AI對(duì)勞動(dòng)力市場(chǎng)的影響：跨國(guó)差異與分配影響+Labor+Market+Exposure+to+AI：Cross-country+Differences+and+Distributional+Implications

LaborMarketExposuretoAI:Cross-country

Differencesand

DistributionalImplications

CarloPizzinelli,AugustusPanton,MarinaM.Tavares,MauroCazzaniga,LongjiLi

WP/23/216

IMFWorkingPapersdescriberesearchin

progressbytheauthor(s)andarepublishedto

elicitcommentsandtoencouragedebate.

TheviewsexpressedinIMFWorkingPapersare

thoseoftheauthor(s)anddonotnecessarily

representtheviewsoftheIMF,itsExecutiveBoard,

orIMFmanagement.

NAr

2023

ARY

OCT

*TheauthorswouldliketothankFlorenceJaumotte,GiovanniMelina,andEmmaRockallforhelpfulcomments.

?2023InternationalMonetaryFund

WP/23/216

IMFWorkingPaper

ResearchDepartment

LaborMarketExposuretoAI:Cross-countryDifferencesandDistributionalImplications

PreparedbyCarloPizzinelli,AugustusPanton,MarinaM.Tavares,MauroCazzaniga,LongjiLi

AuthorizedfordistributionbyFlorenceJaumotte

October2023

IMFWorkingPapersdescriberesearchinprogressbytheauthor(s)andarepublishedtoelicit

commentsandtoencouragedebate.TheviewsexpressedinIMFWorkingPapersarethoseofthe

author(s)anddonotnecessarilyrepresenttheviewsoftheIMF,itsExecutiveBoard,orIMFmanagement.

ABSTRACT:ThispaperexaminestheimpactofArtificialIntelligence(AI)onlabormarketsinbothAdvancedEconomies(AEs)andEmergingMarkets(EMs).WeproposeanextensiontoastandardmeasureofAI

exposure,accountingforAI'spotentialaseitheracomplementorasubstituteforlabor,wherecomplementarityreflectslowerrisksofjobdisplacement.Weanalyzeworker-levelmicrodatafrom2AEs(USandUK)and4

EMs(Brazil,Colombia,India,andSouthAfrica),revealingsubstantialvariationsinunadjustedAIexposure

acrosscountries.AEsfacehigherexposurethanEMsduetoahigheremploymentshareinprofessionalandmanagerialoccupations.However,whenaccountingforpotentialcomplementarity,differencesinexposure

acrosscountriesaremoremuted.Withincountries,commonpatternsemergeinAEsandEMs.Womenand

highlyeducatedworkersfacegreateroccupationalexposuretoAI,atbothhighandlowcomplementarity.

Workersintheuppertailoftheearningsdistributionaremorelikelytobeinoccupationswithhighexposurebutalsohighpotentialcomplementarity.

RECOMMENDEDCITATION:Pizzinelli,C.,A.Panton,M.M.Tavares,M.Cazzaniga,andL.Li,(2023)“Labor

MarketExposuretoAI:Cross-countryDifferencesandDistributionalImplication.”IMFWokringPaper23/216

JELClassificationNumbers:

J23,O33

Keywords:

Artificialintelligence;Employment;Occupations;EmergingMarkets

Author’sE-MailAddress:

cpizzinelli@,apanton@,mmendestavares@,mauro98cazzaniga@,lli4@,

LaborMarketExposuretoAI:Cross-country

DifferencesandDistributionalImplications

CarloPizzinelli*

IMF

AugustusPanton

IMF

MarinaM.Tavares*

IMF

MauroCazzanigaLongjiLi

FGV-SPIMF

September22,2023

Abstract

ThispaperexaminestheimpactofArtificialIntelligence(AI)onlabormarketsinbothAdvancedEconomies(AEs)andEmergingMarkets(EMs).WeproposeanextensiontoastandardmeasureofAIexposure,accountingforAI’spotentialaseitheracom-plementorasubstituteforlabor,wherecomplementarityreflectslowerrisksofjobdisplacement.Weanalyzeworker-levelmicrodatafrom2AEs(USandUK)and4EMs(Brazil,Colombia,India,andSouthAfrica),revealingsubstantialvariationinunadjustedAIexposureacrosscountries.AEsfacehigherexposurethanEMsduetoahigheremploymentshareinprofessionalandmanagerialoccupations.However,whenaccountingforpotentialcomplementarity,differencesinexposureacrosscountriesaremoremuted.Withincountries,commonpatternsemergeinAEsandEMs.WomenandhighlyeducatedworkersfacegreateroccupationalexposuretoAI,atbothhighandlowcomplementarity.Workersintheuppertailoftheearningsdistributionaremorelikelytobeinoccupationswithhighexposurebutalsohighpotentialcomple-mentarity.

Keywords:Artificialintelligence,Employment,Occupations,EmergingMarketsJELCodes:J23,J23,O33

*Correspondingauthors:CarloPizzinelliandMarinaM.Tavares,InternationalMonetaryFund,70019thSt.NW,Washington,DC,20431,USA.Email:

cpizzinelli@

mmendestavares@

1TheauthorswouldliketothankFlorenceJaumotte,GiovanniMelina,AlexanderCopestake,andEmmaRockallforhelpfulcomments.Disclaimer:TheviewsexpressedinthisstudyarethesoleresponsibilityoftheauthorsandshouldnotbeattributabletotheInternationalMonetaryFund,itsExecutiveBoard,oritsmanagement.

1

1Introduction

TherapiddevelopmentofArtificialIntelligence(AI)hassparkedconsiderablediscus-

sionregardingitsimpactonlabormarkets.1

Byautomatingtasks,personalizingexperiences,andimprovingqualitycontrol,AIcoulddramaticallyenhanceproductivityacrossvarioussectors,presentinganunprecedentedrevolutionintheworkplace.Despitethispromisingoutlook,theswiftprogressofAI,coupledwithcontinuedR&D,createssubstantialuncer-

taintysurroundingitssocioeconomicimplications(LaneandSaint-Martin,

2021;

Agrawal

etal.

,

2018).EconomistslargelyagreethatAIcouldbolstersocietalwealthinthelongrun,

yetconcernspersistoveritspotentialtodisruptemploymentinmanyindustries.

Inthisfast-evolvinglandscape,threesignificantareasofuncertaintystandout.First,itremainsunclearhowAItechnologiesmightserveaseithersubstitutesorcomplementsforhumanlaborinspecifictasksandoccupations,ultimatelyleadingto“winnersandlosers”in

thejobmarket(Autor,

2022).Second,thereisinterestinunderstandinghowexposuretoAI

variesacrosscountries,andinparticularwhethertherearesystematicdifferencesbetweenAdvancedEconomies(AEs)andEmergingMarkets(EMs).Third,withincountries,exposuretotherisksandbenefitsofAIislikelytodifferacrossdemographicgroupsandskilllevels,makingimplicationsforeconomicdisparitiesdifficulttopredict.

Inthispaper,weofferpreliminaryinsightsintothesequestions.First,weproposeanadjustmenttoastandardmeasureofAIoccupationalexposure(AIOE)tocaptureAI’spotentialtocomplementorsubstituteforlaborineachoccupation.Second,weapplyboththeoriginalmeasureandthecomplementarity-adjustedonetolaborforcemicrodatafromsixcountries,withaparticularemphasisonEMs.OuranalysisshedslightondifferencesinexposuretoAIacrosscountries,disentanglingthosewithgreaterpotentialtobenefitfromcomplementarityandthoseatgreaterriskfromsubstitution.Finally,withineachcountry,weexaminehowexposurevariesacrossdemographicgroups,skilllevels,andtheincomedistribution.

Recentresearchhasfocusedon“exposure”toAIacrossthespectrumofoccupations.TheproposeddefinitionsofexposureconsiderhowAIapplicationsoverlapwiththehuman

1IthasbeenarguedthatAIfulfillsthedefinitionofaGeneral-PurposeTechnology(GPT)andthereforeholdsthepotentialtospurasustainedwaveofeconomicgrowthandinnovation.

Lipseyetal.

(2005)define

aGPTasatechnologythat(i)iswidelyused,(ii)hasthepotentialforcontinuousinnovation,(iii)generatescomplementaryinnovations.ExamplesofGPTsarethesteamengine,electricity,andtheinternet.Scholars

generallyagreethatAI,asasuiteoftechnologies,isaGPT(Agrawaletal.,

2018)andpotentiallysome

ofitsindividualsub-fields,suchasGenerativeAIandMachineLearning,individuallyfulfillthedefinition

(Goldfarbetal.,

2023)

.

2

abilitiesneededtoperformagivenoccupation(asintheAIOEindexof

Feltenetal.,

2021,

2023)orcouldsignificantlyacceleratetheperformanceoftasksineachjob(Eloundouetal.,

2023

;

BriggsandKodnani,

2023).Sodefined,thisconceptpurposelyremainsagnostictothe

potentialforAItoserveaseitherasubstituteorcomplementforhumanlaborinkeytasksandpossiblytoreplaceanoccupationaltogether.Giventhelargedegreeofuncertaintyregardingfutureinnovationsandtheirapplicationtospecificproductiveprocesses,precisepredictionsarechallengingandrequiresignificantcaveats.Nevertheless,itisimportantforacademicsandpolicymakerstoconsidertheconsequencesofAI’sinteractionswitheachoccupation.Forinstance,workersinoccupationsmorevulnerabletosubstitutionbyAIwillbemorelikelytoexperienceadverseincomeshockswhilethoseincomplementedoccupationscouldexperiencehigherreturnstotheirlabor.SuchexercisewouldallowforaninformeddiscussionofhowAImayposegreaterrisksofadverselabormarketoutcomesforsomeworkersandgreateropportunitiesforothers,drawingaggregateimplicationsforitseconomy-wideimpact.

ThispaperthuscontributestothedebateonhowAImayimpactthelabormarketbyproposinganextensiontothewidelyusedAIOccupationalExposure(AIOE)measureby

Feltenetal.

(2021)toaccountforpotentialcomplementarity.

Tothisaim,wefirstbuildanindexofpotentialforAIcomplementarityattheoccupationlevelbasedonthesamedatasourceusedbytheseauthors,theOccupationalInformationNetwork(O*NET)repository.Specifically,wedrawontwoareasofO*NET:workcontextsandoccupations’“jobzones”.Theformercapture“physicalandsocialfactorsthatinfluencethenatureofwork”,andhenceareinformativeofthelikelihoodthatkeyactivitiesofanoccupationwouldbeassignedtoAIwithouthumansupervision-thatis,asasubstitutetolabor.Forinstance,societyispresumablylesslikelytofullydelegatetoAIincontextsinwhichtherearegraveconsequencestoerrors,likepilotinganairplaneordiagnosingdiseases.Meanwhile,jobzonesreflecttheamountofeducationandtrainingrequiredtoperformanoccupation.LongertrainingmayentailgreaterabilitytointegratetheknowledgeneededtooperateAIintotheskillsetofanoccupation,translatingintogreaterpotentialtousethetechnologytosupporthumantasks.

Equippedwiththisindex,wethenconstructacomplementarity-adjustedAIoccu-pationalexposure(C-AIOE)measure,wheretheexposureofoccupationsismitigatedbytheirpotentialforcomplementarity.Inthisalternativemeasure,ahighervalueofexposuremorecloselycorrespondstogreaterriskofsubstitutionandhenceofanadverselabormarketeffectfromAI.Wefindthatsomehigh-skilloccupationalgroupswithhighexposuretoAI,suchasprofessionalsandmanagers,alsoholdthehighestpotentialforcomplementarityandthushavelowC-AIOEvalues.Meanwhile,clericalsupportoccupationsarehighlyexposed

3

buthaveonaveragelowcomplementarity,thereforescoringhighestintheC-AIOEmeasure.

AsecondquestionconcernsthemagnitudeofdisparitiesinAIexposureacrosscoun-triesandwhether,withineachcountry,similarpatternsemergeinhowexposureisdistributedacrossthelaborforce.MostoftheanalysisofexposuresofarhasfocusedonAdvancedEconomies(AEs),withonlylimiteddiscussionofEmergingMarkets(EMs).Thislattergroupofcountries,encompassingawiderangeofdiverseeconomicrealities,ischaracterizedbydistinctlabormarketcompositionswithrespecttooccupationsandworkerdemographics.LabormarketexposuretoAIinEMs,anditsdifferenceswithAEs,hencedeserveadeeperdiscussion.

Thesecondcontributionofthispaperisthustoprovideadetailedcross-countryanalysisofAIexposureusingworker-levelmicrodatafromsixeconomies:twoadvancedeconomies(UKandUS)andfourEMs(Brazil,Colombia,India,SouthAfrica).WecombinemicrodatafromrecentlaborforcesurveyswiththeAIOEandC-AIOEmeasuresataverygranularoccupationallevel(morethan400ISCO-08codes)topaintadetailedpictureofAIexposurebothacrosscountriesandwithineachcountry.Theuseofmicrodataalsoallowsforadeeperanalysisofheterogeneitythroughoutthelabormarketofindividualcountries,basedondemographicgroupsandalongtheincomedistribution,uncoveringsimilaritiesanddifferencesinexposurepatternsinAEsandEMs.

Themainfindingscanbesummarizedasfollows.Therearesubstantialcross-countrydisparitiesinthebaselineAIOE,withEMsgenerallyexhibitinglowerexposurelevelsthanAEs.Thisvariationprimarilyhingesondifferentemploymentcompositions,withAEschar-acterizedbylargerproportionsofhigh-skilloccupationssuchasprofessionalsandmanagers.Inlinewiththefindingsofpreviousstudies,theseprofessionsarethemostexposedtoAI

duetotheirhighconcentrationofcognitive-basedtasks(Feltenetal.,

2021,

2023;

Briggs

andKodnani

,

2023;

Eloundouetal.,

2023)

.However,becausethosehigh-skilloccupationsalsoshowhigherpotentialforAIcomplementarity,thesecross-countrydisparitiesintermsofpotentiallydisruptiveexposurediminishsignificantlyoncecomplementarityisfactoredin.Nevertheless,AEsremainmoreexposedevenundertheC-AIOEmeasure.Meanwhile,EMswithalargeshareofagriculturalemployment,likeIndia,remainrelativelylessexposedunderbothmeasures,asoccupationsinthissectorhaveverylowbaselineexposuretoAI.Overall,theresultssuggestthattheimpactofAIonlabormarketsinAEsmaybemore“polarized,”astheiremploymentstructurebetterpositionsthemtobenefitfromgrowthopportunitiesbutalsomakesthemmorevulnerabletolikelyjobdisplacements.

4

Ouranalysisuncoverswithin-countrydisparitiesinAIexposure,bothadjustedandunadjusted,acrossdemographicvariablessuchasgender,education,andage,amongbothEMsandAEs.Thesepatternsexhibitnotableparallelsacrosscountries.WomenaremoreexposedtoAIthanmeninalmostallcountriesinoursample,primarilyduetotheirpre-dominantemploymentinmiddle-skillserviceandretailoccupations,whichbeararelativelyhigherexposurethanmanuallaborroles.TheonlyexceptionisIndia,wherewomenhavelowerexposurethanmenduetotheirsubstantialemploymentinagriculture.Intermsofeducationalattainment,inbothAEsandEMsworkerswithatleastacollegedegreearemoreexposedthanthosewithlowereducationalcredentials.However,theformeralsocarryagreaterpotentialtobenefitfromAIduetotheirconcentrationinprofessionalandman-agerialjobs.Nocommonresultsemergewithrespecttoage,mostlikelyduetocomplexinteractionswithcountry-specificseculartrendsineducationalattainmentandfemalelaborforceparticipation.

Withrespecttoexposureacrossthedistributionofearnings,asignificantfindingemerges.High-incomeworkersaremoreexposedtoAI.However,consistentwiththeirgener-allyhighereducationalattainment,thisdifferenceismostlyaccountedforbyemploymentinoccupationswithhighpotentialcomplementarity.Meanwhile,employmentinhigh-exposurebutlow-complementarityjobsisevenlydistributedacrossthedistribution.Thisresultsug-geststhatwhilethepotentialadverseimpactmaybemoreevenlyspreadacrosstheincomedistribution,thebenefitsarepredominantlyconcentratedatthetop.

OurpaperrelatestothegrowingnumberofworksontheimpactofAIonlabormarkets.Themajorityofempiricalstudiesfocusindetailonvariationinexposureexclusively

intheUS(Feltenetal.,

2021,

2023;

Eloundouetal.,

2023;

Webb,

2020).2

OECD

(2023),

Albanesietal.

(2023),

BriggsandKodnani

(2023),

Gmyreketal.

(2023)provideacross

-

countryperspective,butonlythelattertwoconsiderexposureinEMs.3

BriggsandKodnani

(2023)conductabroadsectoralanalysisextrapolatingfromcoarseindustry-levelmeasures

ofexposureconstructedfortheUS.

Gmyreketal.

(2023)havealargecoverageofEMs

andlow-incomeeconomiesattheoccupationallevelwithvaryingdegreesofgranularity.Usingmicrodata,ourworkinsteadconductsagranularcomparisonofEMsandAEsbothattheaggregatelevelandwithincountries.WethusdelvedeeperintopatternsofAIexposureacrossdemographicgroupsandtheincomedistribution,providingamorerefined

2Brynjolfssonetal.

(2018)study“automation”oftasksbutfocusonMachineLearning,whichisan

importantbutsmallsubsetAI.

3Copestakeetal.

(2023)areanexampleofanempiricalstudyoftheearlyimpactofAIonasingleEM

economy.

5

identificationofpotential“winners”and“l(fā)osers”inEMs.

Severalstudieshavemademethodologicalcontributionsbydevelopingmeasuresof

occupation-levelexposuretoAI(Feltenetal.,

2023;

Eloundouetal.,

2023;

Webb,

2020;

BriggsandKodnani,

2023)

.ThroughtheO*NETrepository,theseworksconstructmea-suresofexposurethataregenerallyagnosticregardingthelikelihoodofAIcomplementingorsubstitutingforhumanlaborinagiventask,activity,oroccupation.Followingthelong-standingliteratureonroutine-biasedautomation,recentworksmakingadistinctionbetween

complementarityandsubstitutionhaveadoptedatask-basedframework(AcemogluandRe-

strepo

,

2018,

2022;

Autoretal.,

2022;

Gmyreketal.,

2023).Despiteitsrigorousconceptual

-izationoftheinteractionsbetweenhumanandmachineabilities,asacknowledgedby

Autor

(2022),thetaskmodelalsohassomelimitationswhenappliedtoAI.First,asthetechnology

continuestodevelop,itisdifficulttosaywhattasksAIcanandcannotperformfullyunsu-pervised.Second,thisapproachholdsanarrowviewonthefactorsdeterminingwhichjobsareexposedtoreplacementfromAI.RecentstudiesfromtheOECD,basedonsurveysofworkersandfirms,clearlyshowtherichvarietyofconcernsandindividualexperiencesinAI

adoption(Laneetal.,

2023;

Milanez,

2023).Ourcontributionisthustoconstructameasure

ofcomplementaritytoAIbyexaminingabroadsetoffactorsbeyondtasks,relatedtothesocialandphysicalcontextinwhichworkisperformed.Wethusprovideamorenuancedviewofwhichoccupationsandworkersfacethegreatestrisksandopportunitiesintheyearsahead.

Ourmethodologynaturallycarriescaveats.First,theselectionofcontextsfromO*NETreliesonourownjudgementofwhichfactorsmatterfortheinteractionbetweenAIandworkers.However,wepresentasetofteststoshowthatthesecontextsarenotallsystematicallyrelatedtoeachotherandthusofferamultifacetedtakeonpotentialcomple-mentarity,factoringinapluralityofangles.WealsotesttherobustnessoftheC-AIOEtodifferentspecificationsoftheadjustment.Furthermore,weacknowledgethattheimportanceofcomplementarityreliesonsocietalviewsandonotherinnovationstosupportAI.AsAItechnologyimprovesinprecisionandgarnersincreasedtrust,thelikelihoodofitsupplantinghumantasks–eveninoccupationscharacterizedbyhighlevelsofresponsibility,criticality,andskills–maygrow.Consequently,theapplicabilityoftheconceptproposedinthispapercoulddecreaseovertime.Toillustratethispoint,wediscussanexerciseinwhichtheweightgiventocomplementarityintheadjustmentcanbealtered.

Beforeconcludingwealsomakefurtherconsiderationsontheinterpretationoftheresultsandthescopeforfutureanalysis.Forinstance,ourproposedadjustmenttotheAIOE

6

measuredoesnotimplythatworkersinexposedoccupationswithhighcomplementaritydonotfaceanyriskofdisplacement.ComplementaritycanonlybeleveragedifindividualworkerspossesstheskillsneededtotakeadvantageofAIasasupportingtechnology.Withoutsuchabilities,workersinthoseoccupationswouldstillfacereducedemploymentprospectseveniftheoccupationasawholemayexperiencerisingdemand.Moreover,ourapproachonlymeasurescross-countrydifferencesbasedonoccupationalcomposition,abstractingfrommacro-factorssuchastheavailabilityofinfrastructureneededtoimplementAIandthepotentialdifferenceinthetaskcompositionofoccupationsacrosscountries.

Theremainderofthepaperisstructuredasfollows.Section

2

introducestheconceptofcomplementarityandproposesapotentialcomplementarity-adjustedexposuremeasure.Section

3

describesthecountry-specificdatasourcesusedfortheanalysis.Sections

4-5

presentthemainfindingsandthesensitivityanalysis.Section

6

providesfurtherdiscussionoftheresults.Finally,Section

7

concludes.

2AIExposureandAdjustingforPotentialComple-

mentarity

Inthissection,wediscusstheimportanceofaddingthepotentialforcomplementarityorsubstitutabilityasadimensionforunderstandinghowAIexposureattheoccupationallevelcanposebothrisksandopportunities.

2.1Motivation

RecentanalyseshavefocusedtheirattentiononAIexposure.Whileitsprecisedefinitionvariesacrossstudies,exposurereflectsthepotentialforAItobeintegratedintoeachoccupationbasedthetasksandskillsthatcharacterizeeachjob.Giventhehighdegreeofuncertaintyoverthefutureofthisfast-pacingandbroadlyapplicabletechnology,theconceptofexposureispurposelyframedasagnosticonthelikelihoodofAIcomplementingorreplacinglaborintheperformanceofagiventaskoroccupation.Forinstance,theAIOEindexby

Feltenetal.

(2021)measuresthedegreeofoverlapbetweenmainAIapplications

andtheabilitiesneededtoperformanoccupationeffectively.4

4InthecontextofGenerativeAI,

Eloundouetal.

(2023)defineexposure“asameasureofwhether

accessto[LargeLanguageModels]wouldreducethetimerequiredforahumantoperformaspecific[workactivity]orcompleteataskbyatleast50percent.”Meanwhile

Webb

(2020)measuresexposurethroughthe

degreeofsimilaritybetweenthedescribedapplicationsofAIpatentsandthetasksdefininganoccupation.Finally,

BriggsandKodnani

(2023)manuallyidentifyworkactivitiesexposedtoAIandwhether,withinan

7

GivenAI’spotentialtoperformhighlycomplexfunctions,understandinghowitcouldaugmentworkersorreducethedemandfortheirlaborisofgreatimportanceforpolicymakersandresearchersalike.Whilesomestudiesdifferentiatebetweensubstitutionandcomplementarity,theybuildthisdistinctiononatask-basedframework.Forinstance,

Gmyreketal.

(2023)definesoccupationsashavinghigh“automation”or“augmentation”po

-tentialbasedonthedistributionoftheAI-automationscoresoftheindividualtasksdefining

eachoccupation.5

Althoughthisapproachhasmerits,itholdsanarrowfocusincategorizingtheinteractionofhumanworkwithatechnologythatwilllikelyhavecomplexrepercussionsinotherrealms.

Ourproposedframeworkthusconceivescomplementarityasdrivenbyasetoffac-tors–social,legal,technical–thatareindependentofexposureitself.ThisdistinctionisconceptuallyillustratedinFigure

1.Workersinoccupationshighlyexposed,butwhereAI

hasthepotentialtoturnintoasupportingtechnology(upperrightquadrant)aremorelikelytoexperienceproductivitygains,conditionalonaccesstothenecessaryinfrastructureandtheappropriateskillstoengagewiththetechnology.Ontheotherhand,workersinhighlyexposedoccupationswithlowerpotentialforcomplementarity,andthusahigherriskofsub-stitution(lowerrightquadrant),mayexperiencealong-lastingfallindemandfortheirlaboralongthelinesofthenegativeshockinflictedbythepastwaveofroutine-biasedautomation,

withreducedemploymentopportunitiesandlowerearnings(AutorandDorn,

2013)

.

AtlowerlevelsofAIexposure(leftquadrants),ahighercomplementaritypotentialmaystillaffecthowAIisintegratedintoeachoccupationbut,giventhelowerscopeforinteractionwithhumanskillsandtasks,itwouldlikelybelessinfluentialforlabordemand.Inthissense,theimportanceofpotentialcomplementarityisconditionalonagivenexposurelevel.

Itisalsoworthnotingthat,whilelowercomplementarityreflectsariskoflowerlabordemandforworkersinagivenoccupation,highercomplementaritydoesnotinitselfsignifynorisksforindividualworkers.ThoseemployedinahighlycomplementaryoccupationwhodonotpossesstheskillsneededtoengagewithAIwouldlikelyfaceloweremploymentopportunitiesandwages.

occupation,suchactivitiesareofalow-enoughlevelofcomplexitythatAIcouldcompletethem.Arguably,thislastmethodologyimpliesaviewonexposurethatisclosertolaborsubstitution.

5Moreprecisely,occupationswherethemeantask-levelautomatabilityscoreishighandthestandarddeviationislowaredefinedasautomatable.Occupationswithalowmeanscoreandhighstandarddeviationaredefinedasaugmentable.

8

Withthesecaveatsinmind,weproposeasimpleadjustmentofAIexposuremeasurestoaccountforcomplementarity.Inwhatfollows,weusetheAIOEindexby

Feltenetal.

(2021)asthebaselinemeasuretoaugmentintoacomplementarity-adjustedAIOE(C-AIOE)

.However,thesameapproachcouldbeappliedtoanymeasurethatdoesnotalreadycapturecomplementarity.

Foragivenoccupationi,letθibeameasureofpotentialcomplementarityofAI.Thebaselineexposurecanbeadjustedasfollows:

C-AIOEi=AIOEi*(1?(θi?θMIN),,(1)

whereθMINistheminimumvalueofθiacrossalloccupations.WeadjustforθMINtoallowthecomplementariymeasuretohavearelativeinterpretationastheoriginalAIOEindex.Thesecondtermontheright-handsidethusrepresentsadownwardadjustmentofAIOErelativetotheoccupationwiththelowestpotentialcomplementarity(θMIN),forwhichtheAIOEandC-AIOEmeasuresconicide.HenceahighervalueoftheC-AIOEindeximpliesagreaterriskofreplacementattheoccupationlevel.

Figure1:AIexposureandComplementarityDiagram

Complementarity

人

LowExposureHighExposure

HighComplementarityHighComplementarity

Exposure

HighExposure

LowComplementarity

LowExposure

LowComplementarity

ItshouldbenotedthattheoriginalAIOEindexby

Feltenetal.

(2021)isameasure