Exposure of occupations to technologies of the fourth industrial revolution

1. Introduction

• Automation can reduce demand for some tasks while raising demand for others and creating new tasks, leaving its overall labor-market effect dependent on the relative size and interaction of these effects.Previous automation waves did not result in long-lasting technology-induced rising unemployment.Source: §1 · p. 2
• Patent-based exposure is intended to reflect actual technology maturity and diffusion, unlike measures of theoretical automation potential.Patents document what existing technology can currently do.Source: §1 · p. 2
• Mapping patents to tasks captures differences in exposure across activities within an occupation and can match patents to activities more accurately than occupation-wide text.For example, robot-engineer tasks such as planning paths, debugging programs, and maintaining robots can be distinguished from general robot patents.Source: §1 · p. 2
• The paper provides task- and occupation-level exposure scores that distinguish 4IR patents from other patents and use patent data since 1970 to examine developments over time.The mapping addresses calls for better data on technological progress in key 4IR technologies.Source: §1 · p. 3

2. Patents as an indicator for technological change

• Industry-level concordances do not provide fine-grained occupation mappings, and industry descriptions can emphasize outputs rather than workers’ activities.The cited datasets are either based on 1990–1993 data or use industry-specific terms such as “manufacture cars” rather than tasks such as “cut sheet metal.”Source: §2 · p. 3
• Webb uses verb-noun pairs and word hierarchies to match patent language to tasks, whereas Kogan et al. use text embeddings and full-text context to compare patents with occupation descriptions.The embedding approach can distinguish phrases such as “diagnose patients condition” and “diagnose machine condition.”Source: §2 · p. 3
• Matching patents to tasks separately can distinguish technologies relevant to different activities and reduce the dominance of general context or frequently occurring tasks in occupation-wide text.A barber’s “sell lotions” task may otherwise be overshadowed by descriptions dominated by hair-related terms.Source: §2 · pp. 3–4

3. Mapping patents to occupations

• The analysis compares patent abstracts with O*NET task descriptions, removing references to tools so matching focuses on the activity rather than the technology currently used.For example, the task “Monitor geothermal operations, using programmable logic controllers” is reduced to monitoring geothermal operations.Source: §3 · p. 4, §3.1 · p. 4, Figure 1 · p. 5

3.1. Data

• The PATSTAT sample contains 15,317,445 English-abstract patent families filed from 1970 to 2020, with the most valuable insights expected for 1970–2015 because of publication delays.Only one patent per family is selected to avoid double counting.Source: §3.1 · p. 4

3.2. Patent task similarity scores

• Task-patent similarity combines TF-IDF word weights with embedding-based word similarity, using detailed work activities to reduce context bias while retaining task-specific weighting.A separate patent-trained embedding helps identify technical language, and the two embeddings are used together to expand coverage while reducing noise.Source: §3.2 · pp. 5–7
• 75% of patents in the sample were mapped to at least one task, and two embedding approaches produced identical overall patterns, supporting robustness to embedding choice.Combining the approaches broadened patent coverage while maintaining a high cutoff to reduce irrelevant matches.Source: §3.2 · p. 7

3.3. Patents of the fourth industrial revolution

• A patent is classified as 4IR when any of its patent classifications falls within the 4IR classes identified by M´eni`ere et al. (2020).Those classes span application domains and technology fields.Source: §3.3 · pp. 7–8

3.4. Aggregation of results

• Occupation exposure aggregation uses log patent counts, weights task scores by O*NET task importance, and aggregates occupations before clusters to limit bias from differing task counts.Tasks can be assigned to multiple task categories when their category z-scores are positive.Source: §3.4 · p. 8

4. Patent task mapping results

• The mapping associates 11,343,380 patents—74% of all patents in the dataset—with tasks; chemistry and biotechnology patents map to few occupations, while computer technology and IT-management patents are highly relevant to many tasks.Example results indicate a good match between task descriptions and patent results, but may include patents capturing the same detailed work activity (DWA) rather than directly related to the task.Source: §4 · p. 8

4.1. Patents over time

• 4IR patents drove strong patent growth from 1990 to 2000; their share then declined, explaining the 2000–04 plateau while non-4IR patent counts continued growing.4IR patents made up 29% of patents in the results, ranging from 12% in 1976 to 38% in 2016.Source: §4.1 · p. 9, Figure 2 · p. 9

4.2. Patents per task

• Most patents map to around 30–60 unique task/DWA combinations, while patent counts per task range from about 2,000 to over 500,000.The task-count distribution is log-normal, with most tasks associated with 32,000 patents; the analysis uses log counts to limit bias from high-count tasks.Source: §4.2 · p. 10

4.3. Evaluation per task type

• Non-routine manual tasks have the most related patents but the lowest 4IR exposure, while analytical tasks’ patent exposure grew strongly through a high 4IR share.Routine cognitive exposure to breakthrough patents increased and routine manual exposure decreased, unlike their 4IR exposure patterns.Source: §4.3 · pp. 10–11

5. Patent exposure score

• Patent-based exposure addresses the gap between technical feasibility and diffusion that existing automation-potential measures do not address.The measure relies on patent data, whose granting criteria include novelty and usefulness.Source: §5 · p. 11

5.1. Description of exposure scores

• Occupation-level 4IR scores weight task exposure by task importance; scores range from below three to above six, with computer and information activities among the high-exposure examples.Interacting with computers and recording information score highly, while staffing organizational units and negotiating score lowest.Source: §5.1 · p. 12
• Low-skilled occupations have lower average 4IR exposure than medium- and high-skilled occupations, although their exposure to overall patents is higher.4IR exposure grew fastest during the 1990s, especially for medium- and high-skilled occupations; from 2007 to 2013 it grew faster among occupations with lower current scores.Source: §5.1 · p. 12

5.2. Patent exposure sub-scores and non-4IR exposure

• The mapping supports exposure sub-scores for each of 367 defined 4IR technologies, including categories such as healthcare, software, and user interfaces.Source: §5.2 · p. 12
• 4IR and non-4IR exposure are inversely related: construction, manufacturing, transportation, agriculture, and hospitality are mainly exposed to traditional technologies, while marketing, finance, administration, education, and law show relatively strong 4IR exposure.Source: §5.2 · p. 13

6. Impact of technologies on the labor market and patent exposure

• The section frames labor-market effects as more than routine replacement: automation, deepening, reinstatement, productivity, and capital accumulation interact, so automation potential need not mean net employment loss.Prior studies cited in the section report that productivity effects can overcompensate replacement effects.Source: §6 · p. 13

6.1. Patent exposure per wage percentile and education

• Non-4IR exposure is highest among middle-wage and low-education skilled occupations, while 4IR exposure is highest for medium-to-high-skilled occupations and correlates with mean income.The wage-percentile comparison is shown in Figure 5.Source: §6.1 · p. 14, Figure 5 · p. 15
• 4IR-manufacturing exposure resembles non-4IR exposure across wage percentiles, software resembles overall 4IR exposure, and AI is more skewed toward high-wage occupations.The paper presents these differences as reasons 4IR sub-scores can support more detailed assessments.Source: §6.1 · p. 14

6.2. Patent exposure and job growth

• Patent exposure has an overall negative, concave association with job growth; exposure measured in 1992 has the highest coefficients for 2012–2018 job growth.The authors say the occupation-level findings require confirmation with microdata and do not account for all labor-market effects of automation.Source: §6.2 · p. 15
• Because occupation-level estimates need confirmation with microdata and omit other automation effects, the overall employment impact may differ from the observed negative relationship.The authors identify deepening automation and capital accumulation as examples of effects not included in the regression.Source: §6.2 · p. 15

6.3. Patent exposure scores as indicators for labor market analysis

• Webb’s and Kogan’s indicators overlap with some 4IR patterns, but their categories and scopes differ; the task-level mapping also accounts for task variety within occupations.Kogan’s breakthrough-patent scope changes over time, whereas this paper’s 4IR score refers to selected 4IR patents.Source: §6.3 · p. 16
• The patent-occupation mapping can support exposure analyses beyond 4IR, including AI, software, robot, and breakthrough patents when technology concordances are available.The authors note that different studies’ exposure scores are difficult to compare because they have different focuses.Source: §6.3 · p. 16
• Historical exposure scores based on current task descriptions may not reflect occupations’ past tasks, so the authors consider them most useful for recent labor-market changes.For the 2012–2018 job-growth analysis, they instead use 2012 task descriptions.Source: §6.3 · p. 17

7. Direct comparison of 4IR exposure and other technology and automation indicators

• 4IR technologies include more than machine learning, spanning core IT hardware, software, and connectivity as well as applications such as predictive maintenance and smart ATMs.The section compares this broad exposure measure with indicators of AI exposure, machine-learning suitability, computerization probability, and automation potential.Source: §7 · pp. 17, 21
• The comparisons distinguish patent-based exposure to existing technological capabilities from forward-looking AI exposure and measures of automation potential.The section notes that existing exposure measures do not distinguish task replacement, reinstatement, and augmentation.Source: §7 · p. 17

7.1. Comparison of 4IR exposure and AIOI index by Felten et al.

• 4IR patent share correlates more strongly with AIOE (r = 0.74) than 4IR exposure does (r = 0.58).Patent share reflects the proportion of related patents that are 4IR patents; exposure also reflects the total number of related patents.Source: Figure 6 · p. 21, §7.1 · pp. 21–22

7.2. Combining 4IR exposure score and suitability for machine learning

• 4IR exposure and SML scores are correlated (r = 0.36), but their quadrants distinguish patent exposure from automation suitability.Quadrant I combines high exposure and suitability; Quadrant II has high suitability but few patents, while Quadrant IV has high exposure but low suitability.Source: §7.2 · pp. 22–24, Figure 7 · p. 23
• Low-education occupations have a stronger 4IR-exposure/SML correlation than others (r = 0.40 versus r = 0.29), and few appear in the potentially augmenting Quadrant IV.The analysis also suggests that Quadrants II and III experienced more 4IR exposure growth since 2001 than Quadrants I and IV.Source: §7.2 · p. 24

7.3. Comparison of exposure scores to computerization probabilities by Frey and Osborn

• CP scores correlate weakly with 4IR exposure (r = -0.13) but more strongly with non-4IR exposure (r = 0.50).High-CP occupations are often low-skilled, with administrative occupations more exposed to 4IR patents and manual occupations more exposed to non-4IR patents.Source: §7.3 · p. 25, Figure 8 · p. 26

7.4. Comparison of exposure scores to MGI automation scores.

• Non-4IR patent exposure correlates with higher MGI automation potential (r = 0.64), while medium-to-high 4IR exposure occupations generally have lower potential.Very high 4IR exposure is an exception; some such occupations may be among the first to feel 4IR automation impacts.Source: §7.4 · pp. 26–27, Figure 9 · p. 27

7.5. Comparison of 4IR exposure and Webb’s exposure scores

• Webb’s AI exposure scores do not correlate with 4IR patent exposure, while software exposure correlates only slightly with non-4IR patent exposure.Differences in patent-task mapping, technology definitions, and task-to-occupation aggregation may explain the discrepancies.Source: §7.5 · p. 27

8. Conclusions and future work

• Figure 10 compares patent exposure z-scores with Webb (2019) percentile scores and reports Pearson correlations of 0.05, 0.44, and -0.84 in its first row and 0.05, 0.71, and 0.42 in its second row.Bubble sizes indicate numbers of jobs.Source: Figure 10 · p. 28
• The indicator maps patents to tasks and occupations and distinguishes 4IR patents from other patents, refining prior approaches to account for task-level exposure and activity context.Its scores reflect actual technology diffusion rather than theoretical automation potential.Source: §8 · pp. 27–28
• Manual-task occupations such as manufacturing and construction have high non-4IR exposure and low 4IR exposure, while many non-manual occupations have a higher 4IR exposure ratio.The paper notes that higher exposure scores may indicate that occupations are more impacted by 4IR technologies.Source: §8 · p. 28
• 4IR exposure has a negative, concave relation to job growth, with higher coefficients for patent exposure 10 and 20 years earlier than for more recent exposure.The authors say further microdata analysis is needed to confirm the findings and more complex modeling to estimate overall labor-market impact.Source: §8 · p. 29
• The mapping provides technology-specific exposure scores and can support other exposure measures when a patent-technology mapping is available.Examples include CAD, augmented reality for surgery, smart office technologies, and measures based on robots or breakthrough patents.Source: §8 · p. 29
• The indicator counts patent exposure, but future work could account for differing patent impact; the validity of patent citations as a proxy for impact is ambiguous.The authors also point to breakthrough patents as a possible alternative approach.Source: §8 · p. 29
• The indicator uses O*Net descriptions situated in the US labor market; regional concordance tables may support its use in other contexts when similarly structured regional data exist.Future work could also use O*Net information about technologies and tools used in tasks to study labor augmentation.Source: §8 · p. 29

Competing interests

• The authors declare no significant competing financial, professional, or personal interests that might have influenced the work.Source: Competing interests · p. 29

Appendix A. Patents per year in our dataset

• Patent counts in the dataset increased until 2013, then declined; the dataset includes patents with application dates between 1970 and 2019.The decline may reflect unpublished patents or patents not yet included in PATSTAT.Source: Figure A.11 · p. 32
• The post-2013 decline may be due to unpublished patents or patents not yet included in PATSTAT.Source: Figure A.11 · p. 32
• Dips in patenting activity are likely related to economic decline, with the Dotcom crisis after 2000 given as an example.Source: Figure A.11 · p. 32

Appendix B. Validity check through comparing both word embeddings

• The patent-specific embedding identifies almost twice as many patents as the general-language embedding, and only 35% of general-language patent-task relations overlap with patent-specific relations.Source: Appendix B · p. 32
• The authors report that results by technology sector differ only slightly between embeddings and show no indication of bias toward particular task types.They interpret these patterns as evidence that embedding choice did not systematically bias the results.Source: Appendix B · p. 32, Table B.5 · p. 33, Table B.6 · p. 33
• The comparison checks whether embeddings bias results toward particular technology sectors or task types.Source: Appendix B · p. 32

Appendix C. Patent per technology fields and task type

• IT-management patents are relevant to most tasks; control, computer, and digital communication technologies are frequently linked to occupations, while chemistry and nanotechnology rarely link directly to tasks.These technology-cluster results are presented as validation of the patent-to-task mapping.Source: Appendix C · p. 34
• Tasks are grouped into four broad activity categories, and the analysis measures each technology field’s share of patents associated with each category.The figure includes only technology fields representing more than 1% of patents, so not all columns sum to 100%.Source: Appendix C · p. 34, Figure C.12 · p. 34
• Information-input tasks have a particularly high share of measurement patents; mental-process and interaction tasks link strongly to computer, IT, communication, and control patents.Work-output tasks have high shares of machine-tool and civil-engineering patents; these links match expectations and suggest the mapping is valid.Source: Appendix C · p. 34

Appendix D.1. Example exposure of occupation to 4IR sub-scores

• The example figure shows occupation exposure to 4IR technology sub-scores; shading indicates each occupation’s share of a technology’s total exposure.For example, the shading indicates what share of CAD patents is related to designers.Source: Figure D.13 · p. 35

Appendix E. Exposure to non-4IR patents

• The appendix includes a figure comparing exposure to 4IR and non-4IR patents, but the supplied caption gives no comparison outcome.Source: Figure E.14 · p. 36

• Non-4IR patent exposure is inversely correlated with education level, whereas 4IR exposure is highest for medium-to-high education levels.Occupations are aggregated by education level using their share of total jobs in BLS data.Source: Appendix F · p. 37, Figure F.15 · p. 37
• The education pattern for non-4IR exposure resembles Webb’s robot and software patent findings, while 4IR exposure peaks at medium-to-high education levels.Webb’s AI patent exposure is higher for higher-educated occupations.Source: Appendix F · p. 37

• Tables G.7 and G.8 examine historic 4IR patent exposure alongside employment change from 2012 to 2018.The supplied captions do not report estimates or outcomes.Source: Table G.7 · p. 38, Table G.8 · p. 39
• The accompanying table notes describe analyses of 720 or 704 observations using BLS labor-market data from 2011 and 2018.Controls IE denotes inclusion of industry and education (Job Zone) controls; significance markers are defined in the notes.Source: Appendix G · pp. 38–39
• Table G.9 examines 4IR patent exposure and employment change from 2012 to 2018 without control variables.Source: Table G.9 · p. 39

Appendix G.1. Comparison of SML and CP scores

• The comparison of CP and SML scores reports a Pearson’s r of 0.078 based on single observations.Scores are grouped by SOC career clusters in the graph for readability.Source: Figure G.16 · p. 40, Appendix G.1 · p. 40

Appendix G.2. 4IR exposure sub-scores per wage percentile

• A figure presents 4IR exposure sub-scores by wage percentile, but the supplied caption reports no pattern or comparison outcome.Source: Figure G.17 · p. 41

Appendix H. 4IR exposure scores

• The method maps patents to tasks to produce task-level exposure scores, then aggregates those scores into occupation-level 4IR exposure scores.The section provides an overview of exposure scores at both levels.Source: Appendix H · p. 42

Appendix H.1. Task- and activity-level 4IR exposure scores

• The appendix reports scores for more than 20,000 tasks, with examples at task and aggregated work-activity levels.Source: Appendix H.1 · p. 42
• Table H.10 gives highest and lowest 4IR exposure scores by work activity, aggregated from task-level scores.Source: Table H.10 · p. 42, Appendix H.1 · p. 42
• Table H.11 presents exposure scores for randomly selected detailed work activities.Source: Table H.11 · p. 42, Appendix H.1 · p. 42