iloWorld Bank：拉丁美洲生成式人工智能的就業(yè)機(jī)會(huì)與數(shù)字鴻溝

ILOWorkingPaper121

July/2024

、BufferorBottleneck?EmploymentExposuretoGenerativeAIandtheDigitalDivideinLatinAmerica

Authors/Pawe?Gmyrek,HernanWinkler,SantiagoGarganta

Copyright?InternationalLabourOrganizationandtheWorldBank2024

ThisisanopenaccessworkdistributedundertheCreativeCommonsAttribution3.0IGOLicense(

/licenses/by/3.0/igo

).Userscanreuse,share,adaptandbuildupontheoriginalwork,asdetailedintheLicense.TheILOandTheWorldBankmustbeclearlycreditedastheownersoftheoriginalwork.TheuseoftheemblemoftheILOandTheWorldBankisnotpermittedinconnectionwithusers’work.

Attribution–Theworkmustbecitedasfollows:Gmyrek,P.,Winkler,H.,Garganta,S.BufferorBottleneck?EmploymentExposuretoGenerativeAIandtheDigitalDivideinLatinAmerica.ILOWorkingPaper121.Geneva:InternationalLabourOfficeandTheWorldBank,2024.

Translations–Incaseofatranslationofthiswork,thefollowingdisclaimermustbeaddedalongwiththeattribution:ThistranslationwasnotcreatedbytheInternationalLabourOrganization(ILO)orTheWorldBankandshouldnotbeconsideredanofficialILOorWorldBanktranslation.TheILOandTheWorldBankarenotresponsibleforthecontentoraccuracyofthistranslation.

Adaptations–Incaseofanadaptationofthiswork,thefollowingdisclaimermustbeaddedalongwiththeattribution:ThisisanadaptationofanoriginalworkbytheInternationalLabourOrganization(ILO)andTheWorldBank.ResponsibilityfortheviewsandopinionsexpressedintheadaptationrestssolelywiththeauthororauthorsoftheadaptationandarenotendorsedbytheILOorTheWorldBank.

ThisCClicensedoesnotapplytonon-ILOorWorldBankcopyrightmaterialsincludedinthispublication.Ifthematerialisattributedtoathirdparty,theuserofsuchmaterialissolelyresponsibleforclearingtherightswiththerightholder.

AnydisputearisingunderthislicensethatcannotbesettledamicablyshallbereferredtoarbitrationinaccordancewiththeArbitrationRulesoftheUnitedNationsCommissiononInternationalTradeLaw(UNCITRAL).Thepartiesshallbeboundbyanyarbitrationawardrenderedasaresultofsucharbitrationasthefinaladjudicationofsuchadispute.

Licensing),1211Geneva22,Switzerland,orbyemailto

rights@

.

AllqueriesonrightsandlicensingshouldbeaddressedtotheILOPublishingUnit(Rightsand

ISBN9789220410028(print),ISBN9789220410035(webPDF),ISBN9789220410042(epub),ISBN9789220410066(mobi),ISBN9789220410059(html).ISSN2708-3438(print),ISSN2708-3446(digital)

/10.54394/TFZY7681

ThedesignationsemployedinILOandWorldBankpublicationsandthepresentationofmaterialthereindonotimplytheexpressionofanyopinionwhatsoeveronthepartoftheILOandTheWorldBankconcerningthelegalstatusofanycountry,areaorterritoryorofitsauthorities,orconcerningthedelimitationofitsfrontiersorboundaries.Detailsat

/disclaimer

ThisarticleisaproductofthestaffoftheWorldBankandtheILO.IthasbeenreleasedbothinTheWorldBankPolicyResearchWorkingPaperSeriesandtheILOWorkingPaperSeries.Theresponsibilityforopinionsexpressedinsignedarticles,studiesandothercontributionsrestssolelywiththeirauthors,andpublicationdoesnotconstituteanendorsementbytheILOorTheWorldBankoftheopinionsexpressedinthem.

Referencetonamesoffirmsandcommercialproductsandprocessesdoesnotimplytheiren-

dorsementbytheILOorTheWorldBank,andanyfailuretomentionaparticularfirm,commercialproductorprocessisnotasignofdisapproval.

InformationonILOpublicationsanddigitalproductscanbefoundat:

/

research-and-publications

ILOWorkingPaperssummarizetheresultsofILOresearchinprogress,andseekto

stimulatediscussionofarangeofissuesrelatedtotheworldofwork.CommentsonthisILO

WorkingPaperarewelcomeandcanbesentto

gmyrek@

.

Authorizationforpublication:RichardSamans,Director,ResearchDepartment

ILOWorkingPaperscanbefoundat:

/research-and-publications#working

-papers

Suggestedcitation:

Gmyrek,P.,Winkler,H.,Garganta,S.2024.BufferorBottleneck?EmploymentExposure

toGenerativeAIandtheDigitalDivideinLatinAmerica,ILOWorkingPaper121(Geneva,

ILOandTheWorldBank,2024).

/10.54394/TFZY7681

01ILOWorkingPaper121

Abstract

Empiricalevidenceonthepotentialimpactsofgenerativeartificialintelligence(GenAI)ismostlyfocusedonhigh-incomecountries.Incontrast,littleisknownabouttheroleofthistechnologyonthefutureeconomicpathwaysofdevelopingeconomies.ThispapercontributestofillthisgapbyestimatingtheexposureoftheLatinAmericanlabourmarkettoGenAI.ItprovidesdetailedstatisticsofGenAIexposurebetweenandwithincountriesbyleveragingarichsetofharmonizedhouseholdandlabourforcesurveys.Toaccountfortheslowerpaceoftechnologyadoptionindevelopingeconomies,itadjuststhemeasuresofexposuretoGenAIbyusingthelikelihoodofaccessingdigitaltechnologiesatwork.ThisisthenusedtoassesstheextenttowhichthedigitaldivideacrossandwithincountrieswillbeabarriertomaximizetheproductivitygainsamongoccupationsthatcouldotherwisebeaugmentedbyGenAItools.Thefindingsshowthatcertaincharacteristicsareconsistentlycorrelatedwithhigherexposure.Specifically,urban-basedjobsthatrequirehighereducation,aresituatedintheformalsector,andareheldbyindividualswithhigherincomesaremorelikelytocomeintointeractionwiththistechnology.Moreover,thereisapronouncedtilttowardyoungerworkersfacinggreaterexposure,includingtheriskofjobau-tomation,particularlyinthefinance,insurance,andpublicadministrationsectors.Whenadjust-ingforaccesstodigitaltechnologies,thefindingsshowthatthedigitaldivideisamajorbarriertorealizingthepositiveeffectsofGenAIonjobsintheregion.Inparticular,nearlyhalfofthepo-sitionsthatcouldpotentiallybenefitfromaugmentationarehamperedbylackofuseofdigitaltechnologies.Thisnegativeeffectofthedigitaldivideismorepronouncedinpoorercountries.

Abouttheauthors

Pawe?GmyrekisaSeniorResearcherattheResearchDepartmentoftheILO.

HernanWinklerisaSeniorEconomistattheWorldBankPovertyandEquityGlobalPracticeforLatinAmericaandtheCaribbean.

SantiagoGargantaisaSeniorResearcherattheCenterforDistributive,LaborandSocialStudies(CEDLAS)oftheNationalUniversityofLaPlata(UNLP).

02ILOWorkingPaper121

Tableofcontents

Abstract

Abouttheauthors

Acronyms

01

01

05

、

Introduction

06

、

1

LACregionandthetheoreticaleffectsofGenAI

08

、

2

Methods

OccupationalexposuretoGenAI

Useofacomputeratwork

15

15

19

、

3

Findings

Cross-countrycomparisonsofthelevelsofexposure

Impactofdigitalinfrastructureonthepotentialoftransformation

Within-countrypatterns

Whichoccupationsdrivetheeffects?

Differentialexposureacrossearningslevels

22

22

26

29

30

32

、

Finaldiscussion

35

Appendix38

References45

Acknowledgements50

03ILOWorkingPaper121

ListofFigures

Figure1.GDPpercapita,populationandincomestatusofLACcountriesinthesample08

Figure2.Automationandaugmentationpotential:LACvsotherregions09

Figure3.Internetcoveragevspercapitaincome:globalandLAC11

Figure4.OccupationsintheLACregion,byISCO1-digitandgender13

Figure5.CoverageofISCO-084-digitmicrodatainSEDLAC(WB)andILOharmonizedmicro-

datacollection17

Figure6.HierarchicalclusteringbasedonISCO2-digitshares,GDP(PPP)andtotalpopulation18

Figure7.TotalexposuretoGenAIbycountry23

Figure8.Automationpotential-detailedbreakdownofsocio-economiccharacteristics24

Figure9.Augmentationpotential-detailedbreakdownofsocio-economiccharacteristics25

Figure10.Jobswithaugmentationpotentialandaccesstocomputeratwork,basedon

PIAACdata27

Figure11.Exposurebycountry,exposuretypeandaccesstodigitalinfrastructure28

Figure12.Exposurebycountry,typeanddetailedcountry-levelcharacteristics30

Figure13.ISCO2-digitoccupationsbytypeofexposureandcountry(shareofexposure>25%)31

Figure14.EarningsofoccupationsexposedtoGenAI,byemploymentstatus(exposure

above25%)33

FigureA1.ComparisonofTechXposurescoresvsGBBscores(meanbyoccupation,z-scores)38

FigureA2.ComparisonofFeltenetal.(2023)MLscoresvsGBBscores(z-scores)38

FigureA3.LabourmarketdistributioninLACcountriesbyISCO-082-digitoccupationsandsex39

FigureA4.RankingofcountriesbythetypeofGenAIexposure40

FigureA5.ComparisonofresultsoncomputerusebetweenPIAAC(atwork)andSEDLAC(at

home)-augmentationcategory40

FigureA6.Jobsinaugmentationcategorythatdonotuseacomputedatwork:totalsby

country40

04ILOWorkingPaper121

ListofTables

Table1.DistributionofAIExposurebyDemographicandSocioeconomicCategoriesin

SEDLACData19

TableA1.IndividualSEDLACobservationsbycountryandyear41

TableA2.EstimatedcoefficientsofcomputeruseatworkfromPIAAC41

TableA3.ResultsofthepooledOLSwithallindividualobservations,withcountry-levelnor-

malizedpopulationweights43

05ILOWorkingPaper121

Acronyms

EMEmergingMarkets

GDPGrossDomesticProduct

GBBGmyrek,BergandBescond(asusedinyourstudyforcitation)

GenAIGenerativeAI

GPT-4GenerativePre-trainedTransformer4

HICHighIncomeCountries

ILOInternationalLabourOrganization

IMFInternationalMonetaryFund

ISCOInternationalstandardClassificationofOccupations

ISCO-08InternationalStandardClassificationofOccupations,2008version

LACLatinAmericaandtheCaribbean

LLMLargeLanguageModels

OECDOrganizationforEconomicCooperationandDevelopment

PIAACProgrammefortheInternationalAssessmentofAdultCompetencies

PPPPurchasingPowerParity

SEDLACSocio-EconomicDatabaseforLatinAmericaandtheCaribbean

TFPTotalFactorProductivity

USUnitedStates

WBWorldBank

WEFWorldEconomicForum

06ILOWorkingPaper121

、Introduction

PublicattentiontoGenerativeAI(GenAI)hasbeenontherisesincetheintroductionoftheconversationalmodels,suchasChatGPT,BardorGemini.TheimpressiveabilitiesoftheLargeLanguageModels(LLM),followedbyotherneuralnetwork-basedAIsystemscapableofgener-atingimageandevenvideofromsimpletextpromptshaveraisedarangeofimportantethi-calandsecurityquestionsfornationalpolicymakersandinternationalcooperationstructures.However,thetopicthatcapturesmostdailyattentionofregularcitizensisthepotentialimpactofthesequicklyadvancingtoolsonjobs.

IntheUnitedStates(US),overhalfofalladultsaremoreworriedthanexcitedaboutAIindailylife,citingthe“l(fā)ossofhumanjobs”astheirmostimportantconcern(FaverioandTyson,2022;PewResearchCenter,2023;Rutgers,2024).InSwitzerland,a2023surveyfocusedspecificallyonGenAIrevealedthatof1,000respondentsalreadyworkingwithacomputer,almosthalf(43%)wereconcernedaboutlosingtheirjobinthenextfiveyears,withthosefrequentlyusingGenAIatworkbeingdisproportionately(69%)moreconcerned(Gramppetal.,2023).ThissuggestsarapiddeparturefrommorepositiveassessmentsofAIinsurveyscollectedbyOECDpriortothearrivalofpubliclyaccessiblechatbotsinlate2022(Laneetal.,2023;OECD,2023).1

Notsurprisingly,thepotentialtransformationthatmightresultfromtheinteractionofGenAIwithlabourmarketshasalsoattractedgrowingattentionamongscholars.Mainresearchques-tionshavecanteredaroundtheimpactonemployment,emergingoccupations,productivityandjobquality.2ArecentpaperfromtheIMFprovidesacomprehensiveoverviewofthisliterature,atthesametimehighlightingthescarcityofstudiesthatgobeyondhigh-incomecountries(HICs)(ComunaleandManera,2024).

Bridgingthisresearchgap,ourstudyprovidesnewevidenceonthepotentialimpactsofGenAIacrosslabourmarketsintheLatinAmericaandtheCaribbean(LAC)region.Buildingontheap-proachdevelopedbyGmyrek,BergandBescond(2023)–GBBhereafter–weprovidenewevidenceonAIexposurebetweenandwithincountriesbyleveragingharmonizedhouseholdandlabourforcesurveysforLACfromtheWorldBank(WB)andtheInternationalLabourOrganization(ILO).Bybuildingonthecomparativestrengthsofthedatasetsfrombothinstitutions,wedevelopacompleteregionaloverview,accompaniedbycountry-levelestimatesofthepotentialoccupationalexposure,withfurtherbreakdownsbydetaileddemographicandlabourmarketcharacteristics.

Animportantcontributionofthisstudyistoprovideafirstattemptatadaptingmeasuresofjobs’exposuretoGenAItothecontextofdevelopingcountries,whereevenworkersinoccupa-tionsthataregenerallyexpectedtobenefitfromGenAImaynotbeabletoreapitsbenefitsduetopooraccesstodigitalinfrastructure.WeimplementthisadjustmentbyestimatingmeasuresofcomputeruseatworkacrossISCO2-digitoccupations,workersandcountry-levelcharacter-isticsbasedonPIAACdataandbysubsequentlyimputingthemintoindividualobservationsincountry-levelsurveysincludedintheSEDLACdatabase.WethenusethismeasuretocreatetwocategoriesamongworkerswhoareexpectedtobenefitfromGenAIusebecauseofthena-tureoftheiroccupations:thosewhohaveaccesstodigitaltechnologies,andthosewhodonot.Thesizeofthelatterisanindicatorofthenumberofworkerswhowillnotbeabletoenjoythe

1InOECD’ssurveyofworkers“fourinfiveworkerssaidthatAIhadimprovedtheirperformanceatworkandthreeinfivesaidithadincreasedtheirenjoymentofwork”(…)“WorkerswerealsopositiveabouttheimpactofAIontheirphysicalandmentalhealth,aswellasitsusefulnessindecisionmaking”(OECD,2023).

2E.g.,seeBrynjolfssonetal.,2023;Huietal.,2023,Berajaetal.,2023,Adams-Prassletal.,2023.

07ILOWorkingPaper121

productivitybenefitsofGenAIeventhoughtheirjobscouldtheoreticallybenefitfromthetrans-formation.Wealsodiscussthedetaileddemographicsofthegroupsthataremostlikelytobenegativelyaffectedbytheseinfrastructurelimitations.

Ourfindingsindicatethatbetween30and40percentofemploymentintheLACisexposedinsomewaytoGenAI.Thisexposureislinkedwiththeeconomicstatusofcountries,suggestingthatincomelevelsareastrongcorrelateofGenAI’simpactonlabourmarkets.Thistotallevelofexposureincludesthreecategories:exposedtoautomation,augmentation,and“thebigun-known”.Thelatterincludesoccupations,which–dependingontheprogressoftechnologyandtheuseofadjacenttechnologicalapplications,suchasLLM-basedagents–couldfallclosertoautomationoraugmentation.

CertaincharacteristicsconsistentlycorrelatewithhigheroverallGenAIexposure.Specifically,ur-ban-basedjobsthatrequirehighereducation,aresituatedintheformalsector,andareheldbyindividualswithhigherrelativeincomesaremorelikelytocomeintointeractionwiththistech-nology.Theshareofjobsexposedtoautomationisrelativelysmallbutnontrivialatabout2to5percentoftotalemployment.Youngerandfemaleworkerstendtofacegreaterautomationexposure,particularlyinthefinance,insurance,andpublicadministrationsectors.Atthesametime,thesharesofjobsthatcouldbenefitfromaproductivetransformationwithGenAIarecon-sistentlyhigherthanthosewithautomationrisksacrossallLACcountries,rangingbetween8and12percentofemploymentacrosscountries.Thisisparticularlythecaseforthejobsineducation,healthandpersonalservices.Inaddition,thesectorsorientedtowardscustomerservice(retail,trade,hotels,restaurants,etc.)faceanelevatedexposureto"thebigunknown".Thiscategoryencompassesthelargest(14-21percent)shareofemploymentinourestimates,demonstratingthat,whiletheconceptofoccupationalexposureiseasiertoestablish,thepreciseeffectsonhowmanyoccupationsmightevolvearehardertopredictforalargeshareoftoday’slabourmarkets.

Finally,wefindthataccesstodigitaltechnologiesisacriticaldeterminantoftheextenttowhichworkerscanharnessthepotentialbenefitsofGenAI.Nearlyhalfofthepositionsthatcouldpo-tentiallybenefitfromaugmentationarehamperedbydigitalshortcomingsthatwillpreventthemfromrealizingthatpotential.Specifically,6.24percentofjobsheldbywomenand6.22percentofthoseheldbymenareaffectedduetothesegaps.Similarlimitationsapplytothejobsinthe“bigunknown”category:eventhoughsomeofthemcouldpotentiallypivottowardsaugmenta-tionthroughincreasingcomplementaritybetweenGenAIandthehumanworkerintheseoccu-pations,thedigitalgapswillpreventlargesharesofthesejobsfromsuchascenario.

Therestofthisstudyisstructuredasfollows:section2providesageneraloverviewoftheLACregionandelaboratesonthetheoreticaleffectsonecouldexpectfromtheinteractionofGenAIwithitslabourmarkets,section3discussesthedataandmethodsappliedtoouranalysis,section4providesadetailedbreakdownofourfindings,withthefinaldiscussionpresentedinsection5.

08ILOWorkingPaper121

、1LACregionandthetheoreticaleffectsofGenAI

ThedefinitionoftheregionofLatinAmericaandtheCaribbean(LAC)canhaveavaryingscopeacrossdifferentinstitutions.Inthecaseofourstudy,werelyonaheuristicapproachofinclud-ingthemaximumnumberofcountriesforwhichwecanfinddataofsufficientqualityintheda-tabasesoftheWB,ILOandanyotherrelevantsources.Thefinalsampleincludes22countries,showninFigure1accordingtotheirincome-basedgroupingusedbytheWBin2022,andtotalpopulation.Theregionisveryheterogeneous,fromverysmallislandsintheCaribbeanwithfew-erthanhalfamillioninhabitants,tocountrieswithlargepopulationssuchasBrazilandMexico.Accordingly,itrangesfromhigh-incomecountriessuchasUruguayandPanamatolower-incomecountriessuchasNicaraguaandHonduras.

、Figure1.GDPpercapita,populationandincomestatusofLACcountriesinthesample

WhilethereisalargebodyofliteratureanalysingtheimpactsoftechnologicalchangeonthelabourmarketoutcomesofLAC(forexample,seeDutzetal.2018),theexpectedincidenceofGenAIislikelytobedifferentfromthatofprevioustechnologicalbreakthroughs.Autor(2024)claimsthatthetransformationalimpactofnewtechnologiesonlabouristhroughthereshapingofhumanexpertise,andheillustratesthishypothesiswithtwoexamples:theadoptionofmassproductioninthe18thand19thcenturies,andtheadoptionofdigitaltechnologiessincethe1960s.Theemergenceofmassproductionchangedthecomplexworkofartisansintoself-con-tainedandsimpletaskscarriedoutbyproductionworkers,usingnewmachinery,andoverseen

09ILOWorkingPaper121

byotherswithhigherlevelsofeducation.Theincreaseddemandforthis“massexpertise”wasaccompaniedbyanincreasingnumberofhigh-schoolgraduates,leadingtotheriseofanewmiddleclass.Later,digitaltechnologiesallowedtocarryoutroutinetasksbyencodingthemindeterministicrules.Non-routinetaskscouldnotbereplacedbythistechnologybecausetheyarenotattainedbylearningrules,butthroughlearningbydoing.Asaresult,digitaltechnologiesgaverisetoanewformofexpertisebyallowingprofessionalstoobtainandprocessinforma-tionmoreefficiently,andtherebyhavingmoretimetointerpretandapplyit.Theroutinejobsreplacedbythistechnologytendedtobeinthemiddleoftheearningsdistribution,whilethenon-routinejobscomplementedbydigitalizationtendedtobeatthetop,leadingtoapolariza-tionofthelabourmarket.AI,incontrast,canperformnon-routinetasksthatoftenrequiretacitknowledge.Forexample,itcanallownon-eliteworkers(suchasnurses)toengageincomplexdecision-making,anditcanautomatesomeofthetaskscarriedoutbyhigh-skillworkerssuchasdoctors,softwareengineersandlawyers.However,asdescribedbelow,thefinalimpactsonjobswilldependonotherfactorsaswell.Forexample,thedirectautomationimpactsofGenAIonjobsmaybeoffsetbypositiveimpactsonproductivity,whichwouldstrengthenlabourdemand.

WhilenopreviousgranularassessmentsofoccupationalexposuretoGenAIexistfortheLACregion,therehavebeencomparisonstootherregionsmadeinbroaderstudies.Forexample,GBB(2023)placeLACsomewhereinthemiddleoftheregionalrankingofpotentialautomationexposure,with2.5percentoftotalemploymentfallingintothiscategory(Figure2).Intermsofaugmentationpotential,thesamestudyrankedLACasthethirdfromthebottom(12.8percentofemployment).Similarly,whiletheWEF(2023)globalstudydidnotprovideaspecificregionalranking,itprojecteda5-yearstructurallabourchurninLACat22percent,slightlybelowtheglobalaverage(23percent).Inotherwords,theLACregioncanbecharacterizedashavingeconomieswithanaveragelevelofexposuretoGenAIthatislessthanthatofthemostindustrializednations,yethigherthanthatfoundinlow-incomeregions,makingitarelevantintermediatebenchmark.

、Figure2.Automationandaugmentationpotential:LACvsotherregions

Intheory,theriseofGenAIanditspotentialpositiveimpactsonlabourproductivitycouldposeasignificantopportunityfordevelopingcountries.Somerecentprivatesectorstudiesevensug-gestthataggregateimpactofwidespreadAIadoptioncouldaddbetween0.1and1.5ppofan-nualproductivitygrowthinHICs,withslightlylowerfiguresestimatedforEmergingMarkets(EM)(GoldmanSachs,2023;McKinsey,2023).SuchprojectionsmightbeparticularlyenticingfortheLACregion,whichhaslonggrappledwithapersistentproductivitygapincomparisontootherareasoftheworld.WhilethedevelopingnationsinAsiaandEuropemanagedtonarrowtheirproductivitygapwiththeUnitedStatesbetween1990and2019,suchgapincreasedfortheLACregionduringthesameperiod(IMF,2022).Recenttrendsalsoraiseconcerns,since,despitesome

10ILOWorkingPaper121

countryvariation(Erumbanetal.,2024),theoverallproductivitygrowthhasbeenalmostzeroinLACeversincethestartoftheglobalproductivityslowdownofthelast10years(Dieppe,2021).Comparedtootherregions,barrierstoinnovationandtechnologyadoptionhavebeenparticu-larlysalientfactorslimitingproductivitygrowthinLAC.

CouldGenAIhelpunlockthisproductivityimpasse?RecentempiricalstudiesfocusedontheuseofGenAIinparticularoccupationalsettingssuggestthatthepositiveimpactsonproductivitycanbelarge.Forexample,Pengetal.(2023)implementedacontrolledexperimentamongprofes-sionalprogrammersandfoundthataccesstoaGenAIassistantreducedthetimetocompleteprogrammingtasksby56percent.Brynjolfssonetal.(2023)findthataccesstoGenAIincreasesproductivityamongcustomersupportworkersintermsofissuesresolvedperhour,whichisdriv-enmostlybytheboostofperformanceamongthenoviceandlow-skillworkers.Similarly,NoyandZhang(2023)findthathavingaccesstoChatGPThelpsimprovetheproductivityofwritingprofessionals,byincreasingthequality