《國際勞工組織：2023生成式AI和工作報告：對工作效率和質量的潛在影響的全球分析（英文版）（55頁）.pdf》由會員分享，可在線閱讀，更多相關《國際勞工組織：2023生成式AI和工作報告：對工作效率和質量的潛在影響的全球分析（英文版）（55頁）.pdf（55頁珍藏版）》請在三個皮匠報告上搜索。

1、 XGenerative AI and Jobs:A global analysis of potential effects on job quantity and qualityAuthors/Pawe Gmyrek,Janine Berg,David Bescond August/2023ILO Working Paper 96Copyright International Labour Organization 2023This is an open access work distributed under the Creative Commons Attribution 4.0 I

2、nternational License(https:/creativecommons.org/licenses/by/4.0/).Users can reuse,share,adapt and build upon the original work,as detailed in the License.The ILO must be clearly credited as the own-er of the original work.The use of the emblem of the ILO is not permitted in connection with users wor

3、k.Attribution The work must be cited as follows:Gmyrek,P.,Berg,J.,Bescond,D.Generative AI and Jobs:A global analysis of potential effects on job quantity and quality.ILO Working Paper 96.Geneva:International Labour Office,2023.Translations In case of a translation of this work,the following disclaim

4、er must be added along with the attribution:This translation was not created by the International Labour Organization(ILO)and should not be considered an official ILO translation.The ILO is not responsible for the content or accuracy of this translation.Adaptations In case of an adaptation of this w

5、ork,the following disclaimer must be added along with the attribution:This is an adaptation of an original work by the International Labour Organization(ILO).Responsibility for the views and opinions expressed in the adaptation rests solely with the author or authors of the adaptation and are not en

6、dorsed by the ILO.This CC license does not apply to non-ILO copyright materials included in this publication.If the material is attributed to a third party,the user of such material is solely responsible for clearing the rights with the right holder.Any dispute arising under this license that cannot

7、 be settled amicably shall be referred to arbitra-tion in accordance with the Arbitration Rules of the United Nations Commission on International Trade Law(UNCITRAL).The parties shall be bound by any arbitration award rendered as a result of such arbitration as the final adjudication of such a dispu

8、te.All queries on rights and licensing should be addressed to the ILO Publishing Unit(Rights and Licensing),1211 Geneva 22,Switzerland,or by email to rightsilo.org.ISBN 9789220395356(print),ISBN 9789220395363(web PDF),ISBN 9789220395370(epub),ISBN 9789220395387(mobi),ISBN 9789220395394(html).ISSN 27

9、08-3438(print),ISSN 2708-3446(digital)https:/doi.org/10.54394/FHEM8239 The designations employed in ILO publications,which are in conformity with United Nations practice,and the presentation of material therein do not imply the expression of any opinion whatsoever on the part of the ILO concerning t

10、he legal status of any country,area or territory or of its authorities,or concerning the delimitation of its frontiers.The responsibility for opinions expressed in signed articles,studies and other contributions rests solely with their authors,and publication does not constitute an endorsement by th

11、e ILO of the opinions expressed in them.Reference to names of firms and commercial products and processes does not imply their en-dorsement by the ILO,and any failure to mention a particular firm,commercial product or pro-cess is not a sign of disapproval.Information on ILO publications and digital

12、products can be found at:www.ilo.org/publnsILO Working Papers summarize the results of ILO research in progress,and seek to stimulate discussion of a range of issues related to the world of work.Comments on this ILO Working Paper are welcome and can be sent to RESEARCHilo.org,bergilo.org.Authorizati

13、on for publication:Richard Samans,Director RESEARCHILO Working Papers can be found at:www.ilo.org/global/publications/working-papersSuggested citation:Gmyrek,P.,Berg,J.,Bescond,D.2023.Generative AI and Jobs:A global analysis of potential ef-fects on job quantity and quality,ILO Working Paper 96(Gene

14、va,ILO).https:/doi.org/10.54394/FHEM823901 ILO Working Paper 96AbstractThis study presents a global analysis of the potential exposure of occupations and tasks to Generative AI,and specifically to Generative Pre-Trained Transformers(GPTs),and the possible implications of such exposure for job quanti

15、ty and quality.It uses the GPT-4 model to estimate task-level scores of potential exposure and then estimates potential employment effects at the global level as well as by country income group.Despite representing an upper-bound estimate of exposure,we find that only the broad occupation of clerica

16、l work is highly exposed to the tech-nology with 24 per cent of clerical tasks considered highly exposed and an additional 58 percent with medium-level exposure.For the other occupational groups,the greatest share of highly ex-posed tasks oscillates between 1 and 4 per cent,and medium exposed tasks

17、do not exceed 25 per cent.As a result,the most important impact of the technology is likely to be of augmenting work automating some tasks within an occupation while leaving time for other duties as op-posed to fully automating occupations.The potential employment effects,whether augmenting or autom

18、ating,vary widely across coun-try income groups,due to different occupational structures.In low-income countries,only 0.4 per cent of total employment is potentially exposed to automation effects,whereas in high-income countries the share rises to 5.5 percent.The effects are highly gendered,with mor

19、e than double the share of women potentially affected by automation.The greater impact is from augmenta-tion,which has the potential to affect 10.4 percent of employment in low-income countries and 13.4 percent of employment in high-income countries.However,such effects do not consider infrastructur

20、e constraints,which will impede the possibility for use in lower-income countries and likely increase the productivity gap.We stress that the primary value of this analysis is not the precise estimates,but rather the in-sights that the overall distribution of such scores provides about the nature of

21、 possible changes.Such insights can encourage governments and social partners to proactively design policies that support orderly,fair,and consultative transitions,rather than dealing with change in a reactive manner.Moreover,the likely ramifications on job quality might be of greater consequence th

22、an the quantitative impacts,both with respect to the new jobs created because of the technology,but also the potential effects on work intensity and autonomy when the technology is integrat-ed into the workplace.For this reason,we also emphasize the need for social dialogue and reg-ulation to suppor

23、t quality employment.About the authorsPawe Gmyrek is Senior Researcher in the Research Department of the ILO.Janine Berg is Senior Economist in the Research Department of the ILO.David Bescond is Data Scientist in the ILOs Department of Statistics.02ILO Working Paper 96Abstract 01About the authors 0

24、1Acronyms 05 XIntroduction 07 X 1 Methods and Data 101.1.ISCO data on occupations and tasks 111.2.Prompt design and sequence 12 X 2 Assessment of the Predictions,Robustness Tests and the Bounds for Analysis 17 X 3 Results 203.1.Automation vs augmentation:distribution of scores across tasks and occup

25、ations 24 X 4 Exposed occupations as a share of employment:global and income-based estimates 304.1.Augmentation vs Automation:ILO microdata 304.2.Augmentation vs Automation:global estimate 324.3.The big unknown 36 X 5 Managing the transition:Policies to address automation,augmentation and the growin

26、g digital divide 385.1 Mitigating the negative effects of automation 385.2 Ensuring job quality under augmentation 395.3 Addressing the digital divide 40 XConclusion 43Appendix 1.Countries with missing ISCO-08 4-digit data:estimation procedure 45References 47Acknowledgements and use of GPT 51Table o

27、f contents03ILO Working Paper 96List of FiguresFigure 1.Mean automation scores by occupation,based on ISCO and GPT tasks 21Figure 2.Tasks with medium and high GPT-exposure,by occupational category(ISCO 1-digit)24Figure 3.Box plot of task-level scores by ISCO 4d,grouped by ISCO 1d 25Figure 4.Augmenta

28、tion vs automation potential at occupational level 27Figure 5.Occupations with high automation potential 28Figure 6.Occupations with high augmentation potential 29Figure 7a.Automation vs augmentation potential:shares of total employment,microdata for 59 countries 30Figure 7b.Automation vs augmentati

29、on potential:shares of total employment in each sex(ILO microdata)31Figure 8.Country coverage based on the level of digits in ISCO-08(ILO data)33Figure 9a.Global estimates:jobs with augmentation and automation potential as share of total employment 34Figure 9b.Automation vs augmentation potential:sh

30、ares of total employment for each sex(global estimate)35Figure 10.Occupations with high automation potential,by ISCO 4-digit and income group 36Figure 11a.The“Big Unknown”:occupations between augmentation and automation potential 37Figure 11b.The“Big Unknown”:share of total employment,by income grou

31、p(global estimate)37Figure 11.Share of population not using the internet 41Figure 12.A classic growth path:income and occupational diversification 4204ILO Working Paper 96List of Tables 11 14 15 17 22 26Table 1.ISCO-08 Structure of occupations and tasks used in the study Table 2.Sample of tasks and

32、definitions from ISCO and predicted by GPT-4 Table 3.Sample of task-level scores(high-income country context)Table 4.a Test of score consistency(100 task-level predictions)Table 4.b Tasks with high automation potential clustered into thematic groups*Table 5.Grouping of occupations based on task-leve

33、l scores Table 6.Microdata coverage by levels ISCO-08:number of countries 3205 ILO Working Paper 96Acronyms3GThird Generation(referring to a generation of standards for mobile telecom-munications)AdaA language model by OpenAI used to generate embeddingsAGIArtificial General IntelligenceAIArtificial

34、IntelligenceANNArtificial Neural NetworkAPIApplication Programming InterfaceATMsAutomated Teller MachinesCPUCentral Processing UnitDLDeep LearningDOLEDepartment of Labor and Employment ESCOEuropean Skills,Competences,Qualifications and OccupationsGPTsGenerative Pre-Trained TransformersGPT-4Generativ

35、e Pre-Trained Transformer 4GPUGraphics Processing UnitHICHigh-Income CountriesICTInformation and Communications TechnologyILOInternational Labour OrganizationISCOInternational Standard Classification of OccupationsISCO-08International Standard Classification of Occupations 2008K-MeansK-Means Cluster

36、ing AlgorithmLFSLabour Force SurveysLICLow-Income CountriesLLMsLarge Language Models06 ILO Working Paper 96LMICLower-Middle-Income CountriesMLMachine LearningNLPNatural Language ProcessingOECDOrganisation for Economic Co-operation and DevelopmentO*NETOccupational Information NetworkOpenAIOpen Artifi

37、cial Intelligence(organizations name)PythonHigh-level programming languageRLReinforcement LearningSDStandard DeviationSMEsSmall and Medium-sized EnterprisesUMICUpper-Middle-Income CountriesUSUnited StatesUSDUnited States DollarUMICUpper-Middle-Income CountriesUSUnited States07 ILO Working Paper 96 X

38、IntroductionEach new wave of technological progress intensifies debates on automation and jobs.Current debates on Artificial Intelligence(AI)and jobs recall those of the early 1900s with the introduc-tion of the moving assembly line,or even those of the 1950s and 1960s,which followed the intro-ducti

39、on of the early mainframe computers.While there have been some nods to the alienation that technology can bring by standardizing and controlling work processes,in most cases,the debates have centred on two opposing viewpoints:the optimists,who view new technology as the means to relieve workers from

40、 the most arduous tasks,and the pessimists,who raise alarm about the imminent threat to jobs and the risk of mass unemployment.What has changed in debates on technology and workers,however,is the types of workers af-fected.While the advances in technology in the early,mid and even late-1900s were pr

41、imarily focused on manual workers,technological development since the 2010s,in particular the rapid progress of Machine Learning(ML),has centred on the ability of computers to perform non-rou-tine,cognitive tasks,and by consequence potentially affect white-collar or knowledge workers.In addition,the

42、se technological advancements have occurred in the context of much strong-er interconnectedness of economies across the globe,leading to a potentially larger exposure than location-based,factory-level applications.Yet despite these developments,to an average worker,even in the most highly developed

43、countries,the potential implications of AI have,until recently,remained largely abstract.The launch of ChatGPT marked an important advance in the publics exposure to AI tools.In this new wave of technological transformation,machine learning models have started to leave the labs and begin interacting

44、 with the public,demonstrating their strengths and weaknesses in daily use.The chat function dramatically shortened the distance between AI and the end user,simultaneously providing a platform for a wide range of custom-made applications and inno-vations.Given these significant advancements,it is no

45、t surprising that concerns over potential job loss have resurged.While it is impossible to predict how generative AI will further develop,the current capabilities and future potential of this technology are central to discussions of its impact on jobs.Sceptics tend to believe that these machines are

46、 nothing more than“stochastic parrots”powerful text summarizers,incapable of“learning”and producing original content,with little future for gen-eral purpose use and unsustainable computing costs(Bender et al.2021).On the other hand,more recent technical literature focused on testing the limits of th

47、e latest models suggests an increasing capability to carry out“novel and difficult tasks that span mathematics,coding,vision,medicine,law,psychology and more”,and a general ability to produce responses exhibiting some forms of early“reasoning”(Bubeck et al.2023).Some assessments go as far as suggest

48、ing that machine learning models,especially those based on large neural networks used by Generative Pre-trained Transformers(GPT,see Text Box 1),might have the potential to eventually become a general-purpose technology(Goldfarb,Taska,and Teodoridis 2023;Eloundou et al.2023).1 This would have multip

49、lier effects on the economy and labour markets,as new products and servic-es would likely spring from this technological platform.As social scientists,we are not in position to take sides in these technical debates.Instead,we focus on the already demonstrated capabilities of GPT-4,including custom-m

50、ade chatbots with retrieval of private content(such as collections documents,e-mails and other material),natu-ral language processing functions of content extraction,preparation of summaries,automated content generation,semantic text searches and broader semantic analysis based on text em-beddings.L

51、arge Language Models(LLMs)can also be combined with other ML models,such as 1The three main characteristics of general-purpose technologies are pervasiveness,ability to continue improving over time,and abil-ity to spawn further innovation(Jovanovic and Rousseau,2005).08 ILO Working Paper 96speech-to

52、-text and text-to-speech generation,potentially expanding their interaction with dif-ferent types of human tasks.Finally,the potential of interacting with live web content through custom agents and plugins,as well as the multimodal(not exclusive to text,but also capable of reading and generating ima

53、ge)character of GPT-4 makes it likely that this type of technology will expand into new areas,thereby increasing its impact on labour.Departing from these observations,this study seeks to add the global perspective to the already lively debate on possible changes that may result in the labour market

54、s as a consequence of the recent advent of generative AI.We stress the focus of our work on the concepts of“exposure”and“potential”,which does not imply automation,but rather lists occupations and associated employment figures for jobs that are more likely to be affected by GPT-4 and similar technol

55、ogies in the coming years.The objective of this exercise is not to derive headline figures,but rather to analyse the direction of possible changes in order to facilitate the design of appropriate policy responses,including the possible consequences on job quality.The analysis is based on 4-digit occ

56、upational classifications and their corresponding tasks in the ISCO-08 standard.It uses the GPT-4 model to estimate occupational and task-level scores of ex-posure to GPT technology and subsequently links these scores to official ILO statistics to derive global employment estimates.We also apply emb

57、edding-based text analysis and semantic clus-tering algorithms to provide a better understanding of the types of tasks that have a high auto-mation potential and discuss how the automating and augmenting effects will strongly depend on a range of additional factors and specific country context.We di

58、scuss the results of this analysis in the broader context of labour market transformations.We put particular focus on the current disparities in digital access across countries of different income levels,the potential for this new wave of technological transformation to aggravate such disparities,an

59、d the ensuing consequences on productivity and income.We also give consider-ation to jobs with highest automation and augmentation potential and discuss gender-specific differences.The analysis does not take into account the new jobs that will be created to accom-pany the technological advancement.T

60、wenty years ago,there were no social media managers,thirty years ago there were few web designers,and no amount of data modelling would have rendered a priori predictions concerning a vast array of other occupations that have emerged in the past decades.As demonstrated by Autor et al.(2022),some 60

61、per cent of employment in 2018 in the United States was in jobs that did not exist in the 1940s.Indeed,the main value of studies such as this one is not in the precise estimates,but rather in understanding the possible direction of change.Such insights are necessary for proactively de-signing polici

62、es that can support orderly,fair,and consultative transitions,rather than dealing with change in a reactive manner.For this reason,we also emphasize the potential effects of technological change on working conditions and job quality and the need for workplace consul-tation and regulation to support

63、the creation of quality employment and to manage transitions in the labour market.We hope that this research will contribute to needed policy debates on digital transformation in the world of work.While the analysis outlines potential implications for different occupational categories,the outcomes o

64、f the technological transition are not pre-determined.It is humans that are behind the decision to incorporate such technologies and it is humans that need to guide the transition process.It is our hope that this information can support the development of policies needed to manage these changes for

65、the benefit of current and future societies.We intend to use this broad global study as an opening to more in-depth analyses at country level,with a particular focus on developing countries.09ILO Working Paper 96 X Text Box 1:What are GPTs?Generative Pre-Trained Transformers belong to the family of

66、Large Language Models a type of Machine Learning mod-el based on neural networks.The“generative”part refers to their ability to produce output of a creative nature,which in language models can take the form of sentences,paragraphs,or entire text structures,with characteristics often un-distinguishab

67、le from that produced by humans.“Pre-trained”refers to the initial training on a large corpus of text data,typically through unsupervised or self-supervised learning,during which the model learns about the text structure by temporarily masking part of the content and trying to minimize errors in the

68、 prediction of the masked words.Following pre-training,such models are further fine-tuned with the use of labelled data and so-called“reinforcement learning”,making them more suitable for specific tasks.This part of training is often perceived as a specialized job,executed by a handful of technical

69、experts.In reality,it is labour intensive and involves many invisible contributors(Dzieza 2023).Its prerequisite is the production of vast amounts of labelled data,typically done by workers on crowdsourcing platforms.“Transformers”refer to the underlying model architecture,which uses numerous mechan

70、isms,such as attention and self-attention frameworks,to develop weights related to the importance of text elements,such as words in a sentence,which are subsequently used for predictions(Vaswani et al.2017).While GPT specifically refers to models developed by OpenAI(GPT-1,2,3 and 4),this type of arc

71、hitecture is used by many more language models already available commercially.The launch of ChatGPT on 30 November 2022 made GPTs more popular among the public,as it made it possible for individuals with no programming knowledge to interact with GPT-3(and eventually GPT-4)through a chatbot function

72、with a human-like tone.For research purposes and more com-plex applications,such language models are typically more powerful when used through an Application Programming Interface(API).An API is a developer access point that relies on a query-response protocol with the use of programming software.In

73、 our case,we rely on a Python script based on OpenAI library,designed to connect to GPT-4 model,provide a fine-tuned prompt and receive a response,which is subsequently stored in a database on our server.This enables bulk processing of large numbers of requests and relies on the GPT-4 model with mor

74、e parameters than what is accessible through the public Chat function.10 ILO Working Paper 96 X1 Methods and Data There are two principal approaches to the analysis of automation of occupations(Georgieff and Hyee 2021).The first is to use data on job vacancies to understand how demand for specific s

75、kills evolves over time.Most studies using this approach harness data from online recruitment plat-forms(Cammeraat and Squicciarini 2021;Acemoglu et al.2022)to measure the frequency of ref-erences to AI(or to any other technology of interest)in the text of the job description.These ref-erences are t

76、hen used as a proxy for the demand for specific skills and,by its extension,a proxy for the rate of technological adoption at the enterprise level.This approach works well in coun-tries with a high online presence in recruitment,though it does not always capture the indus-tries affected as a result

77、of subcontracting.The approach,however,is less well suited for a global study covering countries with less online presence,as most vacancies are not advertised on on-line platforms but recruited through other means of communication(Georgieff and Hyee 2021).The second approach is to focus on occupati

78、onal structures,with the idea of estimating the au-tomation potential of tasks or skills that make up a given job.The advantage of this method is that such occupational classifications can easily be linked to official labour market statistics,which is of particular importance for understanding globa

79、l,regional and income-based differ-entials.This strand of literature is rich,but frequently misunderstood,especially when it comes to communicating its findings to the public,as media interpretations tend to blur the distinc-tion between automation potential and actual deployment in the workplace.Fo

80、r example,Frey and Osbornes(2013,2017)influential study has been cited over 12,000 times,often for differ-ent types of doomsday pronouncements,even though the authors were clear about the distinc-tion between potential and predicted effects.A range of studies follow this research tradition,attemptin

81、g to calculate different types of occupational automation scores in OECD countries(Brynjolfsson,Mitchell,and Rock 2018;Felten,Raj,and Seamans 2018;Felten,Raj,and Seamans 2019;Acemoglu and Restrepo 2020;Fossen and Sorgner 2022)or even combining occupational and job posting data(Georgieff and Hyee 202

82、1).Some authors have also taken up the challenge of producing better estimates for developing countries(Balliester and Elsheikhi 2018),often by trying to link detailed occupational data and automation scores from the US with less structured datasets available for lower-income countries(Aly 2020;Carb

83、onero et al.2023).Calculating occupational scores typically involves development of a rubric,which defines a scor-ing method based on pre-established criteria to capture possible impacts from the technology of interest.The rubric is then applied to occupations or occupational tasks,to generate task-

84、or occupation-specific scores.One of the challenges of this approach emerges in covering a wide range of technologies.While some tasks could be very well suited for automation with a particu-lar type of AI(for example,routine non-cognitive tasks in a factory setting),the same technology could be com

85、pletely useless in other areas that require cognitive abilities.Attempting to cover the wide range of systems that currently fall into the AI category would require squeezing the assessments into one matrix of overall technological capabilities.In this study,we focus exclusively on LLMs with similar

86、 capabilities as the latest GPT models.We build upon the method recently demonstrated by Eloundou et al.(2023),and replicated by Eisfeldt et al.(2023),which relies on the use of sequential API calls to GPT-4 model for the purpose of es-timating task and occupational-level automation scores concernin

87、g this particular technology.We observe that their study demonstrates an astonishing proximity of GPT-4 predictions to the judgements made by a group of AI experts(albeit with a hard-to-determine level of possible bias on the human side).Applying a similar approach to the International Standard Clas

88、sification of Occupations(ISCO-08),we conduct some 25,000 high frequency API calls,fine-tuned at the level of occupational definitions,job titles,tasks and country income classifications.We combine the resulting score matrix with what has long been the comparative advantage of the International 11 I

89、LO Working Paper 96Labour Organization(ILO):the ability to translate expert knowledge about occupations into glob-al,regional and country-level employment estimates.1.1.ISCO data on occupations and tasksThe current ISCO-08 relies on a hierarchical structure,reflected in a system of digits.The highes

90、t 1-digit level covers 10 different types of occupational groups that can be further broken down into lower-level sub-groups,each time represented by an increasing number of digits.The most detailed,4-digit level captures 436 occupations(See Table 1).While the publicly available ILO statistics are a

91、t the 2-digit ISCO-08 level,the ILO holds a wealth of additional information from labour force surveys(LFS)and other national surveys in the ILO Harmonized Microdata collection.Its statistical repository contains microdata on employment at the 4-digit ISCO level for some 73 countries,and 3-digit emp

92、loyment data for over 117 countries.This gives us access to a sizeable repository of harmonized survey data that can be used to ana-lyse labour market information in a wide range of countries,including the detailed distributions of employment across occupations.The internal processing of LFS data al

93、so captures additional parameters of interest,such as variations in job titles that belong to each ISCO 4-digit category across different countries.As of 2023,there are some 7,500 jobs titles mapped to ISCO at 4-dig-its,which we also use as a robustness test for our analysis(see Section 3).X Table 1

94、.ISCO-08 Structure of occupations and tasks used in the studyISCO-08 1-digit codeISCO-08 1-digit full labelNr of distinct 1-digit codesNr of distinct 2-digit codesNr of distinct 3-digit codes Nr of distinct 4-digit codesTotal ISCO tasksTotal GPT tasks0Armed forces occupations13330301Managers14113123

95、63102Professionals1627927519203Technicians and associate pro-fessionals1520845808404Clerical support workers148291632905Service and sales workers1413402694006Skilled agricultural,forestry and fishery workers139181411807Craft and related trades work-ers1514665036608Plant and machine operators,and ass

96、emblers1314402804009Elementary occupations161133200330Total10431304363,1234,360To build the principal data frame of tasks and occupations,we use as our foundation Part III of the official ISCO-08 documentation,which provides detailed definition and description of tasks for each of the 436 ISCO-08 4-

97、digit occupations(ILO 2023b).These tasks are devised with a global perspective and used to describe similar occupations that can be identified in LFS,other household surveys and censuses,and other non-statistical sources,such as data derived from 12 ILO Working Paper 96administrative records.This me

98、ans that they are designed to provide a common denominator for the variability of tasks under a given occupation.The number of ISCO-08 tasks assigned to each occupation can vary anywhere from 4 to 14.The data frame with a full list of ISCO-08 occu-pations and tasks constitutes the starting point of

99、our estimations.1.2.Prompt design and sequenceWe develop a Python script that uses the OpenAI library to loop over the ISCO-08 task structure and conduct a series of sequential API calls to the GPT-4 model,using a range of prompts that we fine-tune for specific queries.Before predicting task-level s

100、cores,we run several initial tests of the GPT-4 model on the overall ISCO dataset,to determine its capacity for processing detailed occupational information.As a first step,we use the GPT-4 model to generate an international definition for each of the ISCO 4-digit codes,and to mark the level of skil

101、ls required for each job,according to the same classification as used in ISCO-08(1 for low level skills,4 for the highest).We design the first GPT-4 API prompt,as follows:2“role”:“system”,“content”:“You are a skills specialist3.You will provide job definitions based on a job title and ISCO code.Foll

102、ow instructions closely.”,“role”:“user”,“content”:“Look at this ISCO code and job title and provide an international stand-ard definition of this job:”+“Do not provide any other content,just the definition of some 100 words that describes what the job is about and which level of ISCO skills it requi

103、res(1-4).”+“ISCO code:”+str(ISCO_08)+“Job Title:”+str(Title)By comparing the result with official ISCO-08 definitions,we examine the models“understanding”of the ISCO-08 structure.We observe that the generated definitions are largely consistent with ISCO-08 and often contain more detailed information

104、,which could potentially be a helpful feature in complementing some of the definitions so far created by humans specialized in this domain.As the next step,we move our tests to the level of tasks.It is likely that the training data of GPT-4 included publicly available information from the O*NET occu

105、pations and their corresponding tasks,as well as the European Skills,Competencies and Occupations(ESCO)and ISCO occupa-tional classifications at the 4-digit level,as the model demonstrates familiarity with the details of these different systems.Yet beyond simply reciting the content of these databas

106、es,GPT-4 seems able to engage in more complex exchanges and develop logical links between different types of occupational classifications and tasks a surprising and useful ability that has been document-ed in other domains of application(Bubeck et al.2023).4We therefore adjust the prompt and request

107、 GPT-4 to generate a set of 10 typical tasks for each of the 436 ISCO-08 4-digit occupations,which we append to the main data frame alongside the official ISCO-08 tasks and definitions.Generating a uniform set of tasks across all occupations provides some analytical benefits.First,considering that G

108、PT-4 has detailed ISCO-08 informa-tion already in its training data,the ten-task requirement helps to avoid a situation where the responses simply mirror what GPT-4 already knows about ISCO-08,but rather pushes the model to provide its“own”perception of tasks that belong to each occupation.Second,in

109、 the next step of prediction,a uniform set of tasks allows us to compute mean occupational automation scores 2The prompt in the example is used as part of ChatCompletion.create()function in OpenAI Python library.3We use a skills and AI specialist for the system role,which is understood by GPT-4 as“s

110、omeone who has in-depth knowledge of artificial intelligence technologies,such as machine learning,deep learning,and AI architectures like GPT.They understand how AI systems work,what they are capable of,and the limitations they have.Moreover,they can assess the skills required for certain jobs or t

111、asks and evaluate the potential for these tasks to be automated by AI.They could provide insights into the extent to which AI might be able to replace or augment human roles in various fields,as well as advise on how people might need to adapt or acquire new skills to remain competitive in an increa

112、singly automated job market.”4We observe that some functionalities of the model could potentially be used to support specialized discussions on updating the standard classifications.For example,semantic clustering and content summarization could be used to extract core tasks from a multitude of curr

113、ently unprocessed country-level inputs,which is a tedious and overwhelming task for humans.13 ILO Working Paper 96based on a more statistically balanced list across all ISCO-08 occupations and compare the result to scores based on a varied number of tasks per occupation in the official ISCO-08 docum

114、entation.Table 2 shows a detailed comparison of ISCO-08 tasks with the tasks derived by GPT-4,using the example of primary school teachers(ISCO 2341).As can be seen,the definition and tasks pre-dicted by GPT-4 are comparable to the expert description provided by the ILO.Some of the GPT-generated tas

115、ks are more generic than those given in ISCO,which could partially be a result of our API request that requires 10 tasks,as opposed to the varied task numbers in the ISCO doc-umentation.5 We suspect that the model,by being prompted to provide a list of ten tasks,might be deriving some additional tas

116、ks in order to complete the list,or it may be breaking down more complex tasks into simpler ones in order to match the required parameters of response.5The full list of ISCO 4-digit descriptions by ISCO and GPT is available from the authors.14 ILO Working Paper 96 X Table 2.Sample of tasks and defin

117、itions from ISCO and predicted by GPT-4ISCO-08Definition ISCODefinition predicted by GPT42341Primary school teachers teach a range of sub-jects at the primary education level.Primary School Teachers(ISCO code:2341)are education professionals responsible for planning,preparing,and delivering lessons

118、and activities to facilitate the cognitive,social,and emotional development of children aged be-tween 5 and 12 years old.They assess students progress,provide a supportive and inclusive learning environment,and collaborate with parents and other education pro-fessionals to ensure the overall well-be

119、ing and success of students.Primary School Teachers require ISCO skill lev-el 4,which includes a bachelors degree in education or a related field,and they must possess pedagogical skills,communication skills,and creativity in teaching method-ologies.ISCO-08Tasks ISCOTasks predicted by GPT42341Prepar

120、ing daily and longer-term lesson plans in accordance with curriculum guidelines;Develop lesson plans and daily schedules for a variety of subjects at the primary education level;Instructing children individually and in groups,using various teaching methods and materials(puters,books,games),adapting

121、to childrens varying needs;Teach basic skills such as reading,writing,and arithmetic to students using various teaching methods and tools;Maintaining discipline and good working hab-its in the classroom;Establish and maintain classroom rules to create a positive and structured learning environment;P

122、lanning and conduct activities with the chil-dren such as sporting activities,concerts and excursions;Collaborate with other teachers and school staff to organ-ize events,activities,and educational programs;Assigning and grading class work and home-work;Provide individualized attention and support t

123、o students who are struggling or require additional help;Preparing,administering,and grading tests and assignments to evaluate childrens pro-gress;Assess students performance and progress through as-signments,quizzes,and standardized tests;Observing and evaluating childrens perfor-mance and behaviou

124、r;Prepare students for the transition to higher education levels by helping them develop necessary study habits and organizational skills;Supervising children during classes and at oth-er times in the school day,including the play-ground during breaks;Adapt teaching methods and instructional materia

125、ls to meet students varying needs and interests;Participating in staff meetings and other ses-sions,and conferring with other teachers con-cerning educational issues;Participate in professional development sessions and workshops to stay updated with new teaching techniques and educational reforms;Pr

126、eparing for and attending parent meetings to discuss childrens progress and problems.Communicate regularly with parents and guardians to dis-cuss students progress and address any concerns.15 ILO Working Paper 96 X Table 3.Sample of task-level scores(high-income country context)ISCO-08ISCO TasksJust

127、ification by GPTGPT-predicted TasksJustification by GPT2341Preparing daily and longer-term lesson plans in accordance with cur-riculum guidelines;0.60GPT technology can aid in the gen-eration of lesson plans and offer suggestions based on curriculum guidelines,but a human teachers expertise and nuan

128、ced understand-ing of their students needs will still be important for crafting effective plans.Develop lesson plans and daily schedules for a variety of subjects at the primary ed-ucation level0.70GPT technology can help in generating content and providing suggestions for lesson plans,but human gui

129、dance is still required for contextual understanding and tailoring the lessons to suit the spe-cific needs of the students at the prima-ry education level.2341Instructing children in-dividually and in groups,using various teaching methods and materials(puters,books,games),adapting to chil-drens vary

130、ing needs;0.30GPT technology can assist in provid-ing instructional materials and adap-tive learning approaches,but the physical presence,emotional con-nection,and real-time adaptability of a human teacher are essential for effectively teaching young children.Teach basic skills such as reading,writi

131、ng,and arith-metic to students using var-ious teaching methods and tools0.30GPT technology can assist in teaching basic skills by providing content and ex-ercises,but it cannot fully replace a hu-man teacher needed for personalized guidance,classroom management,and social-emotional development.2341M

132、aintaining discipline and good working habits in the classroom;0.15GPT technology can assist in moni-toring and providing feedback,but it cannot fully automate maintaining discipline and good working habits in the classroom because human in-teraction and physical presence are essential for effective

133、 discipline and enforcing rules.Establish and maintain class-room rules to create a posi-tive and structured learning environment0.20Establishing and maintaining classroom rules involves understanding the unique social dynamics of a specific group of students,which GPT technology may struggle to ass

134、ess comprehensively and adapt to.2341Planning and conduct activities with the chil-dren such as sporting activities,concerts and excursions;0.25GPT technology can contribute to idea generation and planning for ac-tivities,but it cannot physically con-duct activities or interact with chil-dren effect

135、ively in real-life situations.Collaborate with other teach-ers and school staff to organ-ize events,activities,and edu-cational programs.0.55GPT technology can aid in planning,communication,and organization,but human interaction and collaboration with other staff members is still essen-tial to succe

136、ssfully implement events and programs.2341Participating in staff meetings and other ses-sions,and conferring with other teachers con-cerning educational is-sues;0.15GPT technology can potentially as-sist in identifying meeting agendas,summarizing discussion points,and providing insights on issues,bu

137、t it cannot replace human interaction and collaboration required in staff meetings and conferring with other teachers.Participate in professional development sessions and workshops to stay updat-ed with new teaching tech-niques and educational re-forms0.30GPT technology can partially provide informa

138、tion and resources for profes-sional development,but human en-gagement and interaction are essential for proper learning and understanding of new teaching techniques and educa-tional reforms.16ILO Working Paper 96As the final step in the data generation process,we run another set of sequential API c

139、alls at the level of individual tasks.We request GPT-4 to generate a score between 0 and 1,representing potential automation with GPT-based technology for each task in the ISCO task collection and in the GPT-generated set of tasks.We provide the occupations ISCO 4-digit code,specify whether the job

140、is located in a high-income or a low-income country and ask the model to justify its deci-sion.After several rounds of fine-tuning,we settled on the following prompt:“role”:“system”,“content”:“You are a skills and AI specialist.”+“You will provide a score of po-tential automation with GPT technology

141、 for a given task.Follow instructions closely.”,“role”:“user”,“content”:“Look at this job task:”+str(Tasks_GPT)+“It is related to ISCO code:”+str(ISCO_08)+“Provide a score of potential automation of this task with GPT technology,given that the job is located in a highlow income country:”+“The score

142、should range 0-1.Provide a score in one line,and a justification in next line.Do not provide any other commentary,only the score and justification.”+“Do not give any ranges just one score for each task.”This exercise results in an ISCO-08 4-digit level data frame,with automation scores predicted for

143、 each ISCO-08 tasks and for GPT-predicted tasks,with separate scores for low-and high-income countries.Each of the task-level scores is accompanied with a short justification generated by GPT-4.Table 3 shows the results for primary school teachers(ISCO-08 2341)in a high-income country.17ILO Working

144、Paper 96 X2 Assessment of the Predictions,Robustness Tests and the Bounds for AnalysisWe approach our predicted task-level scores with scepticism.However,following a manual re-view,at a large scale of 3,123 tasks across all ISCO-08 occupations,we find no evidence of bias in one direction:highly auto

145、matable tasks such as typing consistently get a high score(above 0.7),whereas tasks requiring manual dexterity consistently get low scores.Moreover,GPT-4 pro-vides a reasonable written explanation of differences across the scores attributed to similar cat-egories(Table 3).We conduct an additional te

146、st of scoring consistency across tasks(whether the model predicts similar level of scores for different types of tasks across multiple runs,based on the same input)and score variability at task level(the range of scores predicted for the same task across multiple runs,based on the same input)by maki

147、ng 100 predictions for 5 tasks randomly selected from all tasks on ISCO-08 list.We then calculate the mean score and standard deviation(SD)for each of the tasks,as shown in Table 4.The scores are highly consistent across different types of tasks,with SDs not exceeding 0.05.This is likely because the

148、 random element in scoring is lower than what it would be in the case of scoring by human respondents,who typically struggle with score uncertainty(e.g.whether a score of 0.2 would be more adequate than 0.15 or 0.25)and tend to have greater variability of opinions.X Table 4.a Test of score consisten

149、cy(100 task-level predictions)ISCO_08TaskMean SD5141Cutting,washing,tinting and waving hair;0.06 0.038122Operating and monitoring equipment which cleans metal articles in preparation for electroplating,galvanizing,enamelling or similar processes;0.11 0.042264Recording information on patients health

150、status and responses to treatment in medical records-keeping systems,and sharing information with other health pro-fessionals as required to ensure continuing and comprehensive care;0.64 0.053313Verifying accuracy of documents and records relating to payments,receipts and other financial transaction

151、s;0.73 0.054411Maintaining library records relating to the acquisition,issue and return of books and other materials.0.73 0.05As a parallel robustness test,we use a slightly modified prompt to generate occupational-level scores for over 7,500 job titles that can be found in different national labour

152、 force surveys,and which aggregate to the 436 ISCO 4-digit occupations.These jobs do not have detailed tasks,but a comparison of occupation level scores with the mean occupational scores generated based on detailed tasks reveals a proximity across the board.In other words,whether we rely on individu

153、al tasks that aggregate to occupations or a much larger pool of job titles to generate predictions,GPT-4 is consistent in the way it scores automation potential.This obviously has to do with its training data,both in terms of originally ingested textual sources and further human-based fine-tuning of

154、 the model.Given the similarity of GPT-4 scores with hu-man-based scoring by AI experts on task-level questions,demonstrated in Eloundou et al.(2023),we believe that our exercise is likely to be estimating the upper bound of the exposure to GPT.This is explained by multiple reasons.18ILO Working Pap

155、er 96First,as recently shown by Karger et al.(2023),tech experts tend to overstate technological ca-pacities and risks in questions concerning broader applications.We believe this is also likely to be true when it comes to full-scale deployment of GPT technology at the workplace,in particular at a l

156、evel that would allow for full elimination of the human component.This is well illustrated in earlier automation studies,which often assigned high scores of displacement potential to rou-tine tasks and even entire occupations,including in garment production.In practice,however,the work continues to

157、be performed by humans due to the challenges of handling highly pliable fabrics and the complexity of skills and dexterity involved in the stitching process(de Mattos et al.2020).Because of GPTs tech-oriented training data and trainers profiles,as well as the litera-ture on automation that most cert

158、ainly was part of its training data,GPT is likely to reflect tech-no-optimism and overstate some task-level scores.GPT-generated scores also do not account for job-level task variation,which can lower occupational-level scores(Arntz et al.,2017).Second,our prompts to GPT focus on technical feasibili

159、ty and ignore important determinants of techno-logical diffusion,including the feasibility of adoption in a given environment which is dependent on constraints such as access to electricity and internet in countries with lower income,or local market dynamics,such as relative cost of labour to techno

160、logy,level of digital literacy or access to finance.Third,despite having generated score predictions from prompts specifying whether the job is in a low-or a high-income country,we find that the difference between the two is too small to justify the use of both datasets.Therefore,for purposes of thi

161、s initial analysis,we use the high-income country scores,with the understanding that the this further contributes to es-timating the upper bound of global exposure,since technological deployment faces additional barriers in lower-income countries.Nevertheless,this theoretical approximation facilitat

162、es an initial global picture of the potential impact on occupations across the globe,for which more detailed and contextualized studies will be needed.Since the tasks in ISCO-08 documentation and those generated by GPT-4 do not correspond di-rectly,we cannot compare the values of automation scores a

163、t the individual task level in the two data sets.Instead,we focus on the occupation level and examine the similarity of the occupa-tional scores,calculated as an arithmetic mean of the task-level scores for each ISCO-08 4-digit occupation.We find that,in general,scores based on tasks previously gene

164、rated by GPT tend to be higher than those attributed to tasks coming directly from ISCO-08.We attribute this differ-ential to the more refined character of ISCO-08 tasks,as opposed to the some of the more ge-neric tasks generated by GPT.In other words,confronted with a higher complexity of tasks cap

165、-tured in the ISCO-08 documentation,GPT-4 seems to attribute lower automation scores,when compared to its own collection of tasks,for which it tends to be more generous with automa-tion potential.We treat the scores related to ISCO-08 tasks as the basis for further analysis,since they are directly l

166、inked to an international standard and associated ILO employment statistics.Since ISCO-08 documentation does not provide any tasks for the first major group of“Armed Forces Occupations”,we use GPT-predicted tasks and scores to include this category in further analysis.In addition,ISCO-08 does not pr

167、ovide tasks for occupations with codes 1439(Services Managers Not Elsewhere Classified),3139(Process Control Technicians Not Elsewhere Classified),3435(Other Artistic and Cultural Associate Professionals),5249(Sales Workers Not Elsewhere Classified),7319(Handicraft Workers Not Elsewhere Classified)a

168、nd 8189(Stationary Plant and Machine Operators Not Elsewhere Classified),which also explains the missing points on ISCO-08 tasks in Figure 1 in the following section.As the catch-all character of these few occupations does not permit the assignation of specific tasks,we drop them from the final anal

169、ysis.Finally,a classic challenge in analysing occupational tasks concerns attributing the share of time needed to execute the individual tasks in a given occupation(Carbonero et al.2023).Time distri-bution likely varies in different country contexts,but unfortunately,the labour force and other surve

170、y data do not provide enough information to make country-level distinctions.The problem of attributing time weights across task-level scores is not exclusive to our attempts and typically appears in the construction of composite indicators related to technology and occupations(e.g.Autor and Dorn 201

171、3;Brynjolfsson,Mitchell,and Rock 2018).One of the reasons why many stud-ies on automation focus on the USA is that the level of detail in the O*NET data facilitates such 19ILO Working Paper 96estimations.For our case,we opt for the most straightforward solution especially given the glob-al focus,whi

172、ch is to apply equal weights to each task-level sub-component or each occupation.20ILO Working Paper 96 X3 ResultsTo further address any potential score imprecision,we establish generous margins for classifica-tions in the calculations that follow,focusing on the extremes of the scoring scale,and in

173、terpret most results at a higher level of aggregate ISCO-08 1-digit categories.Given the range of the estimated index(0-1),we consider scores below 0.25 as representing very low exposure and those between 0.25 and 0.5 as low exposure.Medium exposure is captured in scores with the range of 0.5-0.75,w

174、hile tasks with scores above 0.75 are considered as high-ly exposed.The same cut-off points are applied to the occupation-level scores,calculated as a mean score of the tasks that belong to each occupation.Figure 1 presents the breakdown,with the two upper limits of exposure marked with horizon-tal

175、lines:0.5 for medium exposure and 0.75 for high exposure.The grey area between the dot-ted lines represents the distance between the scores for each occupation based on ISCO-08 and GPT-predicted tasks.This illustration reveals a consistency among the scoring based on ISCO-08 and GPT-generated tasks,

176、with highest exposure found amongst clerical support workers,fol-lowed by technicians and associate professionals,and by professionals.While these occupations have no official common category,they are broadly associated with“knowledge work”(Berg and Gmyrek 2023).21ILO Working Paper 96 X Figure 1.Mea

177、n automation scores by occupation,based on ISCO and GPT tasks22ILO Working Paper 96In addition,the broad category of managers,which for the most part falls underneath the 0.5 cut-off,nonetheless approximates the line of medium exposure.The results for service and sales workers are more mixed,with so

178、me occupations surpassing the threshold but most others falling below.Plant and machine operators and assemblers,elementary occupations,craft and related trades workers and skilled agricultural,forestry and fishery workers have more limited exposure.What drives these results?To answer this question,

179、we apply machine learning techniques to the analysis of ISCO-08 tasks that have been classified as having a high level of exposure.First,we group and sort all tasks with the highest exposure scores and use the OpenAI Ada model to assign embeddings for each task through sequential task-level API call

180、s.6 We then perform semantic clustering of the tasks,based on the K-Means algorithm and a visual inspection of re-sults,which suggests five principal thematic clusters.Once the clusters have been attributed,we engineer another set of API calls to GPT-4 and request the model to provide the common se-

181、mantic denominator for each thematic cluster.Table 4.b presents the result of this exercise,with the corresponding tasks in each cluster and their individual scores.X Table 4.b Tasks with high automation potential clustered into thematic groups*Thematic GroupSample TasksScoreAdministrative and Commu

182、nication TasksMaking appointments for clients;0.80Dealing with routine correspondence on their own initiative.0.80Arranging to buy and sell stocks and bonds for clients;0.80Photocopying and faxing documents;0.80Addressing circulars and envelopes by hand.0.80Customer Service and CoordinationIssuing t

183、ickets for attendance at sporting and cultural events;0.80Selecting area for fishing,plotting courses and computing navigational positions us-ing compass,charts and other aids;0.80Taking reservations,greeting guests and assisting in taking orders;0.80Determining most appropriate route;0.80Making and

184、 confirming reservations for travel,tours and accommodation;0.85Data Management and Record KeepingMaintaining records of stock levels and financial transactions;0.80Initiating records for newly appointed workers and checking records for complete-ness;0.85Importing and exporting data between differen

185、t database systems and software;0.80Operating electronic or computerized control panel from a central control room to monitor and optimize physical and chemical processes for several processing units;0.80Preparing invoices and sales contracts and accepting payment;0.806Embeddings are a vectoral high

186、-dimensional representation of the text,generated by an LLM.Standard Ada embeddings have 1536 dimensions.23ILO Working Paper 96Thematic GroupSample TasksScoreInformation Processing and Language ServicesTaking dictation and recording other matter in shorthand;0.80Translating from one language into an

187、other and ensuring that the correct meaning of the original is retained,that legal,technical or scientific works are correctly ren-dered,and that the phraseology and terminology of the spirit and style of literary works are conveyed as far as possible;0.80Converting information into codes and classi

188、fying information by codes for data-pro-cessing purposes;0.80Keying in processing instructions to programme electronic equipment;0.80Recording,preparing,sorting,classifying and filing information;0.90Providing Information and Responding to InquiriesResponding to inquiries about problems and providin

189、g advice,information and as-sistance;0.80Describing and providing information on points of interest and exhibits and re-sponding to questions;0.80Preparing and reporting short-term or long-term weather maps,forecasts and warnings relating to atmospheric phenomena such as cyclones,storms and other ha

190、zards to life and property and disseminating information about atmospheric con-ditions through a variety of media including radio,television,print and the Internet;0.80Determining customer requirements and advising on product range,price,delivery,warranties and product use and care;0.80Responding to

191、 inquiries concerning services provided and costs for room and equip-ment hire,catering and related services;0.80*Clustering relies on semantic proximity,based on K-means clustering of task embeddings.Cluster names have been as-signed by sending all tasks withing a cluster to GPT4 API and requesting

192、 a common group heading and identification of similarities.As the next step,we calculate the share of tasks with high and medium exposure in each ISCO 1-digit grouping.Figure 2 reveals in stark terms the degree of exposure among clerical support workers,among whom some 24 per cent of all tasks fall

193、into the highly exposed category.If we also account for tasks with medium-level exposure(58 per cent of all tasks),a full 82 per cent of clerical job tasks are exposed at an above-average level.This stands in contrast to the other oc-cupational groups,in which the highest share of highly exposed tas

194、ks oscillates between 1 and 4 per cent,and where the medium-exposed tasks do not exceed 25 percent.7 Even assuming large margins of error,the result is still striking.7Armed forces are absent from the figure,since they do not have any tasks scored at the level of medium and high exposure.24ILO Worki

195、ng Paper 96 X Figure 2.Tasks with medium and high GPT-exposure,by occupational category(ISCO 1-digit)3.1.Automation vs augmentation:distribution of scores across tasks and occupationsIn this next section,we analyse how the exposure to GPT-like technology could potentially affect occupations.Will the

196、 technology replace most tasks within an occupation,provoking job loss?Or could it be used to automate the more routine tasks,leaving time for more gratifying activities?To probe these questions,we turn to the analysis of the distribution of tasks for each of the 4-digit ISCO-08 occupations.Figure 3

197、 provides a visual representation of task scores for the ISCO 1-digit group of managers and clerical support workers.It shows that for the manager category,most occupations have a task-level score distribution somewhere on both sides of the medium expo-sure line of 0.5,with more tasks falling into l

198、ow-level exposure.In contrast,for clerical support workers,many occupations have an entire task distribution that falls to the right of the medium exposure threshold of 0.5.25ILO Working Paper 96 X Figure 3.Box plot of task-level scores by ISCO 4d,grouped by ISCO 1dTo determine whether the technolog

199、y has a greater potential for automation or augmenta-tion across all ISCO-08 4-digit occupations,we use a method similar to Carbonero et al.(2023).Considering an occupation as a collection of tasks with different levels of exposure to a particu-lar technology,we focus on two principal parameters of

200、the task scores distribution:(i)the mean score for a given occupation,and(ii)its standard deviation(SD).Jobs with a high mean score and a low standard deviation fall into the category of high automation potential,as the majority of the occupations tasks have high exposure scores.Jobs with a high aug

201、mentation potential are at the other extreme as they have a low occupation-level mean score,but a high standard devi-ation of the task scores.These jobs are composed of some tasks that are difficult to automate,and others that can be automated more easily.In such cases,technology is likely to have a

202、n augmenting effect,taking away some of the more exposed tasks,but still requiring the human element for the overall performance of the job(Table 5).26ILO Working Paper 96X Table 5.Grouping of occupations based on task-level scoresLow MeanHigh MeanHigh SDAugmentation potentialThe big unknownLow SDNo

203、t affectedAutomation potentialTo ensure a clear separation of the occupations with high augmentation and automation poten-tial,we apply a simple formula focussed on the extremes of this distribution.Let i and i denote the mean and standard deviation of the task-level scores for a given occupation i,

204、respectively.We define an occupation to have Augmentation potential if the following conditions are satisfied:0.4 i and i+i 0.5 (1.1)Similarly,an occupation is said to have automation potential if it fulfils these criteria:i 0.6 and i-i 0.5 (1.2)Figure 4 provides two visual representations of this g

205、rouping:the top panel pools all occupa-tional scores into one sample,while the bottom panel provides a more detailed breakdown by occupational category at ISCO-08 1-digit level.The blue trend line illustrates the relationship be-tween the two plotted variables:the occupation-level mean on the horizo

206、ntal axis and the SD of task-level scores on the vertical axis.Close to the start of the axes,mean scores and SD grow simultaneously,but the scores in this group have a low overall mean and hence low exposure.As the SD begins to plateau in the middle section around 0.2,the mean scores reach the leve

207、ls closer to 0.5,meaning that the sum of these two components starts to significantly exceed the middle exposure threshold of 0.5.As the SD begins to drop to some 0.1,the occupational scores arrive at the level of 0.6 and higher,meaning that the difference between the mean and the SD would still put

208、 such scores well above the middle exposure limit of 0.5.27ILO Working Paper 96 X Figure 4.Augmentation vs automation potential at occupational levelBased on these definitions,we produce two separate lists of occupations,one with a high auto-mation potential and one with a high augmentation potentia

209、l.Figure 5 lists occupations that not only have a high mean score across their tasks,but which also have a low SD,suggesting that the tasks scores do not move far from the overall mean.This means that such jobs are mostly com-posed of tasks that could eventually be automated,provided that other cond

210、itions are in place.28ILO Working Paper 96 X Figure 5.Occupations with high automation potentialFigure 6,in turn,presents the occupations that have a low mean score and a high SD,with the sum of the mean and the SD reaching above the limit of medium exposure.Such jobs are most likely to experience a

211、n augmenting effect of GPT technologies,while still retaining an important human component.29ILO Working Paper 96 X Figure 6.Occupations with high augmentation potential30ILO Working Paper 96 X4 Exposed occupations as a share of employment:global and income-based estimates4.1.Augmentation vs Automat

212、ion:ILO microdataNow that we know which occupations have the greatest potential for automation and augmen-tation from generative AI technology with similar properties as GPT,we can proceed with deriv-ing employment estimates globally and by country income groups.To do this,we use the ILO Harmonized

213、Microdata collection,which enables extracting detailed country-level employment information.We use microdata for 59 countries that report 4-digit microdata in ISCO-08 format:8 low-income countries(LIC),24 lower-middle-income countries(LMIC),19 upper-middle-income countries(UMIC)and 8 high-income cou

214、ntries.We take the latest year available for each country and calculate the share of each occupation belonging to our automation and augmentation cate-gories in the total employment in that country,with further disaggregation by sex.Subsequently,we construct income-group profiles,by calculating the

215、weighted mean of those automation and augmentation shares within each income group,as visualized in Figure 7a.8 X Figure 7a.Automation vs augmentation potential:shares of total employment,microdata for 59 countriesSeveral elements stand out in this comparison.First,occupations with high augmentation

216、 poten-tial constitute a significantly larger share of the total employment in each income group than the jobs with high automation potential.In the LMICs,such jobs have the highest share of the em-ployment distribution,with 14.4 per cent of total employment classified in this category.Second,augmen

217、tation-related jobs have a fairly equal gender distribution,with the shares of such jobs being held by men visibly higher only in the LMICs.8We rely on weighted means as our instrument of choice for the most balanced approach to country-level differences withing groups(see Appendix for detailed form

218、ulas).To ensure that the results are not affected by extreme differences in the distribution of values within groups,we also test calculations based on weighted-median.Since the results are stable and very similar in both cases,we keep the weighted mean as the main calculation method.31ILO Working P

219、aper 96Contrasting with that,occupations with high automation potential show significant differences across income groupings of countries and the visible trend is that they increase their share in the overall employment together with the countries income levels.In the LICs,only some 0.4 per cent of

220、total employment falls into this category,whereas in the HICs the share of such oc-cupations rises to 5.5 per cent.In addition,the share of female participation in these occupa-tions also grows with countries income levels,and in the HICs it is more than double the male share of total employment.Thi

221、s effect becomes even more apparent if we present the jobs with high automation and aug-mentation potential as a share of total employment for each sex.As demonstrated in Figure 7b,in high-income countries,jobs with high automation potential constitute 8.5 per cent of female em-ployment,compared to

222、3.9 per cent of male employment.In addition,the share of jobs with high augmentation potential is visibly higher among women than among men in all income groups.X Figure 7b.Automation vs augmentation potential:shares of total employment in each sex(ILO microdata)32ILO Working Paper 964.2.Augmentatio

223、n vs Automation:global estimateOur next step is to expand this initial estimation to the global level,with the same type of in-come-based country groupings.For this,we benchmark to the ILO modelled estimates data se-ries,which includes employment estimates for 189 countries(ILO 2023a).One of the mai

224、n challenges of producing this type of global employment figure concerns the sample representativeness for each income group.Since only 59 countries report occupational data disaggregated at the 4-digit level of ISCO-08,data for other countries needs to be estimat-ed.Fortunately,the availability of

225、country microdata increases significantly at lower-digit ISCO-08 levels.We thus exploit this greater data availability and move up the cascading structure of ISCO-08 system with each stage of estimations(see Table 6).X Table 6.Microdata coverage by levels ISCO-08:number of countries Income GroupISCO

226、-08 1-digitISCO-08 2-digitISCO-08 3-digitISCO-08 4-digitHIC4440348UMIC34302119LMIC42352924LIC2117138World1411229759We start by calculating the share of jobs in categories of automation and augmentation potential in total employment for each of the 59 countries with available 4-digit data.We then cal

227、culate the weighted mean for each income group,as previously done for Figure 7.As the next step,we calculate for these countries the share of these isolated jobs in the total jobs covered by a high-er-digit category,in this case ISCO-08 at 3-digit level.Subsequently,we calculate the weighted mean of

228、 these shares at ISCO-08 3-digit for each of the income groups and apply these to esti-mate the number of jobs in the countries for which we have ISCO-08 3-digit data,but for which ISCO 4-digit data was missing.We then repeat an analogical procedure moving up the data cov-erage ladder,that is,from I

229、SCO 3-digit to 2-digit and,finally,from 2-digit to 1-digit.At this level we arrive at an estimation that relies on data available for 141 countries,which ensures a broad coverage of data points from ILOs repository(Figure 8).The final batch of 48 countries still miss-ing at this point is estimated u

230、sing the same method,thereby aligning our calculations with the total employment figures in the official global employment estimates of the ILO for 2021,avail-able for 189 countries.99See Appendix for details.33ILO Working Paper 96 X Figure 8.Country coverage based on the level of digits in ISCO-08(

231、ILO data)1010Total refers to countries and income groupings used in ILO modelled estimates(https:/ilostat.ilo.org/resources/concepts-and-definitions/classification-country-groupings/).34ILO Working Paper 96 X Figure 9a.Global estimates:jobs with augmentation and automation potential as share of tota

232、l employ-ment Given the data limitations,the exact numbers presented in Figure 9a should be read as an in-dication of a general trend,based on the best employment estimate that can be produced at the global level for a selection of 4-digit ISCO-08 occupations.More importantly,the global esti-mate co

233、nfirms the trends already observed based on the analysis of microdata for 59 countries(Figures 7a-b).Specifically,it confirms that the number of jobs in the augmentation category is significantly higher than the number of jobs that have a high automation potential.Calculating the global figures lead

234、s to an adjustment in the ranking of income groups in the augmentation category,with UMICs and HICs having the largest share of employment with high augmenta-tion potential(13.5 and 13.4 per cent respectively)and the LICs having the lowest share(10.4 per cent).This means that,once the size and emplo

235、yment distribution aspects of individual coun-tries are considered in the estimate,globally,the share of jobs potentially exposed to automa-tion with generative AI of similar properties as the current GPT technology grows with income,but so does the share of jobs that have a high potential of experi

236、encing augmenting effects.In other words,wealthier countries are likely to face both more disruptive effects in the techno-logical transition and higher net gains from the process.We discuss these differential effects in more detail in section 6.1.The global estimates also confirm the strong gender

237、effect observed in the microdata(Figure 7b).When we disaggregate the estimate to shares of female and male employment(Figure 9b),we observe that 3.7 per cent of all female employment in the world is in jobs that are potentially au-tomatable with generative AI technology,compared with only 1.4 per ce

238、nt of male employment.In high-income countries,the share of potentially affected female jobs is 7.8 per cent,more than double the 2.9 per cent of male jobs for that income group.At the same time,the share of jobs 35ILO Working Paper 96with high augmentation potential is also greater among female tha

239、n male jobs across all income groups.This suggests that any form of technological transition would have a strongly gendered effect,with a badly managed process disproportionately harming women,and a well-managed transition potentially creating important opportunities in terms of womens empowerment.X

240、 Figure 9b.Automation vs augmentation potential:shares of total employment for each sex(global esti-mate)To further illustrate the origins of these discrepancies,it is helpful to consider a 4-digit break-down of the occupational structures across country groups.Figure 10 presents a selection of ISCO

241、 4-digit occupations with high automation potential,based on the mean share of each oc-cupation in total employment,for each income group.While the low number of responses un-derpinning some of the bars would not qualify this breakdown as statistically representative,it still provides useful insight

242、 into the overall differences in the employment structures of countries with different income levels.36ILO Working Paper 96 X Figure 10.Occupations with high automation potential,by ISCO 4-digit and income groupWe can observe that the general trend is for the share of clerical occupations to grow wi

243、th in-come,which explains the disproportionately higher potential automation effects in wealthier economies.For example,jobs of secretaries,accounting and bookkeeping clerks,or bank tell-ers and cashiers enjoy a nearly linear relationship between the countrys income and the share of employment they

244、take.This clearly reflects the general trend of the last decade,which saw many call centre and client service jobs outsourced to locations outside high-income countries.In addition,as previously discussed,such jobs are disproportionately held by women and this pattern remains visible across occupati

245、ons even at the very detailed breakdown to ISCO-08 4-dig-its.There are,however,a few notable exceptions to this rule.For example,occupations of con-tact centre salespersons and data entry clerks are relatively more present in the middle-income countries than in the high-income countries,while the jo

246、bs of application programmers are strongly dominated by men.4.3.The big unknownThe breakdown of occupations into high automation and augmentation potential provided a helpful framework to discuss the extremes of scores distribution,thereby minimizing the risk of statistical overlaps between the two

247、groups.Nevertheless,this left an important group of occu-pations,located between the automation and augmentation out of focus of the discussion.We refer to these jobs,illustrated in Figure 11a-b with green points,as“the big unknown”,since our framework and data do not allow for a clear-cut classific

248、ation of this group.In general,such jobs have a high occupational mean score,and a high variance of tasks-level scores,which means that their exposure to GPT technology can have varied and idiosyncratic effects.Depending on the technological progress of generative AI,as well as the applications buil

249、t on top of the tech-nology,some of the tasks might become more automatable,while new tasks could emerge in these professions,pushing them closer to the augmentation or automation cluster or,the more likely scenario,having them evolve into new occupations.While we refrain from speculating on the dir

250、ection of this evolution,we find it important to quantify the share of employment belong-ing to this group.37ILO Working Paper 96 X Figure 11a.The“Big Unknown”:occupations between augmentation and automation potential X Figure 11b.The“Big Unknown”:share of total employment,by income group(global est

251、imate)As illustrated in Figure 11b,these occupations constitute a nontrivial share of the global em-ployment,with some 8.6 per cent and 281 million workers falling into this category.While in the low-income and middle-income countries such jobs are to a larger extent held by men,in UMICs and HICs,wo

252、men dominate this share of total employment.38ILO Working Paper 96 X5 Managing the transition:Policies to address automation,augmentation and the growing digital divideThe estimates presented in the preceding section suggest that the recent progress in machine learning,in particular developments aro

253、und LLMs,is likely to have disruptive effects on labour markets,with larger effects in high-income countries and specific occupational groups.Still much remains unknown with respect to the progress and limitations of this and similar technologies,which will ultimately determine its overall impact.Ta

254、king GPTs current capabilities at face value and applying it to the distribution of labour markets around the world gives us an indicative pic-ture that suggests greater potential for job augmentation as opposed to automation.This find-ing represents a continuum with previous waves of technological

255、progress,despite recurring bouts of anxiety(Autor 2015,Cherry 2020).Nevertheless,policies are needed to manage the transition of those workers affected by auto-mation,in addition to managing the potential effects on job quality for those workers affected by augmentation.Indeed,both scenarios require

256、 building and strengthening systems of social dialogue,including workplace consultation.Policy attention is also needed for those countries that lack the requisite physical infrastructure and skills to benefit from the new technology.5.1 Mitigating the negative effects of automationThe analysis reve

257、aled that higher-income countries will experience the greatest effects from au-tomation as a result of the important share of share of clerical and para-professional jobs in the occupational distribution.Middle-and low-income countries will be less exposed,though cer-tain occupations that are potent

258、ially exposed to automation,such as call centre work11,figure prominently in some of these countries,particularly India and the Philippines,which dominate the worlds call centre industry.In the Philippines,a half million people were employed in call centres in 2016,of whom 53 percent were women(DOLE

259、,2018).12 The challenges,and consequences,of such adjustments should not be underestimated.For example,a study of the effects of automation on Dutch workers during 2010-2016,found that workers made redundant as a result of automation experienced a 5-year cumulative wage in-come loss of 9 per cent of

260、 an annual wage(Bessen et al.,2019).The losses were only partially offset by various benefits systems,despite the relatively robust Dutch unemployment insurance system.Workers experiencing such effects in countries with less developed insurance systems and which lack job training and job placement s

261、ervices,or where there are high levels of unem-ployment,are more vulnerable.Consultation and negotiation between employers and workers is critical for managing the tran-sition process as it encourages redeployment and training over job loss.The ILOs Employment Protection Convention(No.158,1982)inclu

262、des provisions on the termination of employment for technological reasons.It advocates,particularly in cases of collective dismissals,special pro-cedural requirements including consultations of the employer with workers representatives,notifications to the competent authorities,undertaking measures

263、to avert or minimize termi-nations and to mitigate their effects,and establishing criteria for selection for termination and 11(4222)contact centre information clerks,(4227)survey and market research interviewers,(5244)customer contact salesperson12In 2023,the ITBPA(IT and Business Processing Associ

264、ation)of the Philippines stated that the sector employed 1.5 million full-time equivalent employees in 2022(ITBPA,2023).39ILO Working Paper 96priority of rehiring.The aim of such requirements is to minimize the negative externalities from dismissal,especially when collective,as well as to better int

265、ernalize the cost of such dismissals and support an orderly process that balances the needs of workers,employers,and societies at large(Aleksnyska and Muller,2020).Social dialogue is also useful for designing and institut-ing social protection and skills development programmes that can help mitigate

266、 the negative effects of automation.One issue that will require specific attention is the gendered effects of the automation.As Figure 9 showed,the potential exposure to automation disproportionately affects the share of wom-ens employment by more than two-fold in high-income countries(7.9 per cent

267、vs 2.9 per cent)and upper-middle-income countries(2.7 per cent v 1.3 per cent).Concentrated job losses in fe-male-dominated occupations could threaten advances made in the past decades in increasing womens labour market participation.The care economy,comprising both health care and education,traditi

268、onally employs a great-er share of women,yet these are also sectors that suffer from underinvestment.According to the ILO(Addati et al.2018),achieving the SDG targets would more than double employment in these sectors from 206 million in 2015 to 475 million in 2030.In addition,some care occupations,

269、such as in long-term care,whose demand is projected to increase substantially in the next dec-ades due to ageing populations,are often characterised by poor working conditions.Meeting the demand for workers in this sector and improving their job quality so that they are a decent source of employment

270、,would be a means to not only provide a potential source of decent work for displaced workers,but would also help meet societies need for more care work.Shifting to these opportunities,however,will require greater investment in the sectors,in addition to train-ing and income support during the trans

271、ition.Another source of policy intervention is to ensure quality of the new jobs created as a result of technological change.The development of AI relies on tagging and repetitive feedback done by humans,in what is known as“microtask”work(Irani 2015;Tubaro,Casilli,and Coville 2020).For LLMs in parti

272、cular,human workers train,mould and evaluate the systems through“reinforce-ment learning”,in order to ensure the safety of such systems as well as improve accuracy(Xu et al.,2023).While no global figures exist on the number of microtask workers,estimates from the mid-2010s suggested 9 million worker

273、s from across the globe(Kuek et al.2015).This figure has most certainly grown since then and is likely to continue to expand,as new and often small players enter the market of LLMs.A recently leaked note from Googles engineers noted that“the barrier to entry for LLM training and experimentation has

274、dropped from the total output of a major research organization to one person,an evening,and a beefy laptop”(Patel and Ahmad 2023).This dramatic decrease in the cost and ease of entry to the LLM market points to an increase in de-mand for domain-specific labelled datasets curated by microtask workers

275、.Much microtask work has been conducted on digital labour platforms,either through crowd-sourcing websites or though businesses processing firms that directly hire workers.Microtask jobs mediated through crowdsourcing platforms,are paid by the task and regulated by civil con-tracts,meaning that the

276、workers have none of the labour protections or social security benefits that come with the employment relationship.The poor working conditions of much platform work prompted ILO constituents to agree to a two-year standard setting discussion beginning on 2025 with a view to crafting an international

277、 labour standard on decent work in the platform economy that can guide national regulation(ILO 2023d).5.2 Ensuring job quality under augmentationTechnology can also affect job quality in its application at the workplace.While the technology can allow the more routine tasks that one does to be automa

278、ted,potentially leaving time for more engaging work,it can also be implemented in a way that limits workers agency or acceler-ates work intensity.Concerns over AIs integration at the workplace has focused on the growth of 40ILO Working Paper 96algorithmic management,essentially work settings in whic

279、h“human jobs are assigned,optimized,and evaluated through algorithms and tracked data”(Lee et al.2015).Algorithmic management is a defining feature of digital labour platforms,but it is also pervasive in offline industries such as the warehousing and logistics sectors.In warehouses an automated,“voi

280、ce-picking”system directs warehouse staff to pick certain products in the warehouse,while using data collection to monitor workers and set the pace of work(Matopoulos 2011).Besides lacking autonomy to organize their work or set its pace,workers also have little ability to provide feedback or discuss

281、 with management about the organization of work(Wood,2021).The integration of generative AI into other fields such as banking,insurance,social services,and customer service more broadly may have similar effect.Technological advancements are often felt more immediately at the workplace level and are

282、usu-ally best addressed at the workplace.As a result,whether the effect of technology on working conditions is positive or negative depends in large part on the voice that workers have in the design,implementation and use of technology.Having such voice relies in turn on the opportu-nities for worke

283、r participation and dialogue.This can take place either through formalized set-tings,such as works councils or guidance provided in collective bargaining agreements,or less formally,in workplaces where there is a high degree of employee engagement,such as in organ-izational structures that support t

284、eamwork,problem-solving and decentralized decision-making(ILO 2023c).Studies on Europe have shown that it is the countries with stronger and more coop-erative forms of workplace consultation,essentially the Nordic countries,followed by Germany,where workers are more open to technological adoption at

285、 the workplace.Yet even in Denmark,focus group discussions with workers on digital integration reveal a desire for greater attention to the implementation and organization of technology at the workplace so as to better meet the needs of end users(Refslund and Borello,2023).In addition to consultatio

286、n at the workplace,there is also need for laws that regulate AIs appli-cation at the workplace.To date,much of the discussion on regulation of AI has ignored its pos-sible effects on working conditions(Moore 2023).Where there has been discussion,the focus has overwhelmingly been on voluntary standar

287、ds of AI ethics,ignoring the uneven power re-lations inherent in working relationships(Cole et al.2022).AI tools may aggravate power rela-tions at the workplace,especially if workers cannot have access to the data used to survey their activities,if there are no mechanisms in place to assess the ex-p

288、ost use of the technology in the workplace,or if decisions on dismissal are taken without proper recourse to conflict resolution mechanisms.Adams-Prassl et al.(2023)advocate for a prohibition of worker monitoring and data collection outside of work(temporally or geographically)or in contexts where i

289、t poses risks to human dignity or the exercise of fundamental rights,in addition to other limitations.The design and application of such regulations is best crafted through tripartite systems,in which workers,employers and governments representatives engage with equal voice.The negotiations should b

290、uild on existing tripartite consultation mechanisms and structures and use the already exist-ing labour rights and norms as the point of departure.Giving the quickly evolving nature of AI and its iterative learning process,mechanisms for ex-post evaluation and tripartite governance need to be built

291、into the regulation.5.3 Addressing the digital divideA potentially more significant consequence of a wider adoption of generative AI products could be an increased divergence in productivity between the high-and low-income countries.Larger shares of jobs falling into the augmentation category sugges

292、t that,at least in near future,gen-erative AI systems similar to GPT are more likely to become productivity tools,supporting and speeding up the execution of some tasks within certain occupations.The digital divide will influ-ence how the benefits of such productivity tools are distributed among soc

293、ieties and countries,with high-income countries and privileged groups likely to reap the biggest rewards.41ILO Working Paper 96Low-income countries,in particular,are at risk of falling behind.While up to 13 per cent of em-ployment in these countries is found in the potential augmentation category,in

294、 practice po-tential benefits of GPT technologies are likely to be limited,as the lack of reliable infrastructure will constrain its application.To begin with,such technology is dependent on access and cost of broadband connectivity,as well as electricity.In 2022,one-third of the global population,c

295、orre-sponding to some 2.7 billion people,still did not have access to the internet(Figure 11).Among the two-thirds that do have access,many would not be able to use GPT technologies due to the limitations in the quality of their connection or the cost of the service.Even more fundamental than the in

296、ternet,reliable electricity provision is often a challenge.According to the World Bank Enterprise Survey,49 percent of registered firms in developing countries experienced electrical outages,averaging 4.5 days per month and lasting 4 hours on average.13 X Figure 11.Share of population not using the

297、internet1413https:/www.enterprisesurveys.org/en/data/exploretopics/infrastructure14Authors calculations based on available country data for the most recent year(ITU 2023).Map created with Datawrapper.The boundaries shown,designations used,and any other information shown does not imply official endor

298、sement or acceptance by the International Labour Organization.42ILO Working Paper 96 X Figure 12.A classic growth path:income and occupational diversificationOn the other hand,with the right conditions in place,a new wave of technology could fuel growth opportunities.In the past,technological advanc

299、ements have spurred new and successful indus-tries in many developing countries.One such example is the M-Pesa money service,which relied on the diffusion of mobile telephones in Kenya.The service,in turn,increased financial inclusion thus helping to propel the growth of SMEs and led to creation of

300、a network of 110,000 agents,40 times the number of bank ATMs in Kenya(Buku and Meredith 2012;de Soyres et al.2018).Similarly,a study of the diffusion of 3G coverage in Rwanda between 2002 and 2019 found that increased mobile internet coverage was positively associated with the employment growth,in-c

301、reasing both skilled and unskilled occupations(Caldarola et al.2022).Hjort and Poulsen(2019)also find positive employment effects,from the arrival of internet in 12 African countries,albeit with a slight bias towards skilled occupations.These gains are attributed to increases in produc-tivity and gr

302、owth of markets that followed increased connectivity.Among the developing countries,further distinction needs to be made.While middle-income countries,are more exposed to the automating effects of GPT technologies,their digital infra-structure and skilled workforce can also be an asset for spawning

303、the growth of complementa-ry industries.Although India and the Philippines are at risk of losing some call centre work,their dominance in business process outsourcing may provide the needed foundation for the devel-opment of new industries.43ILO Working Paper 96 XConclusionIn this paper,we attempted

304、 to quantify some of the potential effects of generative AI on occu-pations from a global perspective.Our study provided a global estimate of the number of jobs in the categories that are most exposed to technologies with similar capabilities as GPT-4,by re-lying on the international standard of ISC

305、O-08 and linking the task-level scores to employment distributions reflected in official ILO statistics.We subsequently discussed the consequences of these findings in the context of differential impacts that can be expected depending on coun-tries income levels.We also highlighted the possible cons

306、equences for job quality,in order to draw attention to this important effect on the world of work that has too often been ignored in discussions of digital technologies impact on labour.The analysis was based on the top threshold of current technological possibilities and relied on three bold assump

307、tions.First,we assumed that the tasks,for which automation scores were estimated,would be executed in the context of a high-income country.This ignores the more limited potential for deployment in lower-income countries,where the necessary infrastructure is typically of lower standard,unreliable and

308、 often more expensive,and where lower skill and wage levels make the costs of technological adoption relatively high.Second,we relied on GPT-4 to predict the scores,which is likely to reflect an apex of technological optimism when it comes to ease of deployment,that in practice is difficult to opera

309、tionalize.Third,without being able to make reliable predictions on future technological progress,we focused on the potential of task automation as of today,without speculating on the numbers of new jobs that might emerge.This approach might have been expected to generate alarming estimates of net jo

310、b loss but it did not.Rather,our global estimates point to a future in which work is transformed,but still very much in existence.Our findings largely align with the evolving body of academic literature concerning previous waves of technological transformations,but some of the trends we identify are

311、 new as a result of our exclusive focus on LLMs,and GPT more specifically.While early studies of potential AI adoption identified low-skill,repetitive and routine jobs as those with the highest potential of automa-tion(e.g.,McKinsey 2016;Frey and Osborne 2017),in which a computer-based system could

312、be coupled with a machine to replace a human in manual production jobs(Autor 2015;Acemoglu and Restrepo 2020),more recent literature has highlighted the ability of Machine Learning sys-tems to improve their performance in non-routine tasks(Brynjolfsson et al.,2018;Ernst et al.,2019;Webb 2019;Lane an

313、d Saint-Martin 2021).We argue that the emergence of GPT reinforces this shifting picture,due to its refined ability to perform cognitive tasks,such as analysing text,drafting documents and messages,or searching through private repositories and the web for additional information.As a consequence,our

314、study indicates that at least in the short run this new wave of automation will focus on a different group of workers,typically associated with“knowledge work”(Surawski 2019).The occupational group with the highest share of tasks exposed to GPT technology are the cler-ical jobs,where the majority of

315、 tasks fall at least into medium-level exposure,and about a quar-ter of tasks are highly exposed to potential automation.As a result of technological progress,many such jobs might never emerge in developing countries,where they traditionally served as a vehicle for increasing female employment.For o

316、ther types of“knowledge work”,exposure is only partial,suggesting a stronger augmentation potential and productivity benefits,rather than job displacement.These findings align with some of the most recent literature on generative AI systems with a global focus.A recent study by McKinsey(2023)points

317、to a similar group of“knowledge work”occupations and tasks as having the highest level of exposure,though with a significantly high-er suggested level of displacement.WEFs global survey,focussed on large enterprises,also lists clerical and administrative jobs among occupations with fastest expected

318、declines(WEF 2023).Estimates provided by Goldman Sachs(2023)suggest a slightly higher level of potential 44ILO Working Paper 96automation than our calculations,but with the general conclusion aligning with our main find-ing that“most jobs and industries are only partially exposed to automation and a

319、re thus more likely to be complemented rather than substituted by AI”.The more moderate effects observed in our estimations stem from several factors.First,we rely on ISCO-08 as the source of tasks and occupations,which is more adequate for a study with a global character than the US-oriented O*NET

320、database.Second,the application of ILOs coun-try-level employment statistics adds important nuance to the actual number of jobs that exists in those categories,bringing out income-based differences that affect the final employment effects at the global level.Third,we do not attempt to make predictio

321、ns on the evolution of the tech-nology.While the growing capabilities of generative AI and the range of secondary applications that can be built on top of this technology are likely to increase the numbers of jobs in both the augmentation and automation categories identified in our paper,our analysi

322、s suggests that the general contours of transformation identified in this study will remain valid for the coming years.Ultimately,we argue that in the realm of work,generative AI is neither inherently good nor bad,and that its socioeconomic impacts will largely depend on how its diffusion is managed

323、.The questions of power balance,voice of the workers affected by labour market adjustments,respect for existing norms and rights,and adequate use of national social protection and skills training systems will be crucial elements for managing AIs deployment in the workplace.Without proper policies in

324、 place,there is a risk that only some of the well-positioned countries and market par-ticipants will be able to harness the benefits of the transition,while the costs to affected work-ers could be brutal.Therefore,for policy makers,our study should not read as a calming voice,but rather as a call fo

325、r harnessing policy to address the technological changes that are upon us.45ILO Working Paper 96Appendix 1.Countries with missing ISCO-08 4-digit data:estimation procedureTo illustrate our estimation method,we use the example of jobs identified as having high au-tomation potential.For an income grou

326、p IG,denote the total employment as TIG.The total em-ployment in each income group is the sum of the total jobs Ji in all the countries i that belong to the income group IG:T=JIGiIGiFor each country i,denote Ai as the number of jobs with high automation potential and Ji as the total number of jobs.T

327、he share of automation jobs Si is then calculated as:S=AJiiiThe weight Wi for each country i in income group IG is defined as the share of the countrys em-ployment in the total employment of that income group:W=JTiiIGThe weighted mean MIG for each income group IG is then the sum of the product of th

328、e weights Wi and the automation job shares Si for all countries i in income group IG:M=W SIGiIGiiFor each ISCO-08 3-digit category d,in country i where 4-digit ISCO-08 data exists,the total num-ber of jobs J3di is given by:J3=J4dikD3kidiwhere J4ki is the total number of jobs in the 4-digit category

329、k that falls under the 3-digit category d in that country.The share S3di of automation jobs in 4-digit category d to the total jobs in the corresponding 3-digit category d in country i is given by:S3=AJ3dididiwhere Adi is the number of automation jobs in the 4-digit category d,and J3di is the total

330、number of jobs in the 3-digit category d in country i.At the next step,each 3-digit share S3di is weighted by the total employment Ei in the country i relative to the total employment EIG in the income group IG.The weighted mean WMSIG for in-come group IG is then calculated as:WMS=E*S3EIGiIGidiiIGiF

331、or each country i with missing 4-digit data but available 3-digit data,the estimated number of automation jobs Ai can then be calculated using the weighted mean share WMSIG of the corre-sponding income group and the total employment Ei in country i:46ILO Working Paper 96A=WMS*EiIGiWe then repeat an

332、analogical procedure moving up the data coverage ladder,that is,from ISCO 3-digit to 2-digit,from 2-digit to 1-digit,and finally to global coverage.47ILO Working Paper 96ReferencesAcemoglu,Daron,David Autor,Jonathon Hazell,and Pascual Restrepo.2022.“Artificial Intelligence and Jobs:Evidence from Onl

333、ine Vacancies.”Journal of Labor Economics 40(S1):S293340.https:/doi.org/10.1086/718327.Acemoglu,Daron,and Pascual Restrepo.2020.“Robots and Jobs:Evidence from US Labor Markets.”Journal of Political Economy 128(6):21882244.https:/doi.org/10.1086/705716.Adams-Prassl,Jeremias,Sangh Rakshita,Halefom Abraha,and M.Six Silberman.2023.“Towards an International Standard for Regulating Algorithmic Managemen