《國際清算銀行：2024生成式AI的崛起對美國勞動力市場的影響分析報告：滲透度、替代效應及對不平等狀況的影響（英文版）（40頁）.pdf》由會員分享，可在線閱讀，更多相關《國際清算銀行：2024生成式AI的崛起對美國勞動力市場的影響分析報告：滲透度、替代效應及對不平等狀況的影響（英文版）（40頁）.pdf（40頁珍藏版）》請在三個皮匠報告上搜索。

1、BIS Working Papers No 1207 The rise of generative AI:modelling exposure,substitution,and inequality effects on the US labour market by Raphael Auer,David Kpfer,Josef vda Monetary and Economic Department September 2024 JEL classification:E24,E51,G21,G28,J23,J24,M48,O30,O33 Keywords:Labour market,Arti

2、ficial intelligence,Employment,Inequality,Automation,ChatGPT,GPT,LLM,Wage,Technology BIS Working Papers are written by members of the Monetary and Economic Department of the Bank for International Settlements,and from time to time by other economists,and are published by the Bank.The papers are on s

3、ubjects of topical interest and are technical in character.The views expressed in them are those of their authors and not necessarily the views of the BIS.This publication is available on the BIS website(www.bis.org).Bank for International Settlements 2024.All rights reserved.Brief excerpts may be r

4、eproduced or translated provided the source is stated.ISSN 1020-0959(print)ISSN 1682-7678(online)The rise of generative AI:modelling exposure,substitution,andinequality effects on the US labour marketRaphael AuerDavid K opferJosefSv edaAugust 21,2024AbstractHow exposed is the labour market to ever-a

5、dvancing AI capabilities,to what extentdoes this substitute human labour,and how will it affect inequality?We address thesequestions in a simulation of 711 US occupations classified by the importance and level ofcognitive skills.We base our simulations on the notion that AI can only perform skills t

6、hatare within its capabilities and involve computer interaction.At low AI capabilities,7%ofskills are exposed to AI uniformly across the wage spectrum.At moderate and high AIcapabilities,17%and 36%of skills are exposed on average,and up to 45%in the highestwage quartile.Examining complementary versu

7、s substitution,we model the impact onside versus core occupational skills.For example,AI capable of bookkeeping helps doctorswith administrative work,freeing up time for medical examinations,but risks the jobs ofbookkeepers.We find that low AI capabilities complement all workers,as side skills aresi

8、mpler than core skills.However,as AI capabilities advance,core skills in lower-wagejobs become exposed,threatening substitution and increased inequality.In contrast to theintuitive notion that the rise of AI may harm white-collar workers,we find that those remainsafe longer as their core skills are

9、hard to automate.JEL codes:E24,E51,G21,G28,J23,J24,M48,O30,O33Keywords:Labour market,Artificial intelligence,Employment,Inequality,Automation,ChatGPT,GPT,LLM,Wage,TechnologyWe thank Ryan Banerjee,Sebastian Doerr,Fiorella De Fiore,Fernando Perez-Cruz,Andras Valko,andseminar participants at the BIS fo

10、r comments and suggestions.We acknowledge the use of GPT4 for editing andas mentioned in the paper.The presented views are not necessarily those of the BIS.Bank for International Settlements and CEPR,raphael.auerbis.orgBank for International Settlements,david.koepferbis.orgBank for International Set

11、tlements,josef.svedabis.org11IntroductionHow will the advancement of generative AI complement and substitute different kinds of humanlabour?Recent breakthroughs have enabled generative AI to mimic human cognitive abilities inmany fields,including in“white collar”professions such as law,medicine,or s

12、cience.Ongoingadvances and integration of the technology into day-to-day applications and workflows raiseurgent policy questions.Understanding how the potential evolution of AI will complement or substitute human skillsis essential for shaping policies to ensure equitable growth and employment stabi

13、lity.The liter-ature has focused on the occupation-level impact of current AI models,1experimental evidenceof productivity impacts(Noy&Zhang,2023;Brynjolfsson et al.,2023;Peng et al.,2023),andthe potential for complementarity and substitution effects of AI technology at a particular stateof AI devel

14、opment(Pizzinelli et al.,2023;Acemoglu&Restrepo,2019,2018c,a).Except forcertain types of freelancers(see e.g.Webb 2020),the broader impact of AI capabilities on thelabour market yet remains to be demonstrated.In this paper,we take a forward-looking approach:we ask the question of“what if”andexamine

15、how an AI of a hypothetical level of capabilities was to expose2different occupations.To shed the first light on the future impact,we build a parsimonious bottom-up quantificationwith a special focus on income distribution.Our analysis proceeds in two steps.In the first step,we build on Eloundou et

16、al.(2023);Felten et al.(2021);Gmyrek et al.(2023);Pizzinelli et al.(2023);Acemoglu(2024)and modelthe exposure to the technology as the capabilities of AI increase.3In the second step,weexamine how these developments could complement or substitute human labour through thelens of their impact on core

17、and side skills.In the first step,we argue that the near-term impact of AI is limited a)to computer-relatedinteractions and b)by the difficulty of the skills that AI can substitute for.In this,we onlyquantify the impact on skills involving interaction with a computer.We,hence,do not takeinto account

18、 the impact of AI on robotics that may substitute for physical work or even socialinteractions.4Our first departure from the literature is to employ an underused part of the O*NETdatabase that classifies skills by their difficulty.Intuitively,an AI of a certain capability levelcan only perform tasks

19、 up to a corresponding skill level.As the capabilities of AI advance,an1See i.e.Webb(2020);Felten et al.(2021);Tolan et al.(2021);Gmyrek et al.(2023);Yang(2022)2Throughout the paper,we use the terms“expose”and“exposure”in a neutral manner,to imply that someparts of a skill,task,or occupation could b

20、e enhanced,performed,or otherwise be affected by an AI.3Similar to these approaches,we take a partial equilibrium perspective and do not take into account theinterplay between skills,relative wages,human capital formation and directed technological change(Acemoglu&Restrepo,2018c).4This is in line wi

21、th Acemoglu(2024),who argues that“AI is nowhere close to being able to perform mostmanual or social tasks”,and we thus assume that it can only perform computer interactions.2increasing share of cognitive skills will hence be exposed to the technology.5We nest this notion of AI capability and skill d

22、ifficulty in a quantitative simulation of 711US occupations from the O*NET database classified by the importance and the required levelof cognitive skills that involve computer interactions.The model predicts that an AI capableof substituting for simple cognitive tasks such as the minimal communicat

23、ion skills requiredfor a truck driver will expose around 7%of all skills.At low levels of AI capability,this effectholds uniformly across the entire wage spectrum,but for heterogeneous reasons.For low-incomeworkers,a substantial share of cognitive computer skills is exposed,but the overall share of

24、timespent on computer interaction is low.For high-income workers,only a small share of cognitivecomputer skills is exposed because of the larger skill requirement.However,the share of timespent using such skills is higher.6As AI capabilities increase,we observe a profound difference in occupational

25、exposure:upto 45%in the upper quartile of the wage distribution are exposed,whereas the exposure of thelowest quartile is around 26%.What does this mean for the income distribution?We note that in line with the literature,“exposure”has a neutral meaning in that some parts of a skill,task,or job coul

26、d be performedby an AI.This may lead to substitution but could also complement via increased productivity.7To shed light on these issues,in the second departure from the literature and step of oursimulations,we examine the extent to which AI might complement or substitute human labour.We focus on th

27、e differential impact on core versus side occupational skills,arguing that AI wouldtend to complement occupations wherever the auxiliary(side)skills necessary for the professionare within its capabilities.For example,if AI can organise meetings,billing,or bookkeeping forlawyers,medical doctors,or sc

28、ientists,this frees up time that can be spent on core activitiesand thus increases productivity.However,a profession may be at risk if the core activity itselfcan be performed by the AI.This exercise suggests that AI may initially complement all professions,as side skills are5We take no position on

29、how fast the evolution of the technology will materialise.Some have argued that AImay soon have dramatic impacts on the labour market(ie Korinek&Juelfs(2022).Others argue that futureadvancement of AI may materialise much slower than expected.For example,Acemoglu(2024)argues thatearly evidence is fro

30、m easy-to-learn tasks with clear outcomes(that AI can optimise for),whereas more profoundproductivity impacts in more subtle contexts may materialise much slower.Perez-Cruz&Shin(2024)argue thatcurrent LLMs are limited in their understanding of human interaction and higher-order beliefs.6For these ex

31、amples,“simple cognitive tasks”correspond to those requiring a skill level of 2.0 in the O*NETdatabase,for example,the minimum social perceptiveness skills required for pile drivers or the minimum speakingskills required for industrial truck operators.“Medium cognitive tasks”correspond to those requ

32、iring a skill levelof 3.0,for example,problem-solving skills of medical appliance technicians or the operations monitoring skillsof registered nurses.“High cognitive tasks”correspond to those requiring a skill level of 4.0,for example,thepersuasion skills of psychiatrists or the active listening ski

33、lls of air traffic controllers.7Svanberg et al.(2024)further note that“exposure”does not mean automation:they survey workers with“end-use”tasks to get a sense of the requirements for automation,and second,they model the cost of a modelcapable of meeting the requirements.Focusing on the automatabilit

34、y of vision,find that only 23%of occupationsthat are“exposed”in the sense of Eloundou et al.(2023);Felten et al.(2021)could today be automated eco-nomically.We note that our measure of exposure is more nuanced than the one in Eloundou et al.(2023);Feltenet al.(2021)as we restrict the impact to skill

35、s involving computer interaction and not only model whether a skillin principle could be automated but also whether the capability level of the AI is sufficient for such automation.3generally less difficult than core skills.For example,an AI only capable of performing simplecognitive tasks has negli

36、gible exposure to core skills,whereas it,on average,exposes around12%of side skills.However,already for moderate AI capabilities,there is divergence acrossthe wage spectrum,with the core cognitive skills of the low-wage workers becoming roughly asexposed to AI as their side skills.In contrast,the up

37、per quartile of the wage distribution stillsees negligible exposure of core skills(5%),whereas side skills are exposed substantially(27%).If AI capabilities are high,around 25%of both side and core skills of the lowest quartile ofthe wage distribution are exposed.In contrast,only 20%of the core but

38、a staggering 62%ofthe side skills of the highest quartile of the income distribution become exposed.On balance,our modelling of the impact on side and core skills hence reverses the notion thatgenerative AI might decrease inequality in the labour market(Noy&Zhang,2023;Brynjolfssonet al.,2023).Despit

39、e being a technology that is exposing white-collar jobs more intensively,thiseffect is focused on the side skills of their professions,while the core skills are not in reach.8Incontrast,a capable AI will also expose the core skills of lower-income workers,thus threateningsubstitution and widening in

40、equality.The balance of this paper is as follows:we relate our approach to the literature in SectionSection 2.Next,Section 3 presents the methodology describing the evolutionary impact of ever-improving AI on occupations.It also serves as an AI exposure dependent on AIs capabilities.Thereafter,we sp

41、lit the AI exposure based on core and side skills Section 4 that are then usedto identify complementarity and substitutional effects for individual occupations.Section 5presents additional robustness analysis,while Section 6 concludes.2Literature reviewHistorically,technological advancements have be

42、en met with both optimism and concern re-garding their implications for the labour market(Bessen,2016).The advent of AI and machinelearning technologies,in general,has intensified these debates,with researchers seeking to un-derstand how these new tools can reshape the labour market and how the impa

43、ct can differfrom previous technological advancements in robotisation or computerisation(Autor,2015).Several recent studies have directly addressed the potential of the latest advancements inAI to significantly impact the current structure of the labour market.Brynjolfsson et al.(2018)argue that mos

44、t occupations in the US include at least some tasks that are suitable for machinelearning applications,and Eloundou et al.(2023)suggests that 80%of the workforce could beaffected by Generative Predictive Transformers(GPTs).While these estimates are staggering,Arntz et al.(2016)argue that the actual

45、vulnerability of jobs to automation is lower whenconsidering the nuanced skills within occupations.Nonetheless,the proliferation of the latestLLMs seems to be non-negligent;Eloundou et al.(2023)further find 19%of US workers in the8Of course,once the capability of the AI becomes extremely high such t

46、hat all skills are within reach,thiseffect abates,and all cognitive workers are in danger of replacement.4US may see at least half of their skills impacted and Hatzius et al.(2023)finds 25%of currentwork skills in US automatable.Current AI capabilities,in some instances,fall short of profound reason

47、ing skills(Perez-Cruz&Shin,2024).However,an important issue regards how the future evolution of AI capabilitiescan enhance labour productivity or crowd out workers.Recent experiments with the latestgeneration of AI show that it can have a positive effect in specific occupations while reducingdiffere

48、nces among workers with varying experience levels.Noy&Zhang(2023)demonstratedthat the use of ChatGPT significantly increases average productivity measured by time spenton tasks and reduces differences between high-and low-skilled workers.Brynjolfsson et al.(2023)studied the introduction of genAI ass

49、istant to the customer support agents and found asignificantly higher number of completed tasks that were more pronounced for novice and low-skilled workers.Peng et al.(2023)suggests coders with access to genAI are capable of completingcoding-oriented tasks up to 55%faster.AI tools can also serve as

50、 the tool to discover potentialimprovements in business systems(Cockburn et al.,2018;Cheng et al.,2022).However,an increase in labour productivity means that less human capital is needed tomaintain the same output,which could lead to layoffs or wage reductions(Acemoglu&Restrepo,2020).In this context

51、,Frey&Osborne(2017)predicted that up to 47%of US employment isat high risk of computerisation.Arntz et al.(2016)however uses a different methodology andestimates an impact of only 9%.Gmyrek et al.(2023)find that genAI could automate 5.1%oftotal employment in high-income countries,whereas low-income

52、countries are not so susceptible.The potential for augmentation is similarly distributed across countries relative to their incomelevels,although the potential to augment is much larger(around four to five times).Noy&Zhang(2023)claim that ChatGPT mostly substitutes for worker effort rather than pure

53、lycomplementing worker skills.Yang(2022)also shows that AI can positively affect productivityand employment but adversely affects the employment of less knowledgeable workers.Somestudies additionally debate the effects relative to gender(Eloundou et al.,2023;Webb,2020;Gmyrek et al.,2023;Aldasoro et

54、al.,2024).Historical experience with innovation shows that in the long-term,the displacement can beoffset by an increase in the range of goods and services offered,see(Autor,2015;Acemoglu&Restrepo,2019).For example,Bessen(2016)shows US labour demand has increased fasterin computerised occupations si

55、nce 1980,although the computerisation led to substitution forother occupations,shifting employment and requiring new skills.Acemoglu et al.(2022)findincreasing demand in AI-exposed occupations in the US since 2015.Automatisation in Japanand the US generated cost savings,allowing larger output in eco

56、nomy(Adachi et al.,2024;Dekle,2020;Acemoglu&Restrepo,2020)that outweighed the displacement effects of humanlabour.Yang(2022)finds that AI technology is positively associated with productivity andemployment in Taiwans electronics industry for the 20022018 period.Acemoglu&Restrepo(2019),Acemoglu&Restr

57、epo(2018a)and Acemoglu&Restrepo(2018c)then focus directly onthe dynamics of displacement and reinstatement of labour due to automation.Based on data5from the US since World War II,Acemoglu&Restrepo(2019)claim that displacement effectsoccur intuitively,but they are counterbalanced by the creation of

58、new tasks in which labour hasa comparative advantage.These then change the task content of production in favour of labourbecause of a reinstatement effect followed by a rise in the labour share and labour demand.Acemoglu&Restrepo(2019)point out that the success of reinstatement is not automatic.Itra

59、ther depends on additional variables such as the supply of new skills,demographics,or labourmarket institutions.9Although previous innovations in automatisation and computerisation,on average,broughteconomic growth,they still reshaped the labour market and introduced new challenges in re-gional labo

60、ur market structures that affected labour distribution across the skill distributionof markets.Autor(2019)documents these effects using US data showing that automation(to-gether with international trade)led to the elimination of the bulk of non-college occupations,further leading to disproportionate

61、 polarisation of urban labour markets.Acemoglu&Restrepo(2022)document that between 50%and 70%of changes in the US wage structure over the lastfour decades are accounted for by workers specialised in routine tasks in industries experiencingrapid automation.Acemoglu&Restrepo(2020)show industrial robot

62、 adoption in the UnitedStates was negatively correlated with employment and wages.These examples pinpoint theimportance of understanding the potential effects of technological advancements to navigate asmooth transition towards a new structure of the labour market.The question remains how much the n

63、ew wave of automation with AI is comparable to previ-ous technological advancements.Previously,automation exposed predominantly manual labourthrough the invention of machines and robots.The transition process to robot-driven produc-tion,therefore,affected at its first stage rather lower-skilled labo

64、ur(Acemoglu&Restrepo,2018b).Evolving AI challenges,however,cognitive tasks and skills and creates a potentialto affect different occupations by either complementing or substituting them.Earlier workby Autor&Dorn(2013)suggests that low-wage occupations faced higher substitution due tocomputerisation.

65、In contrast,high-wage occupations were complemented by technology.Webb(2020)then focuses on the newer innovation in AI and states it is directed at high-skilled tasks,effectively affecting the higher-wage quantiles.A similar conclusion is reached by Eloundou et al.(2023)and Pizzinelli et al.(2023).W

66、ebb(2020)argues that the impact of AI is different fromthe effects of software innovation,which exposed mid-wage occupations(in line with Michaelset al.(2014).Pizzinelli et al.(2023)emphasise high complementarity in the upper tail of theearnings distribution by AI,leading to a productivity boost ins

67、tead of job displacements.Theeffects of AI also differ geographically.Pizzinelli et al.(2023);Gmyrek et al.(2023);Albanesiet al.(2023)show that more developed countries are more exposed to AI as their labour marketsare more oriented to cognitive tasks.However,as AI significantly progresses,research

68、also needsto account for the evolution of technology to fully understand its potential effects.Examining9In a similar vein,Aldasoro et al.(2024)show in a general equilibrium model that the output effects of AImay primarily arise via the indirect impact on demand and associated changes in relative pr

69、ices rather than viathe direct initial productivity boost from AI adoption.6the impact of developing AI through the lens of wage distribution seems to be advantageous toformulate targeted policy responses(Furman&Seamans,2019).As the advancements in AItechnology progress,their interaction might chang

70、e rapidly.3Measuring AI exposure:data and methodologyPredicting the impact of AI on the labour market is challenging,as the integration of thetechnology into real-life applications is still in its infancy,and only some synthetic benchmarkson the potential quality and efficiency improvements on certa

71、in aspects of work are available(see i.e.Tolan et al.(2021);Peng et al.(2023);Noy&Zhang(2023).Particularly,the rapidlyevolving capabilities of AI are a major source of uncertainty.In the face of these uncertainties,we construct a parsimonious bottom-up model centred on an“AI capability”parameter,whi

72、challows us to simulate the effects of evolving AI.The model is built on the skill and occupationlevel and later aggregated to the industry or wage-quantile level.In this section,we show how we construct the AI Share Automatability(AISA)Index thatdepends on the sophistication of the AI(defined as“AI

73、 capability”above).This index rests ontwo main assumptions:1.In the short to medium term,automation will affect occupational activities with computerinteraction as opposed to social interactions or physical labour.2.The skills required for performing the occupations are heterogeneous in their diffic

74、ultylevel.For a skill to be impacted in a certain occupation,its difficulty level needs to bewithin the capabilities of the AI.We utilise data from O*NET version 27.2 and the 2022 Occupational Employment and WageStatistics(OEWS)Survey from the US Bureau of Labor Statistics.These datasets detail arou

75、nd800 different occupations(of which we can use 711 after joining across the skills tables andemployment statistics)across 22 industries,providing average income,employment numbers,and ratings for up to 35 cognitive skills for each occupation in terms of required skill level(1-6)and importance(1-5).

76、Furthermore,the data includes detailed task descriptions10for eachoccupation(on average,we have 24 task descriptions for each of the 711 occupations).In the description of our model,we will use subscripts to denote the different levels ofaggregation:the lowest level s for the skill,o for the occupat

77、ion and the highest aggregationlevels i for the industry or w for the wage quantile.The skill level Lo,sis distinct for a givenoccupation o and skill s.For instance,the occupation of Biophysicists requires a level of 4.75in the skill mathematics,while the importance of this skill Io,sis 3.88.10https

78、:/www.onetcenter.org/dictionary/21.0/text/task_statements.html(release number 21.0)73.1Only computer interaction is automatable with AIIn this paper,we only examine the impact of AI on automating tasks that require skills involvingcomputer interaction.Jobs performed on computers are,in the short and

79、 medium run,muchmore likely to incorporate AI applications compared to those involving physical labour.Weacknowledge that also physical labour may,in the future,be prone to automation throughimproved machines and robotics.However,modelling the impact of such developments is out ofthe scope of the an

80、alysis at hand.Similarly,we expect social interaction to require higher degreesof social acceptance before widespread automation materialises.Certainly,cost-effectiveness andimproved social skills of the AI will speed up the process,yet,as for physical labour,we expectlonger timescales.We construct

81、a measure of the share of the time spent on computer interactions based onabout 19,000 detailed task descriptions available in the O*NET database.Based on the de-scriptions of each occupation,we instructed GPT-4 to estimate the time spent with i)computerinteraction,ii)social interaction,and iii)phys

82、ical labour.The exact prompt is shown in theBox A1,and one example of task description is provided to the ChatGPT-4 in Table A1.Notethat computer interaction represents working on a computer that commonly does not includecommunication via e-meetings or other similar social interaction.In general,Cha

83、tGPT-4 proves very high comparability with conventional human-based pro-cedures for categorisation purposes.Eloundou et al.(2023)uses both approaches(human-andGPT4-based)to directly identify occupational AI exposure,finding a very high correlation be-tween human assessments and GPT4-based self-asses

84、sments.11Gmyrek et al.(2023)followstheir approach employing ChatGPT in exploring genAI effects on the labour market worldwide.We also cross-validate the results obtained with ChatGPT-4 by comparing the fraction oftime spent on computer interaction to the AIOE indicator by Felten et al.(2021)in the S

85、ection 5.Figure 1 shows the resulting average times spent in each of these interaction types perindustry(weighted by employment per occupation):Let Ti,odenote the share of time thatoccupation o spends in industry i on computer interaction.The average for the industry iweighted by the employment numb

86、ers Ni,ois calculated as:Ti=PoOTi,o Ni,oPoONi,oThe distribution in Figure 1 shows that the typical office professions such as“Businessand financial operations”or“Architecture and engineering”display a very large proportion incomputer interactions,“Sales”and“Personal care”a large proportion of social

87、 interaction and“Production”,“Construction”and“Transportation and farming”a large fraction of physicallabour.11Eisfeldt et al.(2023)further builds on their findings.8Figure 1:Time spent on technological,physical,and social interaction acrossindustriesNote:This figure presents the fraction of time sp

88、ent on i)computer interaction,ii)social interaction,and iii)physicallabour(see main text for the details of the data construction)in US industries.The baseline simulations of this paperassume that only computer interaction Tiis exposed to AI.3.2Occupational skills need to be within the AIs capabilit

89、ies to be automat-ableWe next measure the impact of the AIs logical capability on the exposure of various skills acrossoccupations.12The O*NET data rates the skill level and importance of around 33 cognitive skills13suchas“reading comprehension”or“mathematics”necessary for all occupations available.

90、Figure 2describes their statistical properties in our dataset.The level variable“indicates the degree,orpoint along a continuum,to which a particular descriptor is required or needed to perform theoccupation”14on a scale from 0(min)to 6(max).The right-hand side of Figure 2 displays a histogram of(di

91、fficulty)“Level variable.”Thislevel can be interpreted as the difficulty level of the skill required to perform the occupation:“While the same skill can be important for a variety of occupations,the amount or level ofthe skill needed in those occupations can differ dramatically.For example,the skill

92、“speaking”12Note that O*NET provides several tables with data classified by level or difficulty,such as abilities and workactivities tables.We test the results for those alternative tables in Section 5.13The number of skills varies at each occupation between 24 and 35 with an average of 32.4 and a m

93、edian of33 skills.14See O*NET website:https:/www.onetonline.org/help/online/scales.Website accessed in June 2024.9is important for both lawyers and paralegals.However,lawyers(who frequently argue casesbefore judges and juries)are required to have a higher level of speaking skill,while paralegalsonly

94、 need an average level of this skill.”15Table A2 in the Appendix presents the occupations most closely matching select skill levelsfor four cognitive skills.For example,“Troubleshooting”skills are required with a low levelfor“Social and Community Service Managers”whereas“Aircraft Mechanics and Servi

95、ce Tech-nicians”require much higher levels to fulfil the high standards in their occupation.“Writersand Authors”do not require high skills in“Mathematics”while“Mathematicians”use it as thesource of their living.Within each occupation,skills are also classified by the“importance”.This variable,“in-di

96、cates the degree of importance a particular descriptor is to the occupation”.The left-handpanel in Figure 2 displays a histogram of the Importance variable.16The possible ratings rangefrom 1(“Not Important”)to 5(“Extremely Important”).Figure 2:Distributions of the skills importance and level variabl

97、es(A)Distribution of the importance Ii,o,s(B)Distribution of the levels Li,o,sNote:The figure shows the distribution of the importance and level variables in the skills dataset from O*NET.Importance:N=23,111;Mean=2.691;SD=0.765;Min=1.04;Max=5.Level:N=23,111;Mean=2.510;SD=1.101;Min=0.060;Max=6.In our

98、 simulations,we assume that cognitive skills that involve computer interactions areexposed once their level comes within the AIs capabilities.17In the main analysis,we do not attempt to directly quantify current AI capabilities(however,we compare our results to the estimates available in the literat

99、ure(subsection 3.3)and add15See O*NET website:https:/www.onetonline.org/help/online/scales.Website accessed in June 2024.16See O*NET website:https:/www.onetonline.org/help/online/scales.Website accessed in June 2024.17Certainly,this is a simplification and does not take into account inherent differe

100、nces in skill levels(forinstance,an AI might excel at maths much sooner than on negotiation)nor the difference in time and effort toautomate different skills with an AI.We also do not account for costs to the implementation that can significantlyaffect the decision to automate the processes,as argue

101、d in(Svanberg et al.,2024).10multiple back-of-the-envelope calculations in Section 5).Rather,we introduce this as a modelvariable(AI),based on which we can construct different scenarios with lower or higher AIcapabilities.Although we can not foresee the speed of AI technology advancement and adoptio

102、n,we are safe to assume that this parameter will grow in the future.On a per-skill level,we construct the binary variable Ai,o,s(AI)to indicate if the skill s ofthe occupation o in industry i is with the AIs capabilities(AI)and thus prone to automation.Ai,o,s(AI)=0if Li,o,s AI,1if Li,o,s AI.The over

103、all share of skills that can be automated for a given occupation(o)as a functionof the AI capability Ai,o(AI)is equal to the weighted average of the impact of each of theoccupations skills,weighted by its importance(Ii,o,s)for the given occupation:Ai,o(AI)=PsSAi,o,s(AI)Ii,o,sPsSIi,o,sThe same share

104、on the industry level is again calculated as an average of the shares of theindividual occupations(Ai,o)weighted by the employment numbers(Ni,o)for each occupation:Ai(AI)=PoOAi,o(AI)Ni,oPoONi,oFigure 3 illustrates the share of skills within AI capabilities across industries for three low(2.0),medium

105、(3.0)and high(4.0)AI capabilities.We see that even moderate levels of AIsurpass more than 50%of the required skills necessary for many labour-intensive occupations,most strikingly,“Buildings and grounds cleaning and maintenance”,where even low AI capabil-ities already surpass 40%of the required skil

106、ls.On the opposite end,traditionally office-proneindustries such as“Engineering”,“Management”,or“Legal”show a comparatively low shareof around 20%of affected skill at medium AI levels and even at high AI capability levels doesnot surpass 80%.11Figure 3:Share of automatable cognitive skills given AI

107、capabilities Ai(AI)Note:The figure shows the share of skills Ai(AI)within reach of the low(2.0),medium(3.0)and high(4.0)AIcapability AIin individual industries i.3.3AISA Index-combining computer interaction and automatability ofskillsThe combination of the time spent with computer interaction Ti,oan

108、d the occupational skillswithin AI capabilities Ai,o(AI)constitutes the AI share of automation(AISA)index:AISAi,o(AI)=Ti,o Ai,o(AI)Again,we aggregate the AISA index to the industry level,weighted by the employmentnumbers.18AISAi(AI)=PoOAISAi,o(AI)Ni,oPoONi,oFigure 4 presents AISAi(AI=3.0)(red)within

109、 individual industries as the combination ofthe fraction of time spent in computer interaction Ti(yellow)and the fraction of skills within theAIs capabilities Ai(AI=3.0)(blue).We see that industries like Production or Farming are“technology-limited”(i.e.,despite simpler skills,much work does not inv

110、olve direct computer18We implicitly assume(statistic)independence between the time spent on computer interaction and individualskills.We acknowledge that this is a simplification,as some of the skills,such as“writing”,are more likelyperformed through computer interactions than others(e.g.“negotiatio

111、n”).Considering the large amount ofuncertainty,we opted for this simple model.However,as more data on real-world AI usage becomes available,weighing the individual skills separately would likely be a point of refinement.12interaction,thus reducing automation potential).Conversely,industries like Leg

112、al services orengineering are“skill-level-limited”;despite high computer interaction,the complexity of skillsoften surpasses the AIs capabilities.At a moderate AI capability of 3.0,we observe a moderateAI impact of 10-25%19across the industries,with slightly greater effects in skill-level-limitedind

113、ustries.A noteworthy exception is the office and administrative support industry,withan almost 40%AI exposure attributable to both high computer usage and relatively low skilldifficulties.Moreover,traditionally well-paid office jobs such as Engineering or Legal have high sharesof computer interactio

114、n and thus would lend themselves to easy employment of AI,assumingthat the AI has the required skill levels and quality.On the other hand,there are traditionallylow-paid(physical)labour-intensive jobs such as“food serving”or“cleaning”,requiring lesscognitive skills,but are difficult to leverage AI i

115、n their tasks because of their working modalities,which are mostly not in direct contact with a computer.Figure 4:AISA Index at AI capability 3.0 at the industry levelNote:The figure presents AI share automation index AISAi(AI=3)as a red bar,computer interaction Tias a bluebar,and share of skills au

116、tomatable by AI Ai(AI=3)as a yellow bar across individual industries i.AISAi,o(AI=3)is calculated by multiplying Ti,oand Ai,o(AI=3)on occupational level.Figure 5 shows the dependency of the AISA index(red bars)on the AI capability for thethree AI levels(low,medium,and high).Per construction,the AISA

117、 can not surpass thefraction of time spent in computer interaction Ti(yellow lines),even at the highest AI capa-bilities.Industries with a high portion of computer interaction,such as Legal or Architecture,19The average AISAavg(AI=3.0)weighted by employment numbers is 18.4%,computer interaction Tavg

118、43.3%and skills below AI capability Aavg(AI=3.0)=56.1%.13converge slowly towards their maximum,whereas occupations with low computer interaction(e.g.Production or Transport)have already almost reached their maximum at medium AIcapabilities.20Figure 5:AISAi(AI)at low(2.0),medium(3.0)and high(4.0)AI c

119、apabilityNote:The figure represents the AISAi(AI)index for three different AI capabilities(2.0,3.0,4.0)at the industry level.The AISA index is dependent on the AI capability(AI)up to a maximum level given by the computer interaction Ti(yellow lines).Our quantification of the AISA index relates to ot

120、her measures of the occupation-levelexposure of AI.Felten et al.(2021)develop an AI-occupational exposure(AIOE)measure thatgauges how important certain AI-exposed tasks are for a given occupation.21Pizzinelli et al.(2023)augment this measure(AIOE measures exposure to AI,but not whether AI complement

121、sor substitutes human labour)by a variable that measures the risk of replacement.The lattervariable is constructed at the occupational level by looking into selected parts of the“workcontext,”defined in O*NET as physical and social factors that influence the nature of work.Webb(2020),on the other ha

122、nd,matches occupational task descriptions with the text of patentsto match the potential impact and a similar approach was done by Yang(2022).Tolan et al.20The overall average AISAavg(AI=2)=7.3%,AISAavg(AI=3)=18.4%,and AISAavg(AI=4)=43.3%.The employment numbers were used as weights(see below).21In m

123、ore detail,Felten et al.(2021)combine three data sources to estimate a measure of AI occupationalexposure(AIOE).First,based on information provided by the Electronic Frontier Foundation,they identify 10applications in which AI had made“meaningful scientific progress”as of the date of writing.These a

124、pplicationsinclude real-time video games,recognition and creation of speech and images,and translation.Second,these 10applications are linked to 52 occupational abilities in the O*NET data based on a crowdsourced survey,resultingin an AI exposure measure for each of the 52 abilities(i.e.,to what ext

125、ent a certain ability will,in total,beexposed to the 10 applications).Lastly,AIOE is calculated as the weighted sum of the ability-level exposures,using the O*NET measures of“importance”and“prevalence”for each ability as weights.14(2021)mixes combinations of tasks and abilities with AI evaluation ta

126、sks from AI benchmarks.Brynjolfsson et al.(2018)chose a different approach.They define exposure by matching anestablished rubric with tasks and direct work activities from the O*NET database.They useda survey to establish the exposure.Eloundou et al.(2023)followed a similar approach but focuson a si

127、gnificant reduction in time to completion.They also tested an alternative approach usingChatGPT(this was also replicated by Eisfeldt et al.(2023).Our simulations raise the question regarding the current level of AI capabilities,which canbe gauged from comparisons of the results of our hypothetical c

128、alculations with the survey-based ones in the literature.The red dotted line in Figure 6 displays AISAavg(AI),defined asthe overall average(weighted across occupational employment as weights)as a function of AIcapability(AI):AISAavg(AI)=PoOAISAi,o(AI)Ni,oPoONi,oFor such comparisons,we note that our

129、approach is conservative in that we assume thatonly computer interaction is automatable.AISAavg(AI)(red dotted line)increases graduallyand tops out at 43.3%,which is the average level of computer interaction across all industriesthat bounds AISA.We compare AISAavg(AI)to four datapoints found in the

130、literature.Hatzius et al.(2023)and Eloundou et al.(2023)suggest that on average 25%22and 30%23of occupations will beexposed respectively.A level of AIof 3.2 and 3.6,respectively,would match these numbers inthe aggregate.An earlier literature had suggested a larger automatisation potential that excee

131、ds the maxi-mum of AISAavg(AI)in our simulations.Frey&Osborne(2017)estimate a maximum exposureof 47%24,while Webb(2020)estimates an exposure of 54.5%25.Some of these higher estimates may rely on the assumption that also types of work otherthan computer interaction are automatable with AI.For better

132、comparability,Figure 6 alsoreports an alternative simulation for AISAavg(AI)that is discussed in the robustness Section 5.This alternative measure assumes that in addition to computer interaction,also a share of socialinteractions,such as communication with clients via e-meetings or taking orders in

133、 a restaurant,can be automated with AI:T25%Sociali,ois calculated assuming that 25%of social interaction canbe automated(if the skill difficulties are in reach of the AIs capabilities).AISAT25%Sociali,oavg(AI)is depicted by the red dashed line in Figure 6 and tops out at 51.8%.22The value represents

134、 AI exposure based on an evaluation of 13 work activities of O*NET.23This value represents the mean AI exposure based on a human assessment of the professions in which AI canreduce the time spent on tasks by at least 50%only with the use of the OpenAIs ChatGPT(alone,or in their“beta”case when it is

135、integrated in a companys systems).24Note that this figure is discussed in the literature.Arntz et al.(2016)revise these estimates using a modifiedmethodology and find that only 9%of jobs in the US are at high risk of automation.25The value represents the exposure to AI.Webb(2020)also models the expo

136、sure to software(50.69%)androbots(48.61%).15Using this line as a benchmark,an AI capability level of 2.9 and 3.1,respectively,would matchthe averages mentioned in Eloundou et al.(2023)and Hatzius et al.(2023).In contrast,re-producing the figure of Frey&Osborne(2017)would require a very high AI capab

137、ility of 4.1(while there is no level of AI capabilities that allows us to reproduce the predictions of Webb(2020).Figure 6:AISA as a function o f AIand comparison to other estimatesNote:The figure presents the level of AISAavg(AI)as a function of AI capability AI.The average AISA is constructed asth

138、e weighted mean across industries,using employment numbers of each occupation as weights.AISAavg(AI)isrepresented by the red dotted line.The figure also depicts an alternative measure from the robustness exercise insubsection 5.1,assuming that AI can automate not only computer interaction but also 2

139、5%of social interactions(T25%Sociali,o).AISAT25%Sociali,oavg(AI)is depicted by the red dashed line.The figure also adds four lines corresponding toaverage AI exposure levels found in the literature(Eloundou et al.(2023)(black dashed vertical line),Hatzius et al.(2023)(brown dotted vertical line),Fre

140、y&Osborne(2017)(black dot-dash horizontal line),and Webb(2020)(browntriangular horizontal line).Figure 7 maps the AISA index across the wage spectrum.To construct the wage spectrum,we used weighted quantiles to take differences in the numbers of employees per occupation Nointo account.Assignment to

141、a wage quantile was done by first indexing(j)all occupations inascending order of their wage Wo,jand calculating the cumulative sum of employees leading upto the index of the particular occupation j=jo,divided by the total number of employees inthe dataset:26po=Pj=joj=1No,jPjJNo,j26Imagine,lining up

142、 all employees based on their wage,then powould denote the relative position as a fractionof the entire length of the line of the last representative of a given occupation o in that line.16Based on that relative position,an occupation is assigned to wage quantile w from a totalof W quantiles if its

143、relative position is in between equally spaced boundaries:w 1W AI,1if Labilitieso,s AIAabilitieso(AI)enables us to calculate a new AISA index and complementarity measureCo(AI)that reflects cognitive abilities instead of skills.32Figure 15 compares Co(AI)using cognitive abilities tables instead of sk

144、ills with the baselineresults.Cognitive abilities(dashed lines)yield outcomes similar to the baseline model(solidlines),with peak complementarity for each quartile closely matching the baseline values andmaintaining the same position relative to AI capability AI.The average complementarityeffect acr

145、oss all quartiles is,however,lower by 2.0pp(11.2%,compared to 13.2%in the baselinemodel).Figure 15:Cognitive abilities instead of skillsNote:The figure compares complementarity effects Cw(AI)of the baseline model with the alternative approach thatuses the cognitive abilities table instead of the ski

146、lls table from O*NET.Each panel shows the evolution of Cw(AI)forindividual wage quartiles against higher progressing AI capabilities AI.The solid lines represent the baseline model andare identical to the results presented in the right panel of Figure 11.The dashed lines represent Cabilitiesw(AI)whe

147、nAabilitieso(AI)is used instead of Askillso(AI).O*NET defines the work activities as“general types of job behaviors occurring in multiplejobs.”33The O*NET website categorises these activities into four groups:i)information input,ii)interacting with others,iii)mental processes,and iv)work output.Thes

148、e categories encom-pass a variety of tasks,from“Coordinating the Work and Activities of Others”and“Developing32Note that we keep our baseline approach to identify core and side variables:the top 33%of cognitiveabilities/work activities are identified as”core”in each occupation.33see https:/www.oneto

149、nline.org/find/descriptor/browse/4.A The website was accessed in June 2024.27Objectives and Strategies”to“Identifying Objects,Actions,and Events”and“Interpreting theMeaning of Information for Others”.Although there is significant variation among these workactivities,we exclude“work output”only due t

150、o its focus on physical activities,which are notrelevant to AI interactions in our scope.We end up with 41 different work activities.Thedifficulty level Lwork act.o,sand importance Iwork act.o,sof these activities are defined similarly to theO*NET skills table,but the difficulty level is set with a

151、range of 0 to 7.We rescaled this toa 0 to 6 range to retain compatibility of the results with our baseline model.The importancevariable Iwork act.o,smaintains the same range as in the skills table,i.e.from 1 to 5.We calculate Awork act.o(AI)in the same fashion as shown previously,i.e.by utilising th

152、elevel Lwork act.o,sand importance Iwork act.o,svariables from the work activities tables from O*NET:Awork act.o(AI)=PsSAwork act.o,s(AI)Iwork act.o,sPsSIwork act.o,s;Awork act.o,s(AI)=0if Lwork act.o,s AI,1if Lwork act.o,s AIFigure 16 compares Co(AI)of the baseline results with the alternative case

153、 that uses re-stricted work activities dataset.The work activities show more variation compared to the resultsfrom cognitive abilities,with peak levels averaging 3.3pp lower and occurring around 0.25 AIlater.Despite these deviations,our findings remain robust to changes in the tables used fromthe O*

154、NET.Figure 16:Work activities instead of skillsNote:The figure compares complementarity effects Cw(AI)of the baseline model with the alternative approach thatuses work activities table instead of skills.Each panel shows the evolution of Cw(AI)for individual wage quartilesagainst the progressing AI c

155、apabilities AI.The solid lines represent the baseline model and are identical to the resultspresented in Figure 11 right panel.The dashed lines represent Cwork act.w(AI)when Awork act.o(AI)is used instead ofAskillso(AI).28The comparisons of AISA between the baseline model and the alternatives that u

156、se eithercognitive abilities or work activities instead of skills reveal no substantial differences betweenthe approaches,as shown in Figure 17.Cognitive abilities are only 3.53 percentage points loweron average,and the coefficient of determination R2reaches 0.98.The comparison with the workactiviti

157、es provides very similar results.Figure 17:AISA in comparison to alternative models(skills-oriented)Note:The figure presents scatter plots of AISA(as used in Figure 7)from the baseline model(x-axis)and AISAcalculated with the alternative setups of the baseline model used for the robustness analysis(

158、y-axis).We consider twodifferent setups:i)cognitive abilities table instead of skills table,and ii)work activities table instead of skills table.Panels also display summaries of simple OLS regression lines.5.3Using different definitions of core and side skillsAs a last robustness exercise,we investi

159、gate the robustness of the results with regard to changingthe definition for core and side skills.The baseline approach defines core skills as the top 33.3%of the most important skills for an occupation.34We construct two alternative models to inspectthe impact of different ratios used to identify t

160、he core skills on complementarity Co(AI).The first model defines core skills as the top 50%based on importance,thus including abroader set of skills.The second model narrows the core skills to the top 20%.Figure 18 com-pares the baseline results with these alternative approaches on the evolution of

161、complementarityCo(AI).When 50%of skills are considered core(dotted lines),Co(AI)peaks at a slightly lowerAI capability(on average by 0.5)and is more pronounced,with a maximum difference of upto 5 percentage points.Conversely,when only 20%of the most important skills are consideredcore(dashed lines),

162、the dynamics reverse:Co(AI)improves more gradually,and the maximumcomplementarity is slightly reduced compared to the baseline model.34This approach mimics O*NET methodology to define core tasks(see subsection 3.2)29Figure 18:Different core skills identification processNote:The figure compares compl

163、ementarity effects Cw(AI)of the baseline model with the two alternative specificationsof core skills.The baseline model considers core skills to be the top 33.3%of the most important skills for an occupation.The first alternative model defines core skills as the top 50%most important skills(broader

164、definition of core skills)andis presented with dotted lines.The second alternative model defines core skills as the top 20%most important skills(amore narrow definition of core skills)and is presented with dashed lines.The solid lines represent the baseline model andare identical to the results pres

165、ented in Figure 11 right panel.Each panel shows the evolution of Cw(AI)for anindividual wage quartile against the progressing AI capabilities AI.6ConclusionIn this paper,we have explored the impact of AIs evolving capabilities on the labour market,focusing on the exposure of 711 US occupational cate

166、gories and the potential of AI to comple-ment or substitute human labour.We model how the share of an occupations skills exposedto AI depends on the difficulty of these skills and the AIs cognitive capability.Our analysisdistinguishes between the impacts on core and side skills and investigates diff

167、erential exposuresacross wage quartiles.We first investigate the overall exposure of industries across occupations and the wage dis-tribution.We find that AI may initially affect occupations uniformly across the wage spectrum,impacting approximately 7%of skills at lower AI capability levels.However,

168、as capabilitiesimprove,up to 45%of skills in the highest wage quartile are susceptible to automation by AI,compared to only 26%in the lowest quartile.Nevertheless,looking into the impact on core and side skills,we find that AI may still leadto increasing inequality as it will tend to substitute low-

169、wage work more easily than high-wagework.We find that low AI capabilities complement all workers,as side skills are simpler thancore skills.However,as AI capabilities advance,core skills in lower-wage jobs become exposed,threatening substitution and increased inequality.In contrast to the intuitive

170、notion that the30rise of AI may harm white-collar workers,we find that those remain safe as their core skillsare hard to automate.Yet workers will be forced to adapt to these heavily AI-supported newworking modalities.They will need to learn to trust AI-enabled services supporting them intheir side

171、tasks while focusing more of their works attention on their core tasks.The simulations in this paper provide a detailed view of who will be impacted by AI andin what ways,offering valuable insights for economic policy formulation.As AI continues toreshape the labour market,several considerations nee

172、d careful evaluation.Enhancing skill de-velopment and training is crucial for workers at risk of AI-driven displacement,which can befocused on adapting workforce capabilities in the affected skill areas and occupations.Trans-parency in AI deployment and involving workers in implementation decisions

173、are essential to en-sure that AI complements rather than replaces human labour.Additionally,establishing safetynets and transition programmes can support those adversely affected by AI.On a broader scale,strengthening international cooperation on AI labour policies allows for a unified approach toma

174、naging AIs impact globally.These considerations are vital for policymakers to balance thebenefits and challenges of AI in the workforce.31ReferencesAcemoglu,D.(2024):“The simple macroeco-nomics of ai.”Economic Policy.Acemoglu,D.,D.Autor,J.Hazell,&P.Re-strepo(2022):“Artificial intelligence and jobs:E

175、vidence from online vacancies.”Journal of La-bor Economics 40(S1):pp.S293S340.Acemoglu,D.&P.Restrepo(2018a):“Artifi-cial intelligence,automation,and work.”In“Theeconomics of artificial intelligence:An agenda,”pp.197236.University of Chicago Press.Acemoglu,D.&P.Restrepo(2018b):“Low-skill and high-ski

176、ll automation.”Journal of Hu-man Capital 12(2):pp.204232.Acemoglu,D.&P.Restrepo(2018c):“The racebetween man and machine:Implications of tech-nology for growth,factor shares,and employ-ment.”American economic review 108(6):pp.14881542.Acemoglu,D.&P.Restrepo(2019):“Automa-tion and new tasks:How techno

177、logy displacesand reinstates labor.”Journal of economic per-spectives 33(2):pp.330.Acemoglu,D.&P.Restrepo(2020):“Robotsand jobs:Evidence from us labor markets.”Journal of political economy 128(6):pp.21882244.Acemoglu,D.&P.Restrepo(2022):“Tasks,automation,and the rise in us wage inequality.”Econometr

178、ica 90(5):pp.19732016.Adachi,D.,D.Kawaguchi,&Y.U.Saito(2024):“Robots and employment:Evidencefrom japan,19782017.”Journal of Labor Eco-nomics 42(2):pp.591634.Albanesi,S.,A.D.Da Silva,J.F.Jimeno,A.Lamo,&A.Wabitsch(2023):“New tech-nologies and jobs in europe.”Technical report,National Bureau of Economi

179、c Research.Aldasoro,I.,O.Armantier,S.Doerr,L.Gambacorta,&T.Oliviero(2024):“Thegen ai gender gap.”Economics Letters 241:p.111814.Arntz,M.,T.Gregory,&U.Zierahn(2016):“The risk of automation for jobs in oecd coun-tries:A comparative analysis.”OECD So-cial,Employment,and Migration Working Pa-pers(189):p

180、.0 1.Autor,D.(2019):“Work of the past,work ofthe future.”Technical report,National Bureauof Economic Research.Autor,D.H.(2015):“Why are there still so manyjobs?the history and future of workplace au-tomation.”Journal of Economic Perspectives29(3):pp.330.Autor,D.H.&D.Dorn(2013):“The growthof low-skil

181、l service jobs and the polarization ofthe us labor market.”American economic review103(5):pp.15531597.Bessen,J.E.(2016):“How computer automa-tion affects occupations:Technology,jobs,andskills.”Boston Univ.school of law,law and eco-nomics research paper(15-49).Brynjolfsson,E.,D.Li,&L.R.Raymond(2023):

182、“Generative ai at work.”Technical re-port,National Bureau of Economic Research.Brynjolfsson,E.,T.Mitchell,&D.Rock(2018):“What can machines learn and whatdoes it mean for occupations and the economy?”In“AEA papers and proceedings,”volume 108,pp.4347.American Economic Association 2014Broadway,Suite 30

183、5,Nashville,TN 37203.Cazzaniga,M.,F.Jaumotte,L.Li,G.Melina,A.J.Panton,C.Pizzinelli,E.J.Rockall,&M.M.Tavares(2024):“Gen-ai:Artificial intel-ligence and the future of work.”Staff DiscussionNotes 2024(001).Cheng,Z.,D.Lee,&P.Tambe(2022):“Innovae:Generative ai for understanding patents and in-novation.”A

184、vailable at SSRN 3868599.Cockburn,I.M.,R.Henderson,&S.Stern(2018):The impact of artificial intelligence oninnovation,volume 24449.National bureau ofeconomic research Cambridge,MA,USA.Dekle,R.(2020):“Robots and industrial labor:Evidence from japan.”Journal of the Japaneseand International Economies 5

185、8:p.101108.Eisfeldt,A.L.,G.Schubert,&M.B.Zhang(2023):“Generative ai and firm values.”Tech-nical report,National Bureau of Economic Re-search.32Eloundou,T.,S.Manning,P.Mishkin,&D.Rock(2023):“Gpts are gpts:An early lookat the labor market impact potential of large lan-guage models.”arXiv preprint arXi

186、v:2303.10130.Felten,E.,M.Raj,&R.Seamans(2021):“Oc-cupational,industry,and geographic exposure toartificial intelligence:A novel dataset and itspotential uses.”Strategic Management Journal42(12):pp.21952217.Frey,C.B.&M.A.Osborne(2017):“The fu-ture of employment:How susceptible are jobs tocomputerisat

187、ion?”Technological forecasting andsocial change 114:pp.254280.Furman,J.&R.Seamans(2019):“Ai and theeconomy.”Innovation policy and the economy19(1):pp.161191.Gmyrek,P.,J.Berg,&D.Bescond(2023):“Generative ai and jobs:A global analysis of po-tential effects on job quantity and quality.”ILOWorking Paper

188、 96.Hatzius,J.et al.(2023):“The potentially large ef-fects of artificial intelligence on economic growth(briggs/kodnani).”Goldman Sachs.Korinek,A.&M.Juelfs(2022):“Preparingfor the(non-existent?)future of work.”OxfordHandbook of AI Governance.Michaels,G.,A.Natraj,&J.Van Reenen(2014):“Has ict polarize

189、d skill demand?ev-idence from eleven countries over twenty-fiveyears.”Review of Economics and Statistics96(1):pp.6077.Noy,S.&W.Zhang(2023):“Experimental ev-idence on the productivity effects of generativeartificial intelligence.”Science 381(6654):pp.187192.Peng,S.,E.Kalliamvakou,P.Cihon,&M.Demirer(2

190、023):“The impact of ai on de-veloper productivity:Evidence from github copi-lot.”arXiv preprint arXiv:2302.06590.Perez-Cruz,F.&H.S.Shin(2024):“Testing thecognitive limits of large language models.”Tech-nical report,Bank for International Settlements.Pizzinelli,C.,A.J.Panton,M.M.M.Tavares,M.Cazzaniga

191、,&L.Li(2023):Labor mar-ket exposure to AI:Cross-country differences anddistributional implications.International Mone-tary Fund.Svanberg,M.,W.Li,M.Fleming,B.Goehring,&N.Thompson(2024):“Beyond AI exposure:Which tasks are cost-effective to automate with computer vision?”Working paper,MIT.Tolan,S.,A.Pe

192、sole,F.Mart nez-Plumed,E.Fern andez-Mac as,J.Hern andez-Orallo,&E.G omez(2021):“Measuring theoccupational impact of ai:tasks,cognitive abil-ities and ai benchmarks.”Journal of ArtificialIntelligence Research 71:pp.191236.Webb,M.(2020):“The impact of artificial intelli-gence on the labor market.”Avai

193、lable at SSRN3482150.Yang,C.-H.(2022):“How artificial intelligencetechnology affects productivity and employment:firm-level evidence from taiwan.”Research Pol-icy 51(6):p.104536.33AAppendixTable A1:Tasks descriptions for the occupation Fundraising Managers fromO*NET databaseNo.Task Description1Assig

194、n,supervise,and review the activities of fundraising staff.2Compile or develop materials to submit to granting or other funding organisations.3Conduct research to identify the goals,net worth,charitable donation history,or otherdata related to potential donors,potential investors,or general donor ma

195、rkets.4Contact corporate representatives,government officials,or community leaders to in-crease awareness of organisational causes,activities,or needs.5Design and edit promotional publications,such as brochures.6Develop fundraising activity plans that maximise participation or contributions andminim

196、ise costs.7Develop strategies to encourage new or increased contributions.8Direct activities of external agencies,establishments,or departments that develop andimplement fundraising strategies and programs.9Establish and maintain effective working relationships with clients,government offi-cials,and

197、 media representatives and use these relationships to develop new fundraisingopportunities.10Establish goals for soliciting funds,develop policies for collection and safeguarding ofcontributions,and coordinate disbursement of funds.11Evaluate advertising and promotion programs for compatibility with

198、 fundraising ef-forts.12Formulate policies and procedures related to fundraising programs.13Manage fundraising budgets.14Plan and direct special events for fundraising,such as silent auctions,dances,golfevents,or walks.15Produce films and other video products,regulate their distribution,and operate

199、filmlibrary.16Write interesting and effective press releases,prepare information for media kits,anddevelop and maintain company internet or intranet Web pages.34Table A2:Examples for Skill Levels:Corresponding Occupations from O*NETdatabaseAI capability/SkillsTroubleshootingCritical ThinkingActive L

200、isteningMathematics1.5Social and CommunityService ManagersCleaners of Vehiclesand EquipmentPressers,Textile,Garment,andRelated MaterialsWriters and Authors2.0General and OperationsManagers-Human ResourcesSpecialists2.5Computer andInformation SystemsManagersCourt Reporters andSimultaneousCaptionersTe

201、rrazzo Workers andFinishersManagers,All Other3.0Industrial ProductionManagersFarm LaborContractorsCooks,Fast FoodAdvertising andPromotions Managers3.5Industrial EngineersProperty,RealEstate,andCommunityAssociation ManagersArchitectural andEngineeringManagersFinancial Managers4.0Electro-Mechanical an

202、dMechatronicsTechnologistsAdvertising andPromotions ManagersChief ExecutivesCost Estimators4.5Aircraft Mechanics andService TechniciansAnesthesiologistsLabor RelationsSpecialistsOperations ResearchAnalysts5.0-Judges,MagistrateJudges,andMagistratesJudges,MagistrateJudges,andMagistratesMathematicians3

203、5Figure A1:Importance of Core and Side Tasks according to O*NET databaseNote:The figure supports our proposition that the skills can be split into core and side skills based on their importanceas core tasks are more important than the side tasks on occupational average.Box A1:Prompt to define fracti

204、ons of time spent in occupations“Below is a list of task descriptions for the profession of occupation:tasks:With this description compile aJSON file with an estimate of how much of the worktime is spent on:1.Working on computer2.Talking to people3.Physical activitiesNote:The prompt was looped acros

205、s all occupations in the dataset.occupation and tasks represent twovariables varying in the prompt.36Figure A2:Comparison between computer interaction and AIOE on individualO*NET occupationsNote:Figure presents the relationship between computer interaction variable and AIOE developed by Felten et al

206、.(2021).The correlation between these variables is notably high at 86.3%.37 Previous volumes in this series 1206 August 2024 Covered interest parity:a forecasting approach to estimate the neutral band Juan R.Hernndez 1205 August 2024 The Measure Matters:Differences in the Passthrough of Inflation Ex

207、pectations in Colombia Andres Sanchez-Jabba and Erick Villabon-Hinestroza 1204 August 2024 Climate Policies,Labor Markets,and Macroeconomic Outcomes in Emerging Economies Alan Finkelstein Shapiro and Victoria Nuguer 1203 August 2024 Strike while the Iron is Hot:Optimal Monetary Policy with a Nonline

208、ar Phillips Curve Peter Karadi,Anton Nakov,Galo Nuno,Ernesto Pasten,and Dominik Thaler 1202 August 2024 Are low interest rates firing back?Interest rate risk in the banking book and bank lending in a rising interest rate environment Lara Coulier,Cosimo Pancaro and Alessio Reghezza 1201 July 2024 Cry

209、pto Exchange Tokens Rodney Garratt,Maarten R.C.van Oordt 1200 July 2024 Financial inclusion transitions in Peru:does labor informality play a role?Jose Aurazo and Farid Gasmi 1199 July 2024 New spare tires:local currency credit as a global shock absorber Stefan Avdjiev,John Burger and Bryan Hardy 11

210、98 July 2024 Sovereign green bonds:a catalyst for sustainable debt market development?Gong Cheng,Torsten Ehlers,Frank Packer and Yanzhe Xiao 1197 July 2024 The gen AI gender gap Iaki Aldasoro,Olivier Armantier,Sebastian Doerr,Leonardo Gambacorta and Tommaso Oliviero 1196 July 2024 Digital payments,i

211、nformality and economic growth Ana Aguilar,Jon Frost,Rafael Guerra,Steven Kamin and Alexandre Tombini 1195 July 2024 The asymmetric and persistent effects of Fed policy on global bond yields Tobias Adrian,Gaston Gelos,Nora Lamersdorf,Emanuel Moench 1194 June 2024 Intelligent financial system:how AI is transforming finance Iaki Aldasoro,Leonardo Gambacorta,Anton Korinek,Vatsala Shreeti and Merlin Stein 1193 June 2024 Aging gracefully:steering the banking sector through demographic shifts Christian Schmieder and Patrick A Imam All volumes are available on our website www.bis.org.