生成式AI和就業(yè)：對工作數(shù)量和質量的潛在全球影響分析+Generative+AI+and+Jobs：A+global+Analysis+of+Potential+Effects+on+Job+Quantity+and+Quality

ILOWorkingPaper96

August/2023

X

GenerativeAIandJobs:Aglobal

analysisofpotentialeffectsonjob

quantityandquality

Authors/Pawe?Gmyrek,JanineBerg,DavidBescond

Copyright?InternationalLabourOrganization2023

ThisisanopenaccessworkdistributedundertheCreativeCommonsAttribution4.0InternationalLicense(

/licenses/by/4.0/

).Userscanreuse,share,adaptandbuildupontheoriginalwork,asdetailedintheLicense.TheILOmustbeclearlycreditedastheown-eroftheoriginalwork.TheuseoftheemblemoftheILOisnotpermittedinconnectionwithusers’work.

Attribution–Theworkmustbecitedasfollows:Gmyrek,P.,Berg,J.,Bescond,D.GenerativeAIandJobs:Aglobalanalysisofpotentialeffectsonjobquantityandquality.ILOWorkingPaper96.

Geneva:InternationalLabourOffice,2023.

Translations–Incaseofatranslationofthiswork,thefollowingdisclaimermustbeaddedalongwiththeattribution:ThistranslationwasnotcreatedbytheInternationalLabourOrganization(ILO)andshouldnotbeconsideredanofficialILOtranslation.TheILOisnotresponsibleforthecontentoraccuracyofthistranslation.

Adaptations–Incaseofanadaptationofthiswork,thefollowingdisclaimermustbeaddedalongwiththeattribution:ThisisanadaptationofanoriginalworkbytheInternationalLabourOrganization(ILO).ResponsibilityfortheviewsandopinionsexpressedintheadaptationrestssolelywiththeauthororauthorsoftheadaptationandarenotendorsedbytheILO.

ThisCClicensedoesnotapplytonon-ILOcopyrightmaterialsincludedinthispublication.Ifthematerialisattributedtoathirdparty,theuserofsuchmaterialissolelyresponsibleforclearingtherightswiththerightholder.

Anydisputearisingunderthislicensethatcannotbesettledamicablyshallbereferredtoarbitra-tioninaccordancewiththeArbitrationRulesoftheUnitedNationsCommissiononInternationalTradeLaw(UNCITRAL).Thepartiesshallbeboundbyanyarbitrationawardrenderedasaresultofsucharbitrationasthefinaladjudicationofsuchadispute.

AllqueriesonrightsandlicensingshouldbeaddressedtotheILOPublishingUnit(RightsandLicensing),1211Geneva22,Switzerland,orbyemailto

rights@

.

ISBN9789220395356(print),ISBN9789220395363(webPDF),ISBN9789220395370(epub),ISBN9789220395387(mobi),ISBN9789220395394(html).ISSN2708-3438(print),ISSN2708-3446(digital)

/10.54394/FHEM8239

ThedesignationsemployedinILOpublications,whichareinconformitywithUnitedNationspractice,andthepresentationofmaterialthereindonotimplytheexpressionofanyopinionwhatsoeveronthepartoftheILOconcerningthelegalstatusofanycountry,areaorterritoryorofitsauthorities,orconcerningthedelimitationofitsfrontiers.

Theresponsibilityforopinionsexpressedinsignedarticles,studiesandothercontributionsrestssolelywiththeirauthors,andpublicationdoesnotconstituteanendorsementbytheILOoftheopinionsexpressedinthem.

Referencetonamesoffirmsandcommercialproductsandprocessesdoesnotimplytheiren-dorsementbytheILO,andanyfailuretomentionaparticularfirm,commercialproductorpro-cessisnotasignofdisapproval.

InformationonILOpublicationsanddigitalproductscanbefoundat:

/publns

ILOWorkingPaperssummarizetheresultsofILOresearchinprogress,andseektostimulatediscussionofarangeofissuesrelatedtotheworldofwork.CommentsonthisILOWorkingPaperarewelcomeandcanbesentto

RESEARCH@

,

berg@

.

Authorizationforpublication:RichardSamans,DirectorRESEARCH

ILOWorkingPaperscanbefoundat:

/global/publications/working-papers

Suggestedcitation:

Gmyrek,P.,Berg,J.,Bescond,D.2023.GenerativeAIandJobs:Aglobalanalysisofpotentialef-fectsonjobquantityandquality,ILOWorkingPaper96(Geneva,ILO).

/10.54394/

FHEM8239

01ILOWorkingPaper96

Abstract

ThisstudypresentsaglobalanalysisofthepotentialexposureofoccupationsandtaskstoGenerativeAI,andspecificallytoGenerativePre-TrainedTransformers(GPTs),andthepossibleimplicationsofsuchexposureforjobquantityandquality.ItusestheGPT-4modeltoestimatetask-levelscoresofpotentialexposureandthenestimatespotentialemploymenteffectsatthegloballevelaswellasbycountryincomegroup.Despiterepresentinganupper-boundestimateofexposure,wefindthatonlythebroadoccupationofclericalworkishighlyexposedtothetech-nologywith24percentofclericaltasksconsideredhighlyexposedandanadditional58percentwithmedium-levelexposure.Fortheotheroccupationalgroups,thegreatestshareofhighlyex-posedtasksoscillatesbetween1and4percent,andmediumexposedtasksdonotexceed25percent.Asaresult,themostimportantimpactofthetechnologyislikelytobeofaugmentingwork–automatingsometaskswithinanoccupationwhileleavingtimeforotherduties–asop-posedtofullyautomatingoccupations.

Thepotentialemploymenteffects,whetheraugmentingorautomating,varywidelyacrosscoun-tryincomegroups,duetodifferentoccupationalstructures.Inlow-incomecountries,only0.4percentoftotalemploymentispotentiallyexposedtoautomationeffects,whereasinhigh-incomecountriesthesharerisesto5.5percent.Theeffectsarehighlygendered,withmorethandoubletheshareofwomenpotentiallyaffectedbyautomation.Thegreaterimpactisfromaugmenta-tion,whichhasthepotentialtoaffect10.4percentofemploymentinlow-incomecountriesand13.4percentofemploymentinhigh-incomecountries.However,sucheffectsdonotconsiderinfrastructureconstraints,whichwillimpedethepossibilityforuseinlower-incomecountriesandlikelyincreasetheproductivitygap.

Westressthattheprimaryvalueofthisanalysisisnotthepreciseestimates,butratherthein-sightsthattheoveralldistributionofsuchscoresprovidesaboutthenatureofpossiblechanges.Suchinsightscanencouragegovernmentsandsocialpartnerstoproactivelydesignpoliciesthatsupportorderly,fair,andconsultativetransitions,ratherthandealingwithchangeinareactivemanner.Moreover,thelikelyramificationsonjobqualitymightbeofgreaterconsequencethanthequantitativeimpacts,bothwithrespecttothenewjobscreatedbecauseofthetechnology,butalsothepotentialeffectsonworkintensityandautonomywhenthetechnologyisintegrat-edintotheworkplace.Forthisreason,wealsoemphasizetheneedforsocialdialogueandreg-ulationtosupportqualityemployment.

Abouttheauthors

Pawe?GmyrekisSeniorResearcherintheResearchDepartmentoftheILO.JanineBergisSeniorEconomistintheResearchDepartmentoftheILO.DavidBescondisDataScientistintheILO’sDepartmentofStatistics.

02ILOWorkingPaper96

Tableofcontents

Abstract

Abouttheauthors

Acronyms

01

01

05

X

Introduction

07

X

1

MethodsandData

1.1.ISCOdataonoccupationsandtasks

1.2.Promptdesignandsequence

10

11

12

X

2

AssessmentofthePredictions,RobustnessTestsandtheBoundsforAnalysis

17

X

3

Results

3.1.Automationvsaugmentation:distributionofscoresacrosstasksandoccupations

20

24

X

4

Exposedoccupationsasashareofemployment:globalandincome-basedestimates

4.1.AugmentationvsAutomation:ILOmicrodata

4.2.AugmentationvsAutomation:globalestimate

4.3.Thebigunknown

30

30

32

36

X

5

Managingthetransition:Policiestoaddressautomation,augmentationandthegrowingdigitaldivide

5.1Mitigatingthenegativeeffectsofautomation

5.2Ensuringjobqualityunderaugmentation

5.3Addressingthedigitaldivide

38

38

39

40

X

Conclusion

43

Appendix1.CountrieswithmissingISCO-084-digitdata:estimationprocedure45

References47

AcknowledgementsanduseofGPT51

03ILOWorkingPaper96

ListofFigures

Figure1.Meanautomationscoresbyoccupation,basedonISCOandGPTtasks21

Figure2.TaskswithmediumandhighGPT-exposure,byoccupationalcategory(ISCO1-digit)24

Figure3.Boxplotoftask-levelscoresbyISCO4d,groupedbyISCO1d25

Figure4.Augmentationvsautomationpotentialatoccupationallevel27

Figure5.Occupationswithhighautomationpotential28

Figure6.Occupationswithhighaugmentationpotential29

Figure7a.Automationvsaugmentationpotential:sharesoftotalemployment,microdata

for59countries30

Figure7b.Automationvsaugmentationpotential:sharesoftotalemploymentineachsex

(ILOmicrodata)31

Figure8.CountrycoveragebasedonthelevelofdigitsinISCO-08(ILOdata)33

Figure9a.Globalestimates:jobswithaugmentationandautomationpotentialasshareof

totalemployment34

Figure9b.Automationvsaugmentationpotential:sharesoftotalemploymentforeachsex

(globalestimate)35

Figure10.Occupationswithhighautomationpotential,byISCO4-digitandincomegroup36

Figure11a.The“BigUnknown”:occupationsbetweenaugmentationandautomationpotential37

Figure11b.The“BigUnknown”:shareoftotalemployment,byincomegroup(globalestimate)37

Figure11.Shareofpopulationnotusingtheinternet41

Figure12.Aclassicgrowthpath:incomeandoccupationaldiversification42

04ILOWorkingPaper96

ListofTables

Table1.ISCO-08Structureofoccupationsandtasksusedinthestudy11

Table2.SampleoftasksanddefinitionsfromISCOandpredictedbyGPT-414

Table3.Sampleoftask-levelscores(high-incomecountrycontext)15

Table4.aTestofscoreconsistency(100task-levelpredictions)17

Table4.bTaskswithhighautomationpotentialclusteredintothematic22

groups*

Table5.Groupingofoccupationsbasedontask-levelscores26

Table6.MicrodatacoveragebylevelsISCO-08:numberofcountries32

05ILOWorkingPaper96

Acronyms

3G

ThirdGeneration(referringtoagenerationofstandardsformobiletelecom-munications)

Ada

AlanguagemodelbyOpenAIusedtogenerateembeddings

AGI

ArtificialGeneralIntelligence

AI

ArtificialIntelligence

ANN

ArtificialNeuralNetwork

API

ApplicationProgrammingInterface

ATMs

AutomatedTellerMachines

CPU

CentralProcessingUnit

DL

DeepLearning

DOLE

DepartmentofLaborandEmployment

ESCO

EuropeanSkills,Competences,QualificationsandOccupations

GPTs

GenerativePre-TrainedTransformers

GPT-4

GenerativePre-TrainedTransformer4

GPU

GraphicsProcessingUnit

HIC

High-IncomeCountries

ICT

InformationandCommunicationsTechnology

ILO

InternationalLabourOrganization

ISCO

InternationalStandardClassificationofOccupations

ISCO-08

InternationalStandardClassificationofOccupations2008

K-Means

K-MeansClusteringAlgorithm

LFS

LabourForceSurveys

LIC

Low-IncomeCountries

LLMs

LargeLanguageModels

06ILOWorkingPaper96

LMIC

Lower-Middle-IncomeCountries

ML

MachineLearning

NLP

NaturalLanguageProcessing

OECD

OrganisationforEconomicCo-operationandDevelopment

O*NET

OccupationalInformationNetwork

OpenAI

OpenArtificialIntelligence(organization'sname)

Python

High-levelprogramminglanguage

RL

ReinforcementLearning

SD

StandardDeviation

SMEs

SmallandMedium-sizedEnterprises

UMIC

Upper-Middle-IncomeCountries

US

UnitedStates

USD

UnitedStatesDollar

UMIC

Upper-Middle-IncomeCountries

US

UnitedStates

07ILOWorkingPaper96

XIntroduction

Eachnewwaveoftechnologicalprogressintensifiesdebatesonautomationandjobs.CurrentdebatesonArtificialIntelligence(AI)andjobsrecallthoseoftheearly1900swiththeintroduc-tionofthemovingassemblyline,oreventhoseofthe1950sand1960s,whichfollowedtheintro-ductionoftheearlymainframecomputers.Whiletherehavebeensomenodstothealienationthattechnologycanbringbystandardizingandcontrollingworkprocesses,inmostcases,thedebateshavecentredontwoopposingviewpoints:theoptimists,whoviewnewtechnologyasthemeanstorelieveworkersfromthemostarduoustasks,andthepessimists,whoraisealarmabouttheimminentthreattojobsandtheriskofmassunemployment.

Whathaschangedindebatesontechnologyandworkers,however,isthetypesofworkersaf-fected.Whiletheadvancesintechnologyintheearly,midandevenlate-1900swereprimarilyfocusedonmanualworkers,technologicaldevelopmentsincethe2010s,inparticulartherapidprogressofMachineLearning(ML),hascentredontheabilityofcomputerstoperformnon-rou-tine,cognitivetasks,andbyconsequencepotentiallyaffectwhite-collarorknowledgeworkers.Inaddition,thesetechnologicaladvancementshaveoccurredinthecontextofmuchstrong-erinterconnectednessofeconomiesacrosstheglobe,leadingtoapotentiallylargerexposurethanlocation-based,factory-levelapplications.Yetdespitethesedevelopments,toanaverageworker,eveninthemosthighlydevelopedcountries,thepotentialimplicationsofAIhave,untilrecently,remainedlargelyabstract.

ThelaunchofChatGPTmarkedanimportantadvanceinthepublic’sexposuretoAItools.Inthisnewwaveoftechnologicaltransformation,machinelearningmodelshavestartedtoleavethelabsandbegininteractingwiththepublic,demonstratingtheirstrengthsandweaknessesindailyuse.ThechatfunctiondramaticallyshortenedthedistancebetweenAIandtheenduser,simultaneouslyprovidingaplatformforawiderangeofcustom-madeapplicationsandinno-vations.Giventhesesignificantadvancements,itisnotsurprisingthatconcernsoverpotentialjoblosshaveresurged.

WhileitisimpossibletopredicthowgenerativeAIwillfurtherdevelop,thecurrentcapabilitiesandfuturepotentialofthistechnologyarecentraltodiscussionsofitsimpactonjobs.Scepticstendtobelievethatthesemachinesarenothingmorethan“stochasticparrots”–powerfultextsummarizers,incapableof“l(fā)earning”andproducingoriginalcontent,withlittlefutureforgen-eralpurposeuseandunsustainablecomputingcosts(Benderetal.2021).Ontheotherhand,morerecenttechnicalliteraturefocusedontestingthelimitsofthelatestmodelssuggestsanincreasingcapabilitytocarryout“novelanddifficulttasksthatspanmathematics,coding,vision,medicine,law,psychologyandmore”,andageneralabilitytoproduceresponsesexhibitingsomeformsofearly“reasoning”(Bubecketal.2023).Someassessmentsgoasfarassuggestingthatmachinelearningmodels,especiallythosebasedonlargeneuralnetworksusedbyGenerativePre-trainedTransformers(GPT,seeTextBox1),mighthavethepotentialtoeventuallybecomeageneral-purposetechnology(Goldfarb,Taska,andTeodoridis2023;Eloundouetal.2023).1Thiswouldhavemultipliereffectsontheeconomyandlabourmarkets,asnewproductsandservic-eswouldlikelyspringfromthistechnologicalplatform.

Associalscientists,wearenotinpositiontotakesidesinthesetechnicaldebates.Instead,wefocusonthealreadydemonstratedcapabilitiesofGPT-4,includingcustom-madechatbotswithretrievalofprivatecontent(suchascollectionsdocuments,e-mailsandothermaterial),natu-rallanguageprocessingfunctionsofcontentextraction,preparationofsummaries,automatedcontentgeneration,semantictextsearchesandbroadersemanticanalysisbasedontextem-beddings.LargeLanguageModels(LLMs)canalsobecombinedwithotherMLmodels,suchas

1Thethreemaincharacteristicsofgeneral-purposetechnologiesarepervasiveness,abilitytocontinueimprovingovertime,andabil-itytospawnfurtherinnovation(JovanovicandRousseau,2005).

08ILOWorkingPaper96

speech-to-textandtext-to-speechgeneration,potentiallyexpandingtheirinteractionwithdif-ferenttypesofhumantasks.Finally,thepotentialofinteractingwithlivewebcontentthroughcustomagentsandplugins,aswellasthemultimodal(notexclusivetotext,butalsocapableofreadingandgeneratingimage)characterofGPT-4makesitlikelythatthistypeoftechnologywillexpandintonewareas,therebyincreasingitsimpactonlabour.

Departingfromtheseobservations,thisstudyseekstoaddtheglobalperspectivetothealreadylivelydebateonpossiblechangesthatmayresultinthelabourmarketsasaconsequenceoftherecentadventofgenerativeAI.Westressthefocusofourworkontheconceptsof“exposure”and“potential”,whichdoesnotimplyautomation,butratherlistsoccupationsandassociatedemploymentfiguresforjobsthataremorelikelytobeaffectedbyGPT-4andsimilartechnologiesinthecomingyears.Theobjectiveofthisexerciseisnottoderiveheadlinefigures,butrathertoanalysethedirectionofpossiblechangesinordertofacilitatethedesignofappropriatepolicyresponses,includingthepossibleconsequencesonjobquality.

Theanalysisisbasedon4-digitoccupationalclassificationsandtheircorrespondingtasksintheISCO-08standard.ItusestheGPT-4modeltoestimateoccupationalandtask-levelscoresofex-posuretoGPTtechnologyandsubsequentlylinksthesescorestoofficialILOstatisticstoderiveglobalemploymentestimates.Wealsoapplyembedding-basedtextanalysisandsemanticclus-teringalgorithmstoprovideabetterunderstandingofthetypesoftasksthathaveahighauto-mationpotentialanddiscusshowtheautomatingandaugmentingeffectswillstronglydependonarangeofadditionalfactorsandspecificcountrycontext.

Wediscusstheresultsofthisanalysisinthebroadercontextoflabourmarkettransformations.Weputparticularfocusonthecurrentdisparitiesindigitalaccessacrosscountriesofdifferentincomelevels,thepotentialforthisnewwaveoftechnologicaltransformationtoaggravatesuchdisparities,andtheensuingconsequencesonproductivityandincome.Wealsogiveconsider-ationtojobswithhighestautomationandaugmentationpotentialanddiscussgender-specificdifferences.Theanalysisdoesnottakeintoaccountthenewjobsthatwillbecreatedtoaccom-panythetechnologicaladvancement.Twentyyearsago,therewerenosocialmediamanagers,thirtyyearsagotherewerefewwebdesigners,andnoamountofdatamodellingwouldhaverenderedaprioripredictionsconcerningavastarrayofotheroccupationsthathaveemergedinthepastdecades.AsdemonstratedbyAutoretal.(2022),some60percentofemploymentin2018intheUnitedStateswasinjobsthatdidnotexistinthe1940s.

Indeed,themainvalueofstudiessuchasthisoneisnotinthepreciseestimates,butratherinunderstandingthepossibledirectionofchange.Suchinsightsarenecessaryforproactivelyde-signingpoliciesthatcansupportorderly,fair,andconsultativetransitions,ratherthandealingwithchangeinareactivemanner.Forthisreason,wealsoemphasizethepotentialeffectsoftechnologicalchangeonworkingconditionsandjobqualityandtheneedforworkplaceconsul-tationandregulationtosupportthecreationofqualityemploymentandtomanagetransitionsinthelabourmarket.

Wehopethatthisresearchwillcontributetoneededpolicydebatesondigitaltransformationintheworldofwork.Whiletheanalysisoutlinespotentialimplicationsfordifferentoccupationalcategories,theoutcomesofthetechnologicaltransitionarenotpre-determined.Itishumansthatarebehindthedecisiontoincorporatesuchtechnologiesanditishumansthatneedtoguidethetransitionprocess.Itisourhopethatthisinformationcansupportthedevelopmentofpoliciesneededtomanagethesechangesforthebenefitofcurrentandfuturesocieties.Weintendtousethisbroadglobalstudyasanopeningtomorein-depthanalysesatcountrylevel,withaparticularfocusondevelopingcountries.

09ILOWorkingPaper96

XTextBox1:WhatareGPTs?

GenerativePre-TrainedTransformersbelongtothefamilyofLargeLanguageModels–atypeofMachineLearningmod-elbasedonneuralnetworks.The“generative”partreferstotheirabilitytoproduceoutputofacreativenature,whichinlanguagemodelscantaketheformofsentences,paragraphs,orentiretextstructures,withcharacteristicsoftenun-distinguishablefromthatproducedbyhumans.“Pre-trained”referstotheinitialtrainingonalargecorpusoftextdata,typicallythroughunsupervisedorself-supervisedlearning,duringwhichthemodellearnsaboutthetextstructurebytemporarilymaskingpartofthecontentandtryingtominimizeerrorsinthepredictionofthemaskedwords.Followingpre-training,suchmodelsarefurtherfine-tunedwiththeuseoflabelleddataandso-called“reinforcementlearning”,makingthemmoresuitableforspecifictasks.Thispartoftrainingisoftenperceivedasaspecializedjob,executedbyahandfuloftechnicalexperts.Inreality,itislabourintensiveandinvolvesmanyinvisiblecontributors(Dzieza2023).Itsprerequisiteistheproductionofvastamountsoflabelleddata,typicallydonebyworkersoncrowdsourcingplatforms.“Transformers”refertotheunderlyingmodelarchitecture,whichusesnumerousmechanisms,suchasattentionandself-attentionframeworks,todevelopweightsrelatedtotheimportanceoftextelements,suchaswordsinasentence,whicharesubsequentlyusedforpredictions(Vaswanietal.2017).

WhileGPTspecificallyreferstomodelsdevelopedbyOpenAI(GPT-1,2,3and4),thistypeofarchitectureisusedbymanymorelanguagemodelsalreadyavailablecommercially.ThelaunchofChatGPTon30November2022madeGPTsmorepopularamongthepublic,asitmadeitpossibleforindividualswithnoprogrammingknowledgetointeractwithGPT-3(andeventuallyGPT-4)throughachatbotfunctionwithahuman-liketone.Forresearchpurposesandmorecom-plexapplications,suchlanguagemodelsaretypicallymorepowerfulwhenusedthroughanApplicationProgrammingInterface(API).AnAPIisadeveloperaccesspointthatreliesonaquery-responseprotocolwiththeuseofprogrammingsoftware.Inourcase,werelyonaPythonscriptbasedonOpenAIlibrary,designedtoconnecttoGPT-4model,provideafine-tunedpromptandreceivearesponse,whichissubsequentlystoredinadatabaseonourserver.ThisenablesbulkprocessingoflargenumbersofrequestsandreliesontheGPT-4modelwithmoreparametersthanwhatisaccessiblethroughthepublicChatfunction.

10ILOWorkingPaper96

X1MethodsandData

Therearetwoprincipalapproachestotheanalysisofautomationofoccupations(GeorgieffandHyee2021).Thefirstistousedataonjobvacanciestounderstandhowdemandforspecificskillsevolvesovertime.Moststudiesusingthisapproachharnessdatafromonlinerecruitmentplat-forms(CammeraatandSquicciarini2021;Acemogluetal.2022)tomeasurethefrequencyofref-erencestoAI(ortoanyothertechnologyofinterest)inthetextofthejobdescription.Theseref-erencesarethenusedasaproxyforthedemandforspecificskillsand,byitsextension,aproxyfortherateoftechnologicaladoptionattheenterpriselevel.Thisapproachworkswellincoun-trieswithahighonlinepresenceinrecruitment,thoughitdoesnotalwayscapturetheindus-triesaffectedasaresultofsubcontracting.Theapproach,however,islesswellsuitedforaglobalstudycoveringcountrieswithlessonlinepresence,asmostvacanciesarenotadvertisedonon-lineplatformsbutrecruitedthroughothermeansofcommunication(GeorgieffandHyee2021).

Thesecondapproachistofocusonoccupationalstructures,withtheideaofestimatingtheau-tomationpotentialoftasksorskillsthatmakeupagivenjob.Theadvantageofthismethodisthatsuchoccupationalclassificationscaneasilybelinkedtoofficiallabourmarketstatistics,whichisofparticularimportanceforunderstandingglobal,regionalandincome-baseddiffer-entials.Thisstrandofliteratureisrich,butfrequentlymisunderstood,especiallywhenitcomestocommunicatingitsfindingstothepublic,asmediainterpretationstendtoblurthedistinc-tionbetweenautomationpotentialandactualdeploymentintheworkplace.Forexample,FreyandOsborne’s(2013,2017)influentialstudyhasbeencitedover12,000times,oftenfordiffer-enttypesofdoomsdaypronouncements,eventhoughtheauthorswereclearaboutthedistinc-tionbetweenpotentialandpredictedeffects.Arangeofstudiesfollowthisresearchtradition,attemptingtocalculatedifferenttypesofoccupationalautomationscoresinOECDcountries(Brynjolfsson,Mitchell,andRock2018;Felten,Raj,andSeamans2018;Felten,Raj,andSeamans2019;AcemogluandRestrepo2020;FossenandSorgner2022)orevencombiningoccupationalandjobpostingdata(GeorgieffandHyee2021).Someauthorshavealsotakenupthechallengeofproducingbetterestimatesfordevelopingcountries(BalliesterandElsheikhi2018),oftenbytryingtolinkdetailedoccupationaldataand