322552801 (2003). https://academic.oup.com/qje/article-pdf/118/4/ 1279/5427313/118-4-1279.pdf

    2003 doi

  20. [28]

    Wilson, H. J. & Daugherty, P. R. Collaborative intelligence: Humans and ai are joining forces.Harv. Bus. Rev. 96, 114–123 (2018)

    2018

  21. [29]

    Davenport, T. H. & Kirby, J. Only humans need apply: Winners and losers in the age of smart machines (Harper Business New York, 2016)

    2016

  22. [30]

    Building the vital skills for the future of work in operations

    McKinsey & Company. Building the vital skills for the future of work in operations. McKinsey & Co. (2023). Accessed: 2024-08-20

    2023

  23. [31]

    Schumpeter, J. A. Capitalism. Social. Democr. (1942)

    1942

  24. [32]

    Should chatgpt be biased? challenges and risks of bias in large language models

    Ferrara, E. Should chatgpt be biased? challenges and risks of bias in large language models. arXiv preprint arXiv:2304.03738 (2023)

    2023 arXiv

  25. [33]

    & Moro, E

    Frank, M., Ahn, Y .-Y . & Moro, E. AI exposure predicts unemployment risk (2023). 2308.02624

    2023 arXiv

  26. [34]

    O*NET® Database Releases Archive

    National Center for O*NET Development. O*NET® Database Releases Archive. www.onetcenter.org/db_releases.html (2024). Accessed: 2024-08-08

    2024

  27. [35]

    Patent and Trademark Office

    U.S. Patent and Trademark Office. Data Download Tables. https://patentsview.org/download/data-download-tables. Accessed: [date]

  28. [36]

    Why 2015 was a breakthrough year in artificial intelligence

    Clark, J. Why 2015 was a breakthrough year in artificial intelligence. Bloom. News 8 (2015)

    2015

  29. [37]

    Wipo technology trends 2019: Artificial intelligence

    WIPO. Wipo technology trends 2019: Artificial intelligence. Geneva: World Intellect. Prop. Organ. (2019)

    2019

  30. [38]

    & Funk, R

    Park, M., Leahey, E. & Funk, R. J. Papers and patents are becoming less disruptive over time. Nature 613, 138–144, DOI: 10.1038/s41586-022-05543-x (2023)

    2023 doi

  31. [39]

    Ni, J. et al. Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models. arXiv preprint arXiv:2108.08877 (2021)

    2021 arXiv

  32. [40]

    & Lordan, G

    Josten, C. & Lordan, G. Automation and the changing nature of work. Plos one 17, e0266326 (2022)

    2022

  33. [41]

    Frey, C. B. & Osborne, M. A. The future of employment: How susceptible are jobs to computerisation? Technol. F orecast. Soc. Chang. 114, 254–280, DOI: https://doi.org/10.1016/j.techfore.2016.08.019 (2017)

    2017 doi

  34. [42]

    & Seamans, R

    Felten, E., Raj, M. & Seamans, R. Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses. Strateg. Manag. J. 42, 2195–2217, DOI: https://doi.org/10.1002/smj.3286 (2021). https: //onlinelibrary.wiley.com/doi/pdf/10.100...

    2021 doi

  35. [43]

    & Salomons, A

    Goos, M., Manning, A. & Salomons, A. Explaining job polarization: Routine-biased technological change and offshoring. Am. Econ. Rev. 104, 2509–26, DOI: 10.1257/aer.104.8.2509 (2014)

    2014 doi

  36. [44]

    S., Ales, L

    Combemale, C., Whitefoot, K. S., Ales, L. & Fuchs, E. R. H. Not all technological change is equal: how the separability of tasks mediates the effect of technology change on skill demand. Ind. Corp. Chang. 30, 1361–1387, DOI: 10.1093/icc/ dtab026 (2021). _eprint: https://academ...

    2021 doi

  37. [45]

    & Shirota, T

    Kikuchi, S., Fujiwara, I. & Shirota, T. Automation and disappearing routine occupations in japan. J. Jpn. Int. Econ. 74, 101338, DOI: https://doi.org/10.1016/j.jjie.2024.101338 (2024)

    2024

  38. [46]

    & SRIDHAR, V

    UPRETI, A. & SRIDHAR, V . The dynamics of task automation and worker adjustment in labor markets: An agent- based approach. Adv. Complex Syst. 25, 2250005, DOI: 10.1142/S0219525922500059 (2022). https://doi.org/10.1142/ S0219525922500059

    2022 doi

  39. [47]

    & Viollaz, M

    Egana-delSol, P., Bustelo, M., Ripani, L., Soler, N. & Viollaz, M. Automation in latin america: Are women at higher risk of losing their jobs? Technol. F orecast. Soc. Chang.175, 121333, DOI: https://doi.org/10.1016/j.techfore.2021.121333 (2022)

    2022

  40. [48]

    & Biagetti, M

    Vannutelli, S., Scicchitano, S. & Biagetti, M. Routine-biased technological change and wage inequality: do workers’ perceptions matter? Eurasian Bus. Rev. 12, 409–450, DOI: 10.1007/s40821-022-00222-3 (2022). 11/13

    2022 doi

  41. [49]

    Aggregation, technological change and the skill structure of the british economy 1951-1991

    Hoskins, M. Aggregation, technological change and the skill structure of the british economy 1951-1991. Appl. Econ. Lett. 9, 251–254, DOI: 10.1080/13504850110060376 (2002). https://doi.org/10.1080/13504850110060376

    2002 doi

  42. [50]

    Black, S. E. & Spitz-Oener, A. Explaining women’s success: Technological change and the skill content of women’s work. The Rev. Econ. Stat. 92, 187–194 (2010)

    2010

  43. [51]

    & Kambourov, G

    Caines, C., Hoffmann, F. & Kambourov, G. Complex-task biased technological change and the labor market. Rev. Econ. Dyn. 25, 298–319, DOI: https://doi.org/10.1016/j.red.2017.01.008 (2017). Special Issue on Human Capital and Inequality

    2017 doi

  44. [52]

    Taking technology to task: The skill content of technological change in early twentieth century united states

    Gray, R. Taking technology to task: The skill content of technological change in early twentieth century united states. Explor. Econ. Hist. 50, 351–367, DOI: https://doi.org/10.1016/j.eeh.2013.04.002 (2013)

    2013 doi

  45. [53]

    Tschang, F. T. & Almirall, E. Artificial intelligence as augmenting automation: Implications for employment. Acad. Manag. Perspectives 35, 642–659, DOI: 10.5465/amp.2019.0062 (2021)

    2021

  46. [54]

    & Reeson, A

    Evans, D., Mason, C., Chen, H. & Reeson, A. Accelerated demand for interpersonal skills in the Australian post-pandemic labour market. Nat. Hum. Behav. 8, 32–42, DOI: 10.1038/s41562-023-01788-2 (2024)

    2024 doi

  47. [55]

    & Mezzanzanica, M

    Colombo, E., Mercorio, F. & Mezzanzanica, M. Ai meets labor market: Exploring the link between automation and skills. Inf. Econ. Policy 47, 27–37, DOI: https://doi.org/10.1016/j.infoecopol.2019.05.003 (2019). The Economics of Artificial Intelligence and Machine Learning

    2019 doi

  48. [56]

    Montobbio, F., Staccioli, J., Virgillito, M. E. & Vivarelli, M. Labour-saving automation: A direct measure of occupational exposure. The World Econ. 47, 332–361 (2024)

    2024

  49. [57]

    D., Guillaume, J.-L., Lambiotte, R

    Blondel, V . D., Guillaume, J.-L., Lambiotte, R. & Lefebvre, E. Fast unfolding of communities in large networks. J. statistical mechanics: theory experiment 2008, P10008, DOI: 10.1088/1742-5468/2008/10/P10008 (2008)

    2008 doi

  50. [58]

    Prolific

    Prolific. Prolific. https://www.prolific.co/ (2024). Accessed: 2024-08-14

    2024

  51. [59]

    & Rock, D

    Eloundou, T., Manning, S., Mishkin, P. & Rock, D. Gpts are gpts: Labor market impact potential of llms. Science 384, 1306–1308, DOI: 10.1126/science.adj0998 (2024)

    2024 doi

  52. [60]

    The Potential Impact of Disruptive AI Innovations on U.S. Occupations

    Josten, C. & Lordan, G. Robots at work: Automatable and non-automatable jobs (Springer, 2020). 12/13 Supplementary Information for “The Potential Impact of Disruptive AI Innovations on U.S. Occupations” Results of Classifying Patents Representative Patent Disruption Index Scor...

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.