1、2 Generative AIs Influence on Employment Patterns June 2024 Mar Carpanelli Staff Data Scientist LinkedIn Matthew Baird Senior Staff Economist LinkedIn Cristian Jara-Figueroa Senior Data Scientist LinkedIn*Generative AIs Influence on Employment Patterns Economic Graph White Paper June 2024 In this st

2、udy,we investigate the differential impacts of Generative Artificial Intelligence(GAI)on workers in the US using LinkedIn data,both overall and by their educational attainment levels.We find a gradual shift away from GAI-disrupted occupations over the past six years towards potentially augmented or

3、insulated roles,particularly pronounced among higher-educated workers.Through an analysis of occupational transitions,we make predictions under various scenarios of GAI impact,examining the proportions of workers within different GAI occupation classifications,transitions from employment to non-empl

4、oyment,and shifts in occupational categories.While workers with higher educational attainment tend to exhibit higher rates of occupational mobility,the differential effects of GAI on employment and occupational changes underscore the importance of considering educational disparities in the context o

5、f technological advancements.*Currently at Cash App.All work was done while at LinkedIn.3 GAIs Influence on Employment Patterns Economic GraphIntroduction Generative AI(GAI)tools,such as Copilot,ChatGPT,and Gemini,have the potential to fundamentally change how people work and their productivity.75%o

6、f knowledge workers using AI at work and 79%of work leaders stating their company needs to adopt AI to stay competitive(Microsoft&LinkedIn,2024).However,not all workers will be equally impacted.In our prior work,we classified each occupation to one of three groups,based on their skill composition:th

7、ose likely to be insulated from the impacts of GAI,those likely to be augmented,and those likely to be disrupted.Our prior report found that GAI is likely to have unequal effects on workers,depending on their skills level(Kimbrough&Carpanelli,2023).Unlike previous technological advancements that pri

8、marily affected workers in lower-skilled roles or in jobs with lower-education requirements,the GAI technological wave will likely affect some of the highest-skilled and highest-educated jobs as well as those with lower educational attainment.There is a rich history of research evaluating the impact

9、 of technological development on workers.For example,while wage inequality has grown over time,and many have attributed this to technological innovation which favors the more skills,higher-paid workers,Card&DiNardo(2002)argue against this hypothesis by examining the data and underlying models.They s

10、uggest the need for a much more nuanced understanding of changes in wage inequality as well as the impact of technology on different groups of workers(see also Acemoglu&Autor,2011;Card&DiNardo,2006).However,automation can indeed exacerbate inequality,especially if the speed of adoption is fast and i

11、s accompanied by the creation of new tasks(Acemoglu&Restrepo,2018).While it is still early days for GAI,there is a fast-growing literature on how this technology may or has impacted employment outcomes across different groups.Based on an occupational task analysis,Eloundou et al.(2023)estimate that

12、80%of US workers could have at least 10%of their tasks automated by large language models,and 19%of these workers could have more than SkillsSkills-based GAI occupation classificationbased GAI occupation classification (GAI(GAI-GroupGroup)Augmented occupationsAugmented occupations are those which us

13、e many skills that are complemented by GAI.For example,software engineers may automate some of their coding work with GAI,focusing more of their time on GAI-complementary skills,such as cross-functional influencing and stakeholder engagement.Examples:Software engineer,data analyst,web designer,nutri

14、tion assistant Disrupted occupationsDisrupted occupations are those which stand to see significant change from GAI,but do not rely as much on GAI-complementary skills.For instance,language translators skills stand to shift from doing translations from scratch to reviewing and certifying machine-gene

15、rated translations,or to specializing on specific legal or literary domains.Examples:Customer service representative,administrative assistant,legal associate Insulated occupationsInsulated occupations are those that have a relatively small proportion of GAI-replicable skills in their core skills.For

16、 example,real estate agents might utilize GAI for writing house descriptions,but core relationship management skills would be insulated from GAI.Examples:teacher,nurse,locksmith Source:Kimbrough&Carpanelli(2023)4 GAIs Influence on Employment Patterns Economic Graph50%of tasks automated.Moreover,thei

17、r work suggests that higher-income workers are particularly exposed to such automation.A separate paper by Felten et al.(2023)agree,and expands the demographic impact analysis to show that highly-educated,highly-paid,white-collar occupations may be most exposed to generative AI.To expand current und

18、erstanding,in this report we focus on the impacts of GAI on workers job outcomes,both overall and depending on their educational attainment.We leverage recent estimates of occupational transitions using data from LinkedIn members profiles,limiting our attention to US workers.The methodologies for sc

19、oring occupations GAI complementarity and replicability with occupations skills,for classifying occupations into groups based on those scores,and for simulating the predicted impacts of GAI on different groups under different scenarios are all described in the appendix.We examine three outcomes:1.Pr

20、oportion of members working within each GAI occupation classification(GAI group)2.Proportion of members who transition from working to not working(which includes retirement and other voluntary separations as well as involuntary separations such as lay-offs)3.Proportion of members who change occupati

21、ons Key findings Historical trends and current ratesHistorical trends and current rates On aggregate,over the past six years,workers have already been trending away from occupations classified as disrupted by GAI,and into occupations potentially augmented or insulated from GAI.While all education gr

22、oups have been moving away from GAI-disrupted occupations,bachelors degree holders had the fastest decline in the share of workers in disrupted roles.Compared to workers with lower educational attainment(high school or associates degrees),higher education workers(bachelors and graduate degree)are mo

23、re likely to be in augmented and insulated occupations,and less likely to be in disrupted occupations.On average,the share of workers in augmented occupations has been increasing.However,workers with higher educational attainment have been trending towards jobs in augmented occupations at the fastes

24、t rate.Workers with higher educational attainment have tended to exit employment at a lower rate but change occupations at a higher rate than lower education workers.We next examine predictions under five different scenarios with respect to the extent to which GAI complements or replicates skills.Th

25、is allows us to investigate potential employment outcomes across the next year depending on GAI impact patterns without making concrete claims to what those GAI impact patterns will be.3 GAIs Influence on Employment Patterns Economic Graph We summarize the main findings of our paper under each scena

26、rio.1 Scenario 1:Status quoScenario 1:Status quo Our baseline from pre-GAI trends,which assumes that the best predictor for occupation transitions and our outcomes is the year before GAI.Scenario Scenario 2 2:Gradual integrationGradual integration If both skills complementarity and skill replicabili

27、ty are only lightly impacted,we predict very small changes relative to status quo.1 Our five scenarios are described by the intensity of impact on complementary and replicability of skills.We acknowledge that what will occur will be more nuanced,with for example the intensity of the impact varying a

28、cross industries,occupations,and times.This may be due to speed of adoption of GAI,barriers to usage,and other factors.It is outside of the scope of this paper to test differentiation of scenarios across the economy or across time.There are very small increases out of disrupted occupations and into

29、insulated occupations and not working.Scenario Scenario 3 3:Heightened exposureHeightened exposure All education groups see a decrease in the share of workers in augmented and disrupted occupations and an increase in the share of workers in insulated occupations.Bachelors degree holders widen their

30、advantage over high school graduates in employment in augmented occupations,and high school graduates become proportionally more likely to work in insulated occupations.The higher likelihood of high school graduates to work in disrupted occupations than bachelors graduates drops substantially due to

31、 an outsized decrease in participation by high school graduates.This is the scenario with the largest increase in the share of members not working.Because the increase from GAI is proportionally bigger for bachelors degree graduates than high school graduates,the gap between these two groups narrows

32、 from 30.4%higher under status quo(2.6%of high school graduates transitioning to non-work the following year,compared to 2.0%of bachelors degree holders)status quo,to 9.2%higher under this scenario(5.4%of high school graduates,compared to 5.0%of bachelors graduates).People switch occupations year-to

33、-year more than under status-quo,which tends to widen even more for higher education workers(e.g.,Potential scenariosPotential scenarios Scenario 1(Status quo):Scenario 1(Status quo):GAI has no impact on demand for workers and observed employment patterns.Scenario 2(Gradual integration):Scenario 2(G

34、radual integration):The impact of GAI is low on both skill complementarity and skill replicability.Scenario 3(Heightened exposure):Scenario 3(Heightened exposure):There is a large impact on skill replicability,exposing occupations which use such skills.The impact on skill complementarity is low.Scen

35、ario 4(Broad augmentation):Scenario 4(Broad augmentation):There is a large impact on how GAI complements skills,and occupations which employ such skills see increases in productivity and demand for workers.The impact on skill exposure is low.Scenario 5(Paradigm shift):Scenario 5(Paradigm shift):Ther

36、e are large impacts of GAI arising from both skill complementarity and skill exposure.Demand for occupations changes depending on their use of complemented or exposed skills.4 GAIs Influence on Employment Patterns Economic Graphunder status quo high school graduates switch occupations 12.3%less than

37、 bachelors degree holders,while under this scenario they switch 14.6%less).Scenario Scenario 4 4:Broad augmentationBroad augmentation The largest share of workers in augmented occupations are under this scenario across education groups,benefiting lower education workers the most(narrowing existing h

38、igh school/bachelors gap from-8.8%to-5.4%).Participation in disrupted occupations decreases for all education groups,but by less than under scenarios 3 and 5.High school graduates higher share working in disrupted occupations compared to bachelors decreases.For all education groups,there is a decrea

39、se in the share not working,and disproportionately for lower education workers.Under status quo,high school graduates are 34.6%more likely to be not working than bachelors graduates.Under scenario 3,that drops to 12.8%.Only small increases in occupation-switching predicted for all education groups w

40、ith no appreciable changes in gaps.This is the only scenario where we have a predicted decrease in the share who exit employment compared to status quo.The high school/bachelors gap decreases somewhat.Scenario Scenario 5 5:Paradigm shiftParadigm shift For all education groups,an increase in workers

41、in augmented occupations relative to status quo,but slightly more for lower education workers.This is the scenario with the lowest participation in disrupted occupations for all education groups,but a large shift in the gaps.Under status quo,high school graduates are predicted to be 7.8%more likely

42、to be in disrupted occupations than bachelors degree holders,but under this scenario,the gap shrinks to 0.8%.Participation in insulated occupations increases for all groups,but more so for lower education workers.High school graduates are predicted to move from 1.3%less likely than bachelors degree

43、holders to work in insulated occupations to 4.0%more likely.We predict slight increases in the proportion exiting employment in all groups,but more so for higher education.The high school/bachelors gap drops in half,from 30.4%to 14.0%.This is the scenario with the most occupation changing,but relati

44、vely similarly across education groups,with only small changes in gaps.Modeling projected gaps In order to explain the intuition behind our approach,we first present one outcomethe share of people in each of the GAI groupsfor the overall population(that is,not by education).Figure 1 presents the est

45、imated shares of workers in each group each month over the last six years.In order to calculate how groups have transitioned between GAI groups over time,we examined the occupations each month,and 5 GAIs Influence on Employment Patterns Economic Graphmap this according to the 2023 categorizations in

46、 Kimbrough and Carpanelli(2023).2 We find over the past six years that workers have been shifting away from jobs that we would in 2023 classify as disrupted by GAI(exemplified by the downward slope on the green line),and shifting towards augmented jobs(see text box on page 1 for examples).Insulated

47、jobs have remained relatively stable,with a highpoint at the start of 2021,potentially due to pandemic restrictions.We next examine US LinkedIn member data from February 2022 through 2024 on US members.We use changes in employment status and which occupation group(of over 600)each person reports wor

48、king in from each month in 2022 to the same month in 2023 for example.For each rolling 12-month span,we calculate the 2 In sensitivity analysis,we test this on a sample of members who have belonged to LinkedIn since at least January 2018,in Appendix Figure A.1.This allows GAI groups.The results are

49、qualitatively similar.share of people in employment status or occupation i who are not working or work in each of the occupations 12 months later.We averaged these ratios across all 12 months to create predictions of how workers will transition in the futurea Markov Chain transition matrix.Note that

50、 we also tested a model where this transition matrix was estimated not just over the period from 2022 to 2023 each month,but over the past six years.The results were very similar,and so in this paper we only report the results from the 2022 to 2023 transition matrix.The appendix contains more detail

51、s on the methodology.From this,we can calculate any number of occupational choice decisions across a one-year period.We take the same matrix generated above of each transition and use it to predict what fraction would be in each occupation one year later,if the same trends continued.This forms Scena

52、rio 1(Status quo),and would intuitively approximately represent a continuation of the above trend line.We then examine four alternative scenarios.These are briefly summarized in the text box above and the assumptions mathematically are explained in the appendix.Here we provide a little more discussi

53、on into the nature of each scenario.Scenario 2(Gradual integration)explores a scenario where GAI does have impacts on how skills are used,but the impact is low for both skill complementarity and skill replicability.In other words,estimates and predictions for worker occupational transitions are base

54、d on the assumption that the impact of both GAI skill Figure 1 Share of members in each GAI group 6 GAIs Influence on Employment Patterns Economic Graphcomplementarity and exposure to GAI-replicable skills is low.In other words,a job with skills which are complemented by GAI but no skills replicated

55、 by GAI would see a slight increase in demand.A job with skills replicated by GAI but none complemented would see a slight decrease in demand.A job with both complemented and replicated skills would have uncertain impacts on the share working in each occupation,and may not change at all.Scenario 3(H

56、eightened Exposure)represents a scenario where the impact is low from GAI skill complementarity,but high from GAI-replicable skill replicability.Estimates and predictions for worker occupational transitions are based on the assumption that the impact of GAI skill complementarity is lowhaving very li

57、ttle impactbut that the impact of GAI-replicable skills is high,which could decrease overall demand for workers in these occupations.This would generally be considered a bad case scenario for the labor market,although as we discuss below,there are some disparities between lower and higher education

58、workers that end up narrowing under this scenario.Scenario 4(Broad augmentation)presents a case where the Impact is high from GAI skill complementarity,but low from GAI-replicable skill exposure.Occupations that utilize these GAI-complementary skills thus see an increase in demand,and there is wider

59、 increases in worker productivity with likely increases in overall employment as a result as well.Scenario 5(Paradigm shift)shows the case when the impact is assumed to be high from GAI skill complementarity as well as from GAI-replicable skill exposure.As suggested by the name,this would lead to th

60、e most dramatic shifts in demand for different occupations(some increasing,others decreasing)depending on which skills are utilized in the job,and how they are complemented or replicated by GAI.These scenarios are predicted by shifting the empirically estimated probabilities of transitioning from oc

61、cupation i to occupation j(for each occupation pair)based on occupation i and occupation js scores on GAI-replicable skills and GAI-complementary skills(Kimbrough&Carpanelli,2023).While we explain this in more detail in the appendix,intuitively,demand for occupation i decreases by a little if it has

62、 a below average score of GAI-complementarity under the scenario of low impact from complementarity,and a lot under the scenario of high impact.Demand for occupation i decreases if it has an above average score on GAI-replicability in the same way.Overall predicted changes GAI groupings We first pre

63、dict how the share of people working in each GAI group would change under the different scenarios.Figure 2 demonstrates the predicted shares of the share of workers in each of the three GAI groups under the four scenarios one year from now,contrasted with status quo.Consider for example the share in

64、 augmented occupations.Under the status quo scenario,the share of workers would be predicted to grow gradually(reflecting the trend over the past few 7 GAIs Influence on Employment Patterns Economic Graphyears observed in Figure 1),and result in 25.6%of the sample working in augmented occupations.3

65、However,under the heightened exposure scenario(the complementarity impact of GAI is low but the exposure impact is high),a larger share of workers will transition out of augmented occupations.Under the gradual integration scenario(low impact of both exposure and complementarity),the bar is almost id

66、entical to status quo for augmented occupations,but is slightly lower for disrupted occupations and higher for insulated occupations.Under the broad augmentation scenario(the impact of GAI is high from complementarity of GAI and low from exposure),the projected fraction of people in 3 Note this is s

67、omewhat different(and in this case,smaller)than the proportions shown in Figure 1.That is because Figure 1 relies on a balanced sample,and so is over-represented by slightly older,more senior people(given they have had to have been members of LinkedIn since 2018).augmented occupations is predicted t

68、o increase relative to status quo.We can also easily contrast across GAI groups.Consider the case again of the broad augmentation scenario.In that case,we predict fewer workers will be employed in disrupted occupations than under status quo,while more will work in augmented and insulated occupations

69、.On the other hand,under the heightened exposure scenario,we predict even fewer people working in disrupted occupations.However,those workers would not primarily transition into augmented occupations this time Figure 2 Predicted shares in each GAI group by scenario in one year 8 GAIs Influence on Em

70、ployment Patterns Economic Graphas in the broad augmentation scenario,but instead to disrupted occupations.It is important to note,the magnitudes we assume under the different scenarios have a degree of subjectivity(what constitutes a large impact versus a small)and are not entirely data-driven,of n

71、ecessity.Instead,they are illustrative of different potential scenarios for use in comparison.For that reason,we would not for example mean to imply we are claiming that if there is low complementarity from GAI but high exposure,the fraction of workers in insulated occupations would increase to 47.3

72、%instead of 43.5%under no change(as shown in Figure 2).Instead,we encourage examination of the overall trends(not levels achieved),and even more so,comparison across educational groups under the different scenarios as done below.Proportion of members exiting employment We next investigate the share

73、of workers who are predicted to move from working to not working.This could occur for many reasons,including unemployment or separation from the labor force,either temporarily(voluntarily employment gaps,including for parental leave)or permanently(e.g.,retirement).We predict that the share of people

74、 exiting employment stays approximately the same under gradual integration,decreases slightly under broad augmentation,increases slightly under the paradigm shift,and increases substantially under heightened exposure(Figure 3).These trends show the importance of which scenario plays out,and the risk

75、s inherent in situations where skills are replicable by GAI as well as the strong insulating effect skills augmentation plays.Proportion of members changing occupations Even if people do not exit work,they may switch occupations.This may occur for many reasons,including voluntary and planned career

76、changes as well as unplanned or layoff-pressured Figure 3 Predicted proportion exiting employment Figure 4 Predicted proportion changing occupations 9 GAIs Influence on Employment Patterns Economic Graphchanges.Figure 4 presents the prediction.Under status quo,we predict 5.7%of the workers would cha

77、nge occupations(given by the typical occupation change in the past year).This is also what would occur under gradual integration.However,under any of the scenarios with a large impact on either skills augmentation or skills replicability,there would be an increase in the share of workers changing oc

78、cupationsthe most in the case of heightened exposure or paradigm shift,at 6.4%.Summary of findings for overall predictions Table 1 helps summarize all of these findings in a more qualitative manner.Predicted changes by education level We turn our attention to the share of workers in each GAI group b

79、y education level,as we did in our example above.Shares over time We first show the historical trends in each educational group,as we did in Figure 1 above for the overall population.While we found in Figure 1 that overall,workers were shifting towards augmented occupations years before GAI,these sa

80、me trends are not repeated for each education group.In fact,we find that lower education workers(high school and associate)have been trending away from augmented occupations since years before,while higher education workers(bachelors and graduate degree holders)have been trending towards augmented o

81、ccupations.This means the gap was already increasing in the occupations poised most to benefit from GAI.On the other hand,the trend shown in Figure 1 of workers moving away from disrupted occupations for years prior to the introduction of GAI is revealed on average for each of the education groups.H

82、owever,bachelors degree holders had the fastest rate of decline in their participation in GAI-disrupted occupations.Finally,bachelors degree holders were least likely to be in insulated occupations,a gap that has widened over time with a shallower growth trajectory than other education groups.Table

83、1 Summary of Predicted Impacts of Scenarios Impact of GAI-replicable skills Low High GAI-complementary skills Low Gradual Gradual integrationintegration Augmented Disrupted Insulated Exit work Change occ Heightened Heightened exposureexposure Augmented Disrupted Insulated Exit work Change occ High B

84、road Broad augmentationaugmentation Augmented Disrupted Insulated Exit work Change occ ParadigmParadigm shiftshift Augmented Disrupted Insulated Exit work Change occ and :changes exceeding 2%from status quo and :changes between 0.1%and 2%from status quo :changes between 0 and 0.1%from status quo 10

85、GAIs Influence on Employment Patterns Economic GraphCurrent levels We first examine the current levels of workers in each of the groups,as shown in Figure 5.We find that overall,the largest share of workers are in insulated occupations(around 2/5ths of workers),followed by workers in disrupted occup

86、ations(around 1/3rd),and the smallest share are in augmented occupations(around 1/4).However,there are important differences across education groups.For example,graduate and associate degree workers are more likely to be in occupations likely insulated from GAI than high school or bachelors degree h

87、olders(such as project manager or teacher for graduate degree workers,and medical and mechanical technician for associate degree workers).On the other hand,high school and bachelors degree holders are the most likely to be in disrupted occupations.Interestingly,bachelors degree holders are both in t

88、he groups most likely to be disrupted and most likely to be augmented,showing how often they are in jobs with complementary skills to GAI.One of the common comparisons we will make in this white paper is between high school graduates and bachelors degree holders.This is shown in the table as well.Bo

89、th groups have around 1/3 of their members working in disrupted occupations(such as administrative assistants and salespersons).However,bachelors degree holders have more workers in augmented occupations(such as software engineers or marketing managers)than high school graduates(27.8%compared to 25.

90、6%),while high school graduates have more workers in insulated occupations(40.7%versus 38.8%).These gaps may seem small,but are in reality non-negligible differences representing an underlying population of millions of workers.Figure 5 Share of workers in each GAI group over time,by education 11 GAI

91、s Influence on Employment Patterns Economic GraphGAI predicted group composition As described above,we next simulate the GAI-occupational group in which members work in the future under the five difference scenarios.We first examine this for participation in occupations belonging to the augmented gr

92、oup.Figure 5 presents the average participation rates one year from now for the five scenarios(including status quo).Figure 6 presents the same data but represented as percent difference between the given education group and bachelors degree holders.This is done for ease of interpretation,to be able

93、 to with facility ascertain how each group fares relative to bachelors degree holders,and how this differs across scenarios.It also emphasizes the goal of the simulations,which is to contrast how the gaps change across scenarios.The values in Figure 5 are completely dependent on our choice of parame

94、ters in the simulations.In the heightened exposure scenario shown in red,all groups have lower participation in augmented occupations relative to status quo.However,Figure 6 demonstrates that this scenario would widen the gap between high school graduates and bachelors degree holders,from 7.7%lower

95、participation for high school graduates down to 9.8%less.On the other hand,if the complementary impact of GAI is high(i.e.,either in broad augmentation or paradigm shift scenarios),then the share of people working in augmented occupations would be predicted to increase over the status quo for all ed

96、ucation groups.However,the increase is proportionally smallest for bachelors degree graduates,as seen by the narrowing of the gap relative to lower education levels.Thus,for measuring the share of workers in augmented occupations in the future,the gap between bachelors degree holders and high school

97、 workers would be largest under the heightened exposure scenario and smallest under the broad augmentation case.Interestingly,the complementarity of skills seems to dominate the replicability of skills for this Figure 6 Share of members in each GAI group by education 12 GAIs Influence on Employment

98、Patterns Economic Graph Figure 7 Share of workers in augmented occupations one year from now,by education Figure 8 Gap in the share of workers in augmented occupations one year from now,relative to bachelors degree holders employment in augmented occupations 13 GAIs Influence on Employment Patterns

99、Economic Graphoutcome between these groups,as the gap would still be narrower in the paradigm shift scenario than under status quo.For the other two GAI occupation classifications(disrupted and insulated),we present the levels(such as is done in Figure 7)for discussion while presenting the gaps(such

100、 as is done in Figure 8)in the appendix.We next examine the share of workers in occupations we predict to be disrupted by GAI,shown in Figures 9 and A.3.Under the gradual integration scenario,we predict very small decreases in the share of workers working in disrupted occupations,as we might expect.

101、This does not have large impacts on the observed gaps;the largest reduced gap is for high school graduates who are 7.8%more likely to work in disrupted occupations than bachelors degree graduates under status quo,but a slightly smaller 6.9%more likely under gradual integration.On the other hand,unde

102、r all three of the other scenarios,each education group sees a bigger decrease in participation in disrupted occupations.This is especially true for bachelors degree holders,which narrows some of the gaps.If the impact replicability of skills is high(heightened exposure scenario or paradigm shift sc

103、enario),then there is a much larger decrease in the share of workers in disrupted occupations across the board.While this doesnt alter the gap between bachelors graduates and either associate degree graduates(about as likely)or graduate degree graduates(more likely),it has a large impact on the gap

104、between high school graduates compared to bachelors degree Figure 9 Predicted share of workers in disrupted occupations one year from now,by education 14 GAIs Influence on Employment Patterns Economic Graphgraduates.They are 7.8%more likely to be in disrupted occupations under status quo,but 0.3%les

105、s likely than bachelors degree graduates under the paradigm shift scenario.The impact under the broad augmentation scenario is not as strong as either of the other two scenarios,although it also leads to less participation in disrupted occupations,especially for bachelors degree holders,narrowing th

106、e gap compared to status quo.We next look at how participation in insulated occupations changes under different scenarios,shown in Figures 10 and A.4.Under all scenarios and for each education group,participation in insulated occupations is predicted to increase relative to status quo,pre-GAI.Howeve

107、r,the extent to which it changes varies greatly.There are once again only minor increases in participation under the gradual integration scenario,and only minor narrowing of gaps closer to zero.The largest changes occur under scenarios when the impact from replicability of skills by GAI is high,name

108、ly the heightened exposure scenario and the paradigm shift scenario.In these cases,workers are more likely to work in insulated occupations.Under the heightened exposure scenario,participation in insulated occupations is higher than any other scenario.This leads to high school graduates being only 1

109、.0%less likely than bachelors degree holders to be 5.1%more likely.On the other hand,it decreases the higher levels that associate and graduate degree holders work in insulated occupations relative to bachelors degree holders.The same is true also for the other case of high impact from skill Figure

110、10 Predicted share of workers in insulated occupations one year from now,by education 15 GAIs Influence on Employment Patterns Economic Graphreplicability,the paradigm shift scenario,although not to quite as large of an extent.For the broad augmentation scenario,when the replicability impact of GAI

111、is low,there is only a small increase in the share of workers in insulated occupations across education groups.This leads to virtually unchanged gaps between education groups.In summary,large changes would happen under the heightened exposure scenario,as fewer individuals would work in augmented and

112、 disrupted occupations while more would work in insulated occupations,and the education gap between high school graduates and bachelors graduates would widen for augmented occupations(in favor of college graduates),widen for insulated occupations(in favor of high school graduates),and narrow to zero

113、 for disrupted occupations(away from a status quo advantage in favor of high school graduates).On the other hand,under the broadened augmentation scenario,participation in augmented occupations would increase across the board of education groups,taking away from participation in disrupted occupation

114、s.Participation in insulated occupations would increase a little across the board.Focusing on the education gap between high school graduates and bachelors degree graduates again,this scenario would narrow the gap for augmented occupation work(from an advantage for college graduates),have little imp

115、act on the already-small gap for insulated occupation work,and narrow the gap for disrupted occupations(from an advantage for high school graduates).Under the paradigm shift scenario,many of the above-discussed impacts of the heightened exposure scenario and the broadened augmentation scenario would

116、 slightly moderate each other while still having sizeable over impacts relative at least to status quo.The primary exception here is for participation in disrupted occupations,wherein the two impacts somewhat amplify each other and lead to even higher departures from working in disrupted occupations

117、 and,at least for the high school-bachelors degree gap,have the largest change.Proportion of individuals exiting employment While workers will switch between occupational groups,some will also change whether they switch to not working at all one year later given shifts.We next simulate and estimate

118、this outcome by scenario and education group.For example,under status quo we estimate that 2.6%of high school graduates will not be working one year from now as shown by LinkedIn profile status.This does not mean that 2.6%will lose their job,as it will include some already not working and some who e

119、xit employment,as well as be reduced by some who move from not working to working.We find similar although somewhat smaller rates under status quo.Under the gradual integration scenario,the proportion not working is approximately the same as in status quo.However,for the heightened exposure scenario

120、,there is a substantial increase in the share of individuals not working across all education groups.This impact is proportionately largest for bachelors degree holders,which leads to the high school graduates having a 34.6%higher rate of not working under status quo decrease to 12.8%under heightene

121、d exposure.16 GAIs Influence on Employment Patterns Economic GraphOn the other hand,the broadened augmentation scenario would decrease the share of people not working relative to status quo for all groups.This too would widen narrow the gap between high school graduates and bachelors degree holders,

122、although by not as much as under the heightened exposure scenario,and for a different reason(decrease in both groups but stronger for high school graduates,instead of increase in both groups but stronger for bachelors degree holders).Finally,under the paradigm shift scenario,the two effects countera

123、ct each other(augmented skills decreasing the proportion not working,while replication of skills increases the proportion).The replication of skills effect dominates,with the overall shares not working increasing slightly over status quo for each education group.The high school-bachelors degree gap

124、is narrowed under this scenario as well,relative to status quo.Proportion of individuals changing occupations We next calculate the proportion of workers who change occupations from one year to the next under the different scenarios,as shown in Figure 12.We find that occupational changing increases

125、under every scenario compared to status quo.However,the largest shifts are in heightened exposure and paradigm shift,both of which reflect a higher degree of impact from skill replicability of GAI.This is true across education levels,with no strong differential.Conclusion This research paper aimed t

126、o investigate the potential influences of GAI on employment patterns in the US under different scenarios.The primary focus is on understanding the differential impacts of GAI on workers and examining Figure 11 Share of members exiting employment,by education 17 GAIs Influence on Employment Patterns

127、Economic Graphpotential shifts in occupational trends over time,with additional attention to how different educational groups may be impacted.The appendix discusses assumptions and limitations of the paper.Over the past six years,a noticeable,steady shift has occurred among workers predating the rel

128、ease of GAI in 2022,with workers moving away from occupations which would later be susceptible to disruption by GAI and towards those that are insulated or augmented.Notably,bachelors degree holders exhibited the fastest decline in occupations which would potentially be disrupted by GAI.4 Educationa

129、l disparities in occupational trends are evident,with lower education workers(high school and associate)trending away from augmented occupations,while bachelors and 4 From our methodology,we are unable to determine the extent to which this is driven by their choice and labor supply,compared to trend

130、s in demand for different occupations.graduate degree holders are trending towards them.This highlights a growing gap in occupations poised to benefit from GAI.However,if GAI has a high impact complementing skills,we predict that the educational gap in augmented occupations would be narrower than if

131、 the trends continued at their historical and diverging trajectories.Predictions reveal many other intriguing patterns.Under the heightened exposure scenario,while participation is predicted to decrease for all groups,the advantage that bachelors degree holders would have over high school graduates

132、in participation would increase.Similarly,the gap in the share working in disrupted occupations(with high school graduates being more likely to work there)would decrease in this scenario.Figure 12 Share changing occupations 18 GAIs Influence on Employment Patterns Economic GraphScenarios with high c

133、omplementary impact of skills tend to favor high school graduates more than bachelors degree holders compared to the status quo.For example,decreasing the augmented occupation advantage over status quo virtually erases the higher rate under status quo of high school graduates working in disrupted oc

134、cupations.halves the higher rate at which high school graduates leave employment each year.Table 2 summarizes the findings descriptively.In most scenarios and for many outcomes,the predicted impacts follow in the same direction across educational groups,although the extent to which each education gr

135、oup differs.Table 2 Predicted outcomes across educational groups All High school Associate degree Bachelors degree Graduate degree Gradual integrationGradual integration Share in augmented occupations Share in disrupted occupations Share in insulated occupations Share exiting working Share changing

136、occupations Heightened exposureHeightened exposure Share in augmented occupations Share in disrupted occupations Share in insulated occupations Share exiting working Share changing occupations Broad augmentationBroad augmentation Share in augmented occupations Share in disrupted occupations Share in

137、 insulated occupations Share exiting working Share changing occupations Paradigm shiftParadigm shift Share in augmented occupations Share in disrupted occupations Share in insulated occupations Share exiting working Share changing occupations and :changes exceeding 2%from status quo;and :changes bet

138、ween 0.1%and 2%from status quo;:changes between 0 and 0.1%from status quo Economic GraphReferencesReferences Acemoglu,D.,&Autor,D.(2011).Skills,tasks and technologies:Implications for employment and earnings.In Handbook of labor economics(Vol.4,pp.10431171).Elsevier.https:/ Acemoglu,D.,&Restrepo,P.(

139、2018).The race between man and machine:Implications of technology for growth,factor shares,and employment.American Economic Review,108(6),14881542.Card,D.,&DiNardo,J.(2006).The impact of technological change on low-wage workers:A review.Working and Poor:How Economic and Policy Changes Are Affecting

140、Low-Wage Workers,113140.Card,D.,&DiNardo,J.E.(2002).Skill-Biased Technological Change and Rising Wage Inequality:Some Problems and Puzzles.Journal of Labor Economics,20(4),733783.https:/doi.org/10.1086/342055 Eloundou,T.,Manning,S.,Mishkin,P.,&Rock,D.(2023).GPTs are GPTs:An Early Look at the Labor M

141、arket Impact Potential of Large Language Models(arXiv:2303.10130).arXiv.http:/arxiv.org/abs/2303.10130 Felten,E.W.,Raj,M.,&Seamans,R.(2023).Occupational heterogeneity in exposure to generative ai.Available at SSRN 4414065.https:/ Kimbrough,K.,&Carpanelli,M.(2023).Preparing the Workforce for Generati

142、ve AI:Insights and Implications(LinkedIn Economic Graph Research Note).https:/ Microsoft,&LinkedIn.(2024).2024 Work Trend Index Annual Report.https:/ Economic GraphAppendixAppendix AcknowledgementsAcknowledgements We gratefully acknowledge the feedback and support of many people in this research,inc

143、luding Gorki De Los Santos,Chris Grant,Karin Kimbrough,Carl Shan,Anne Trapasso,and Amie Wong.MethodMethodologyology Data and Privacy This body of work represents the world seen through LinkedIn data,drawn from the anonymized and aggregated profile information of LinkedIns one billion members around

144、the world.As such,it is influenced by how members choose to use the platform,which can vary based on professional,social,and regional culture,as well as overall site availability and accessibility.In publishing these insights from LinkedIns Economic Graph,we want to provide accurate statistics while

145、 ensuring our members privacy.As a result,all data show aggregated information for the corresponding period following strict data quality thresholds that prevent disclosing any information about specific individuals.Generative AI classifications(replicated from Kimbrough&Carpanelli 2023)GAIGAI-repli

146、cable and GAIreplicable and GAI-complementary skillscomplementary skills We identify GAI-replicable and GAI-complementary skills with the following steps:1.We ask ChatGPT 3.5(Feb 2023)the following prompts:a.GAI-replicable skills:What are the 100 top skills that AI technologies(ChatGPT,Dall-E,LaMDA,

147、etc.)can perform very well?b.GAI-complementary skills:What are the 100 top skills that can currently exclusively be performed by humans?We map these lists to LinkedIns taxonomy with LinkedIns taxonomy API,and we refine matches Economic Graph2.We expand coverage further by applying skill similarities

148、 based on skill embeddings to score kills that resemble those flagged in each list,and by manually reviewing the skills in the popular skill groups containing the skills from the previous steps.3.For external validation,we ingest and map to our taxonomy three exposure scores from the academic litera

149、ture(Webb(2019);Felten,Raj,&Seamans(2023),and Felten,Raj,&Seamans(2021).We use these scores to train a model that learns which skills contribute more to these three rankings,and we use this model to score all skills in LinkedIns taxonomy.SkillsSkills-based occupation classificationbased occupation c

150、lassification For each occupation,we calculate the percentage of skills that are GAI-replicable and GAI-complementary based on each occupations skills genome.An occupations skills genome is the ranking of its top 100 most relevant skills,based on a TF-IDF model.In this model,skills are relevant when

151、 they tend to be disproportionately added by members in this occupation compared to other occupations.We classify each occupation into Augmented,Disrupted,or Insulated from GAI,based on their GAI-replicable and GAI-complementary medians.Occupations with above median GAI-replicable skills and above-m

152、edian GAI-complementary skills are classified as Augmented,occupations with above median GAI-replicable skills and below-median GAI-complementary skills are classified as Disrupted,and occupations with below median GAI-replicable skills are classified as Insulated.Simulations of outcomes In order to

153、 simulate outcomes,we first estimated a transition matrix from each occupation or not working to each occupation or not working one year later.We do so using data from hundreds of millions of US members profiles and job histories.For example,for the transition from occupation to occupation,we take t

154、he sample of all people in occupation in year and then calculate what fraction of them are working in occupation one year later.This estimated transition matrix forms the basis for status quo.We test both using only the most recent year(i.e.,for each month in 2022,examining the occupation of the ind

155、ividual one year later in 2023 of the same month),as well as across the past six years in sensitivity checks.We then calculate a counterfactual transition matrix for each of the four scenarios.Let!be the estimated transition probability under status quo(what we observe in the data)for Economic Graph

156、transitioning from occupation to occupation over the course of one year,and be the overall transition matrix.Intuitively,we estimate the shift in transitions based on changes in demand for occupation.We do so using the formula#=1+$#%#Here,:occupation js%of skills complemented by GAI :occupation js%o

157、f skills exposed to GAI$#:the extent to which skills complemented by GAI increase demand for the occupation%#:the extent to which skills exposed by GAI decrease demand for the occupation Thus,having many skills that are complemented by GAI increases demand for the occupation;having many skills expos

158、ed to GAI decreases demand for the occupation.We use this to calculate the transition matrix elements under the counterfactual scenario by&#/=#(#!(*!The scaling(#!(*is necessary to reflect the need for the rows of the transition matrix to sum to one.For example,if there is an occupation with no expo

159、sed or complemented skills,#=1 and there is no shift in demand for workers for that occupation.However,if people from occupation tend to transition into highly complemented occupations,for which demand went up,then the demand to the first occupation would decrease because of the relative shifts.The

160、same would conversely hold for if transitions tended to happen to highly exposed occupations,which would lead to an increase in transitions into the first occupation.Our approach does not account for any general equilibrium impacts that may occur as workers move between occupations.We test four hypo

161、thetical scenarios:$#%#Gradual integration 0.1 0.1 Heightened exposure 0.1 0.9 Broad augmentation 0.9 0.1 Paradigm shift 0.9 0.9 Using these scenarios,we can simulate out shares of employees in each occupation and not working one year from now,by using the initial shares in each occupation(and not w

162、orking)and the counterfactual transition matrix+.From that,we can estimate several Economic Graphoutcomes.The ones we examine are the fraction of workers in each GAI classification,the fraction transitioning to not working,and the fraction who change occupations over the course of the year.The taxon

163、omy we use for occupation includes just over 600 unique occupations.Assumptions and limitations There are a number of assumptions made in the approach of this paper which may impact the findings.No general equilibrium effectsNo general equilibrium effects We assume that shifts in shares of workers i

164、n occupations do not shift demand for those workers due to a larger supply,for example.Instead,Types of occupations in each GAI group are fixedTypes of occupations in each GAI group are fixed We assume that over time,occupations do not shift how much the skills used in the occupation are complemente

165、d or replicated by GAI,and also hold the skills used in each occupation constant.This is almost certainly incorrect in the long-term,and thus our projections should only be viewed with a short-term time horizon.Introduction of new occupationsIntroduction of new occupations If GAI itself creates new

166、occupations,or if technology in other areas creates new occupations,we will miss the shift into and out of these.This is another reason why we only look at one-year projections.Data estimation of Markov transition matrixData estimation of Markov transition matrix The Markov transition matrix is base

167、d off of one year(Dec 2021 through Dec 2022,and one year later for each month).We also tested it with six years prior instead of only one.The goal was to estimate transitions prior to introduction of GAI to serve as status quo.No No heterogeneity in impact of GAI across industries,occupations,or tim

168、eheterogeneity in impact of GAI across industries,occupations,or time We have modeled the impact as a function only of the current top skills in an occupation and how those are complemented or replicated by GAI.However,some occupations may have more reaction to GAI than other occupations that use so

169、me of the same skills,depending on such factors as rate of adoption of GAI in the occupation and industry or digital literacy within the occupation.We also do not model shifts over time,which is why we focus only on a one-year Economic GraphFigure A.1 Share of members in each GAI panel,balanced pane

170、l Figure A.2 Trend with balanced panel Economic GraphFigure A.3 Gaps in outcomes for disrupted occupations,given education group versus bachelors degree holders Figure A.4 Gaps in outcomes for insulated occupations,given education group versus bachelors degree holders Economic GraphFigure A.5 Gaps in outcomes for exiting employment,given education group versus bachelors degree holders Figure A.6 Gaps in outcomes for changing occupations,given education group versus bachelors degree holders