生成式人工智能的崛起:对美国劳动力市场的暴露、替代和不平等效应进行建模_第1页
生成式人工智能的崛起:对美国劳动力市场的暴露、替代和不平等效应进行建模_第2页
生成式人工智能的崛起:对美国劳动力市场的暴露、替代和不平等效应进行建模_第3页
生成式人工智能的崛起:对美国劳动力市场的暴露、替代和不平等效应进行建模_第4页
生成式人工智能的崛起:对美国劳动力市场的暴露、替代和不平等效应进行建模_第5页
已阅读5页,还剩66页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

BISWorkingPapersNo1207

TheriseofgenerativeAI:modellingexposure,

substitution,andinequalityeffectsontheUSlabour

market

byRaphaelAuer,DavidKöpfer,JosefŠvéda

MonetaryandEconomicDepartment

September2024

JELclassification:E24,E51,G21,G28,J23,J24,M48,O30,O33

Keywords:Labourmarket,Artificialintelligence,Employment,Inequality,Automation,ChatGPT,GPT,LLM,Wage,Technology

BISWorkingPapersarewrittenbymembersoftheMonetaryandEconomicDepartmentoftheBankforInternationalSettlements,andfromtimetotimebyothereconomists,andarepublishedbytheBank.Thepapersareonsubjectsoftopicalinterestandaretechnicalincharacter.TheviewsexpressedinthemarethoseoftheirauthorsandnotnecessarilytheviewsoftheBIS.

ThispublicationisavailableontheBISwebsite

()

.

©BankforInternationalSettlements2024.Allrightsreserved.Briefexcerptsmaybereproducedortranslatedprovidedthesourceisstated.

ISSN1020-0959(print)ISSN1682-7678(online)

1

TheriseofgenerativeAI:modellingexposure,substitution,and

inequalityeffectsontheUSlabourmarket*

RaphaelAuertDavidK¨opfer‡Josefv´eda§

August21,2024

Abstract

Howexposedisthelabourmarkettoever-advancingAIcapabilities,towhatextentdoesthissubstitutehumanlabour,andhowwillitaffectinequality?Weaddressthesequestionsinasimulationof711USoccupationsclassifiedbytheimportanceandlevelofcognitiveskills.WebaseoursimulationsonthenotionthatAIcanonlyperformskillsthatarewithinitscapabilitiesandinvolvecomputerinteraction.AtlowAIcapabilities,7%ofskillsareexposedtoAIuniformlyacrossthewagespectrum.AtmoderateandhighAIcapabilities,17%and36%ofskillsareexposedonaverage,andupto45%inthehighestwagequartile.Examiningcomplementaryversussubstitution,wemodeltheimpactonsideversuscoreoccupationalskills.Forexample,AIcapableofbookkeepinghelpsdoctorswithadministrativework,freeinguptimeformedicalexaminations,butrisksthejobsofbookkeepers.WefindthatlowAIcapabilitiescomplementallworkers,assideskillsaresimplerthancoreskills.However,asAIcapabilitiesadvance,coreskillsinlower-wagejobsbecomeexposed,threateningsubstitutionandincreasedinequality.IncontrasttotheintuitivenotionthattheriseofAImayharmwhite-collarworkers,wefindthatthoseremainsafelongerastheircoreskillsarehardtoautomate.

JELcodes:E24,E51,G21,G28,J23,J24,M48,O30,O33

Keywords:Labourmarket,Artificialintelligence,Employment,Inequality,Automation,ChatGPT,GPT,LLM,Wage,Technology

*WethankRyanBanerjee,SebastianDoerr,FiorellaDeFiore,FernandoPerez-Cruz,AndrasValko,andseminarparticipantsattheBISforcommentsandsuggestions.WeacknowledgetheuseofGPT4foreditingand

asitanel.lti,erlcieoftheBIS.

‡BankforInternationalSettlements,david.koepfer@§BankforInternationalSettlements,josef.sveda@

2

1Introduction

HowwilltheadvancementofgenerativeAIcomplementandsubstitutedifferentkindsofhumanlabour?RecentbreakthroughshaveenabledgenerativeAItomimichumancognitiveabilitiesinmanyfields,includingin“whitecollar”professionssuchaslaw,medicine,orscience.Ongoingadvancesandintegrationofthetechnologyintoday-to-dayapplicationsandworkflowsraiseurgentpolicyquestions.

UnderstandinghowthepotentialevolutionofAIwillcomplementorsubstitutehumanskillsisessentialforshapingpoliciestoensureequitablegrowthandemploymentstability.Theliter-

aturehasfocusedontheoccupation-levelimpactofcurrentAImodels,1

experimentalevidence

ofproductivityimpacts(Noy&Zhang,

2023;

Brynjolfssonetal.,

2023;

Pengetal.,

2023),and

thepotentialforcomplementarityandsubstitutioneffectsofAItechnologyataparticularstate

ofAIdevelopment(Pizzinellietal.,

2023;

Acemoglu&Restrepo,

2019,

2018c,a)

.Exceptforcertaintypesoffreelancers(seee.g.

Webb

2020),thebroaderimpactofAIcapabilitiesonthe

labourmarketyetremainstobedemonstrated.

Inthispaper,wetakeaforward-lookingapproach:weaskthequestionof“whatif”and

examinehowanAIofahypotheticallevelofcapabilitieswastoexpose2

differentoccupations.Toshedthefirstlightonthefutureimpact,webuildaparsimoniousbottom-upquantificationwithaspecialfocusonincomedistribution.

Ouranalysisproceedsintwosteps.Inthefirststep,webuildon

Eloundouetal.

(2023);

Feltenetal.

(2021);

Gmyreketal.

(2023);

Pizzinellietal.

(2023);

Acemoglu

(2024)andmodel

theexposuretothetechnologyasthecapabilitiesofAIincrease

.3

Inthesecondstep,weexaminehowthesedevelopmentscouldcomplementorsubstitutehumanlabourthroughthelensoftheirimpactoncoreandsideskills.

Inthefirststep,wearguethatthenear-termimpactofAIislimiteda)tocomputer-relatedinteractionsandb)bythedifficultyoftheskillsthatAIcansubstitutefor.Inthis,weonlyquantifytheimpactonskillsinvolvinginteractionwithacomputer.We,hence,donottakeintoaccounttheimpactofAIonroboticsthatmaysubstituteforphysicalworkorevensocialinteractions

.4

OurfirstdeparturefromtheliteratureistoemployanunderusedpartoftheO*NETdatabasethatclassifiesskillsbytheirdifficulty.Intuitively,anAIofacertaincapabilitylevelcanonlyperformtasksuptoacorrespondingskilllevel.AsthecapabilitiesofAIadvance,an

1Seei.e

.Webb

(2020);

Feltenetal.

(2021);

Tolanetal.

(2021);

Gmyreketal.

(2023);

Yang

(2022)

2Throughoutthepaper,weusetheterms“expose”and“exposure”inaneutralmanner,toimplythatsomepartsofaskill,task,oroccupationcouldbeenhanced,performed,orotherwisebeaffectedbyanAI.

3Similartotheseapproaches,wetakeapartialequilibriumperspectiveanddonottakeintoaccountthe

interplaybetweenskills,relativewages,humancapitalformationanddirectedtechnologicalchange(Acemoglu

&Restrepo,

2018c)

.

4Thisisinlinewith

Acemoglu

(2024),whoarguesthat“AIisnowhereclosetobeingabletoperformmost

manualorsocialtasks”,andwethusassumethatitcanonlyperformcomputerinteractions.

3

increasingshareofcognitiveskillswillhencebeexposedtothetechnology.5

WenestthisnotionofAIcapabilityandskilldifficultyinaquantitativesimulationof711USoccupationsfromtheO*NETdatabaseclassifiedbytheimportanceandtherequiredlevelofcognitiveskillsthatinvolvecomputerinteractions.ThemodelpredictsthatanAIcapableofsubstitutingforsimplecognitivetasks–suchastheminimalcommunicationskillsrequiredforatruckdriver–willexposearound7%ofallskills.AtlowlevelsofAIcapability,thiseffectholdsuniformlyacrosstheentirewagespectrum,butforheterogeneousreasons.Forlow-incomeworkers,asubstantialshareofcognitivecomputerskillsisexposed,buttheoverallshareoftimespentoncomputerinteractionislow.Forhigh-incomeworkers,onlyasmallshareofcognitivecomputerskillsisexposedbecauseofthelargerskillrequirement.However,theshareoftimespentusingsuchskillsishigher

.6

AsAIcapabilitiesincrease,weobserveaprofounddifferenceinoccupationalexposure:upto45%intheupperquartileofthewagedistributionareexposed,whereastheexposureofthelowestquartileisaround26%.

Whatdoesthismeanfortheincomedistribution?Wenotethatinlinewiththeliterature,“exposure”hasaneutralmeaninginthatsomepartsofaskill,task,orjobcouldbeperformedbyanAI.

Thismayleadtosubstitutionbutcouldalsocomplementviaincreasedproductivity.7

Toshedlightontheseissues,intheseconddeparturefromtheliteratureandstepofoursimulations,weexaminetheextenttowhichAImightcomplementorsubstitutehumanlabour.Wefocusonthedifferentialimpactoncoreversussideoccupationalskills,arguingthatAIwouldtendtocomplementoccupationswherevertheauxiliary(side)skillsnecessaryfortheprofessionarewithinitscapabilities.Forexample,ifAIcanorganisemeetings,billing,orbookkeepingforlawyers,medicaldoctors,orscientists,thisfreesuptimethatcanbespentoncoreactivitiesandthusincreasesproductivity.However,aprofessionmaybeatriskifthecoreactivityitselfcanbeperformedbytheAI.

ThisexercisesuggeststhatAImayinitiallycomplementallprofessions,assideskillsare

5Wetakenopositiononhowfasttheevolutionofthetechnologywillmaterialise.SomehavearguedthatAImaysoonhavedramaticimpactsonthelabourmarket(ie

Korinek&Juelfs

(2022))

.OthersarguethatfutureadvancementofAImaymaterialisemuchslowerthanexpected.Forexample,

Acemoglu

(2024)arguesthat

earlyevidenceisfromeasy-to-learntaskswithclearoutcomes(thatAIcanoptimisefor),whereasmoreprofoundproductivityimpactsinmoresubtlecontextsmaymaterialisemuchslower.

Perez-Cruz&Shin

(2024)arguethat

currentLLMsarelimitedintheirunderstandingofhumaninteractionandhigher-orderbeliefs.

6Fortheseexamples,“simplecognitivetasks”correspondtothoserequiringaskilllevelof2.0intheO*NETdatabase,forexample,theminimumsocialperceptivenessskillsrequiredforpiledriversortheminimumspeakingskillsrequiredforindustrialtruckoperators.“Mediumcognitivetasks”correspondtothoserequiringaskilllevelof3.0,forexample,problem-solvingskillsofmedicalappliancetechniciansortheoperationsmonitoringskillsofregisterednurses.“Highcognitivetasks”correspondtothoserequiringaskilllevelof4.0,forexample,thepersuasionskillsofpsychiatristsortheactivelisteningskillsofairtrafficcontrollers.

7Svanbergetal.

(2024)furthernotethat“exposure”doesnotmeanautomation:theysurveyworkerswith

“end-use”taskstogetasenseoftherequirementsforautomation,andsecond,theymodelthecostofamodelcapableofmeetingtherequirements.Focusingontheautomatabilityofvision,findthatonly23%ofoccupationsthatare“exposed”inthesenseof

Eloundouetal.

(2023);

Feltenetal.

(2021)couldtodaybeautomatedeco

-nomically.Wenotethatourmeasureofexposureismorenuancedthantheonein

Eloundouetal.

(2023);

Felten

etal.

(2021)aswerestricttheimpacttoskillsinvolvingcomputerinteractionandnotonlymodelwhetheraskill

inprinciplecouldbeautomatedbutalsowhetherthecapabilityleveloftheAIissufficientforsuchautomation.

4

generallylessdifficultthancoreskills.Forexample,anAIonlycapableofperformingsimplecognitivetaskshasnegligibleexposuretocoreskills,whereasit,onaverage,exposesaround12%ofsideskills.However,alreadyformoderateAIcapabilities,thereisdivergenceacrossthewagespectrum,withthecorecognitiveskillsofthelow-wageworkersbecomingroughlyasexposedtoAIastheirsideskills.Incontrast,theupperquartileofthewagedistributionstillseesnegligibleexposureofcoreskills(5%),whereassideskillsareexposedsubstantially(27%).

IfAIcapabilitiesarehigh,around25%ofbothsideandcoreskillsofthelowestquartileofthewagedistributionareexposed.Incontrast,only20%ofthecorebutastaggering62%ofthesideskillsofthehighestquartileoftheincomedistributionbecomeexposed.

Onbalance,ourmodellingoftheimpactonsideandcoreskillshencereversesthenotionthat

generativeAImightdecreaseinequalityinthelabourmarket(Noy&Zhang,

2023;

Brynjolfsson

etal.,

2023)

.Despitebeingatechnologythatisexposingwhite-collarjobsmoreintensively,thiseffectisfocusedonthesideskillsoftheirprofessions,whilethecoreskillsarenotinreach

.8

Incontrast,acapableAIwillalsoexposethecoreskillsoflower-incomeworkers,thusthreateningsubstitutionandwideninginequality.

Thebalanceofthispaperisasfollows:werelateourapproachtotheliteratureinSection

Section2.

Next,

Section3

presentsthemethodologydescribingtheevolutionaryimpactofever-improvingAIonoccupations.ItalsoservesasanAIexposuredependentonAI’scapabilities.Thereafter,wesplittheAIexposurebasedoncoreandsideskills

Section4

thatarethenusedtoidentifycomplementarityandsubstitutionaleffectsforindividualoccupations.

Section5

presentsadditionalrobustnessanalysis,while

Section6

concludes.

2Literaturereview

Historically,technologicaladvancementshavebeenmetwithbothoptimismandconcernre-

gardingtheirimplicationsforthelabourmarket(Bessen,

2016)

.TheadventofAIandmachinelearningtechnologies,ingeneral,hasintensifiedthesedebates,withresearchersseekingtoun-derstandhowthesenewtoolscanreshapethelabourmarketandhowtheimpactcandiffer

fromprevioustechnologicaladvancementsinrobotisationorcomputerisation(Autor,

2015)

.

SeveralrecentstudieshavedirectlyaddressedthepotentialofthelatestadvancementsinAItosignificantlyimpactthecurrentstructureofthelabourmarket.

Brynjolfssonetal.

(2018)

arguethatmostoccupationsintheUSincludeatleastsometasksthataresuitableformachinelearningapplications,and

Eloundouetal.

(2023)suggeststhat80%oftheworkforcecouldbe

affectedbyGenerativePredictiveTransformers(GPTs).Whiletheseestimatesarestaggering,

Arntzetal.

(2016)arguethattheactualvulnerabilityofjobstoautomationislowerwhen

consideringthenuancedskillswithinoccupations.Nonetheless,theproliferationofthelatestLLMsseemstobenon-negligent;

Eloundouetal.

(2023)furtherfind19%ofUSworkersinthe

8Ofcourse,oncethecapabilityoftheAIbecomesextremelyhighsuchthatallskillsarewithinreach,thiseffectabates,andallcognitiveworkersareindangerofreplacement.

5

USmayseeatleasthalfoftheirskillsimpactedand

Hatziusetal.

(2023)finds25%ofcurrent

workskillsinUSautomatable.

CurrentAIcapabilities,insomeinstances,fallshortofprofoundreasoningskills(Perez-Cruz

&Shin,

2024)

.However,animportantissueregardshowthefutureevolutionofAIcapabilitiescanenhancelabourproductivityorcrowdoutworkers.RecentexperimentswiththelatestgenerationofAIshowthatitcanhaveapositiveeffectinspecificoccupationswhilereducingdifferencesamongworkerswithvaryingexperiencelevels.

Noy&Zhang

(2023)demonstrated

thattheuseofChatGPTsignificantlyincreasesaverageproductivitymeasuredbytimespentontasksandreducesdifferencesbetweenhigh-andlow-skilledworkers.

Brynjolfssonetal.

(2023)studiedtheintroductionofgenAIassistanttothecustomersupportagentsandfounda

significantlyhighernumberofcompletedtasksthatweremorepronouncedfornoviceandlow-skilledworkers.

Pengetal.

(2023)suggestscoderswithaccesstogenAIarecapableofcompleting

coding-orientedtasksupto55%faster.AItoolscanalsoserveasthetooltodiscoverpotential

improvementsinbusinesssystems(Cockburnetal.,

2018;

Chengetal.,

2022)

.

However,anincreaseinlabourproductivitymeansthatlesshumancapitalisneededto

maintainthesameoutput,whichcouldleadtolayoffsorwagereductions(Acemoglu&Restrepo,

2020)

.Inthiscontext,

Frey&Osborne

(2017)predictedthatupto47%ofUSemploymentis

athighriskofcomputerisation.

Arntzetal.

(2016)howeverusesadifferentmethodologyand

estimatesanimpactofonly9%.Gmyreketal.

(2023)findthatgenAIcouldautomate5.1%of

totalemploymentinhigh-incomecountries,whereaslow-incomecountriesarenotsosusceptible.Thepotentialforaugmentationissimilarlydistributedacrosscountriesrelativetotheirincomelevels,althoughthepotentialtoaugmentismuchlarger(aroundfourtofivetimes).

Noy

&Zhang

(2023)claimthatChatGPTmostlysubstitutesforworkereffortratherthanpurely

complementingworkerskills.

Yang

(2022)alsoshowsthatAIcanpositivelyaffectproductivity

andemploymentbutadverselyaffectstheemploymentoflessknowledgeableworkers.Some

studiesadditionallydebatetheeffectsrelativetogender(Eloundouetal.,

2023;

Webb,

2020;

Gmyreketal.,

2023;

Aldasoroetal.,

2024)

.

Historicalexperiencewithinnovationshowsthatinthelong-term,thedisplacementcanbe

offsetbyanincreaseintherangeofgoodsandservicesoffered,see(Autor,

2015;

Acemoglu

&Restrepo,

2019)

.Forexample,

Bessen

(2016)showsUSlabourdemandhasincreasedfaster

incomputerisedoccupationssince1980,althoughthecomputerisationledtosubstitutionforotheroccupations,shiftingemploymentandrequiringnewskills.

Acemogluetal.

(2022)find

increasingdemandinAI-exposedoccupationsintheUSsince2015.AutomatisationinJapan

andtheUSgeneratedcostsavings,allowinglargeroutputineconomy(Adachietal.,

2024;

Dekle,

2020;

Acemoglu&Restrepo,

2020)thatoutweighedthedisplacementeffectsofhuman

labour.

Yang

(2022)findsthatAItechnologyispositivelyassociatedwithproductivityand

employmentinTaiwan’selectronicsindustryforthe2002–2018period.

Acemoglu&Restrepo

(2019),

Acemoglu&Restrepo

(2018a)and

Acemoglu&Restrepo

(2018c)thenfocusdirectlyon

thedynamicsofdisplacementandreinstatementoflabourduetoautomation.Basedondata

6

fromtheUSsinceWorldWarII,

Acemoglu&Restrepo

(2019)claimthatdisplacementeffects

occurintuitively,buttheyarecounterbalancedbythecreationofnewtasksinwhichlabourhasacomparativeadvantage.Thesethenchangethetaskcontentofproductioninfavouroflabourbecauseofareinstatementeffectfollowedbyariseinthelabourshareandlabourdemand.

Acemoglu&Restrepo

(2019)pointoutthatthesuccessofreinstatementisnotautomatic

.Itratherdependsonadditionalvariablessuchasthesupplyofnewskills,demographics,orlabourmarketinstitutions

.9

Althoughpreviousinnovationsinautomatisationandcomputerisation,onaverage,broughteconomicgrowth,theystillreshapedthelabourmarketandintroducednewchallengesinre-gionallabourmarketstructuresthataffectedlabourdistributionacrosstheskilldistributionofmarkets.

Autor

(2019)documentstheseeffectsusingUSdatashowingthatautomation(to

-getherwithinternationaltrade)ledtotheeliminationofthebulkofnon-collegeoccupations,furtherleadingtodisproportionatepolarisationofurbanlabourmarkets.

Acemoglu&Restrepo

(2022)documentthatbetween50%and70%ofchangesintheUSwagestructureoverthelast

fourdecadesareaccountedforbyworkersspecialisedinroutinetasksinindustriesexperiencingrapidautomation.

Acemoglu&Restrepo

(2020)showindustrialrobotadoptionintheUnited

Stateswasnegativelycorrelatedwithemploymentandwages.Theseexamplespinpointtheimportanceofunderstandingthepotentialeffectsoftechnologicaladvancementstonavigateasmoothtransitiontowardsanewstructureofthelabourmarket.

ThequestionremainshowmuchthenewwaveofautomationwithAIiscomparabletoprevi-oustechnologicaladvancements.Previously,automationexposedpredominantlymanuallabourthroughtheinventionofmachinesandrobots.Thetransitionprocesstorobot-drivenproduc-

tion,therefore,affectedatitsfirststageratherlower-skilledlabour(Acemoglu&Restrepo,

2018b)

.EvolvingAIchallenges,however,cognitivetasksandskillsandcreatesapotentialtoaffectdifferentoccupationsbyeithercomplementingorsubstitutingthem.Earlierworkby

Autor&Dorn

(2013)suggeststhatlow-wageoccupationsfacedhighersubstitutiondueto

computerisation.Incontrast,high-wageoccupationswerecomplementedbytechnology.

Webb

(2020)thenfocusesonthenewerinnovationinAIandstatesitisdirectedathigh-skilledtasks,

effectivelyaffectingthehigher-wagequantiles.Asimilarconclusionisreachedby

Eloundouetal.

(2023)and

Pizzinellietal.

(2023)

.

Webb

(2020)arguesthattheimpactofAIisdifferentfrom

theeffectsofsoftwareinnovation,whichexposedmid-wageoccupations(inlinewith

Michaels

etal.

(2014))

.

Pizzinellietal.

(2023)emphasisehighcomplementarityintheuppertailofthe

earningsdistributionbyAI,leadingtoaproductivityboostinsteadofjobdisplacements.TheeffectsofAIalsodiffergeographically.

Pizzinellietal.

(2023);

Gmyreketal.

(2023);

Albanesi

etal.

(2023)showthatmoredevelopedcountriesaremoreexposedtoAIastheirlabourmarkets

aremoreorientedtocognitivetasks.However,asAIsignificantlyprogresses,researchalsoneeds

toaccountfortheevolutionoftechnologytofullyunderstanditspotentialeffects.Examining

9Inasimilarvein,

Aldasoroetal.

(2024)showinageneralequilibriummodelthattheoutputeffectsofAI

mayprimarilyariseviatheindirectimpactondemandandassociatedchangesinrelativepricesratherthanviathedirectinitialproductivityboostfromAIadoption.

7

theimpactofdevelopingAIthroughthelensofwagedistributionseemstobeadvantageousto

formulatetargetedpolicyresponses(Furman&Seamans,

2019)

.AstheadvancementsinAItechnologyprogress,theirinteractionmightchangerapidly.

3MeasuringAIexposure:dataandmethodology

PredictingtheimpactofAIonthelabourmarketischallenging,astheintegrationofthetechnologyintoreal-lifeapplicationsisstillinitsinfancy,andonlysomesyntheticbenchmarksonthepotentialqualityandefficiencyimprovementsoncertainaspectsofworkareavailable(seei.e.

Tolanetal.

(2021);

Pengetal.

(2023);

Noy&Zhang

(2023))

.Particularly,therapidlyevolvingcapabilitiesofAIareamajorsourceofuncertainty.Inthefaceoftheseuncertainties,weconstructaparsimoniousbottom-upmodelcentredonan“AIcapability”parameter,whichallowsustosimulatetheeffectsofevolvingAI.Themodelisbuiltontheskillandoccupationlevelandlateraggregatedtotheindustryorwage-quantilelevel.

Inthissection,weshowhowweconstructtheAIShareAutomatability(AISA)IndexthatdependsonthesophisticationoftheAI(definedas“AIcapability”above).Thisindexrestsontwomainassumptions:

1.Intheshorttomediumterm,automationwillaffectoccupationalactivitieswithcomputerinteractionasopposedtosocialinteractionsorphysicallabour.

2.Theskillsrequiredforperformingtheoccupationsareheterogeneousintheirdifficultylevel.Foraskilltobeimpactedinacertainoccupation,itsdifficultylevelneedstobewithinthecapabilitiesoftheAI.

WeutilisedatafromO*NETversion27.2andthe2022OccupationalEmploymentandWageStatistics(OEWS)SurveyfromtheUSBureauofLaborStatistics.Thesedatasetsdetailaround800differentoccupations(ofwhichwecanuse711afterjoiningacrosstheskillstablesandemploymentstatistics)across22industries,providingaverageincome,employmentnumbers,andratingsforupto35cognitiveskillsforeachoccupationintermsofrequiredskilllevel(1-6)andimportance(1-5).

Furthermore,thedataincludesdetailedtaskdescriptions10

foreachoccupation(onaverage,wehave24taskdescriptionsforeachofthe711occupations).

Inthedescriptionofourmodel,wewillusesubscriptstodenotethedifferentlevelsofaggregation:thelowestlevelsfortheskill,ofortheoccupationandthehighestaggregationlevelsifortheindustryorwforthewagequantile.TheskilllevelLo,sisdistinctforagivenoccupationoandskills.Forinstance,theoccupationofBiophysicistsrequiresalevelof4.75intheskillmathematics,whiletheimportanceofthisskillIo,sis3.88.

10/dictionary/21.0/text/task_statements.html

(releasenumber21.0)

8

3.1OnlycomputerinteractionisautomatablewithAI

Inthispaper,weonlyexaminetheimpactofAIonautomatingtasksthatrequireskillsinvolvingcomputerinteraction.Jobsperformedoncomputersare,intheshortandmediumrun,muchmorelikelytoincorporateAIapplicationscomparedtothoseinvolvingphysicallabour.Weacknowledgethatalsophysicallabourmay,inthefuture,bepronetoautomationthroughimprovedmachinesandrobotics.However,modellingtheimpactofsuchdevelopmentsisoutofthescopeoftheanalysisathand.Similarly,weexpectsocialinteractiontorequirehigherdegreesofsocialacceptancebeforewidespreadautomationmaterialises.Certainly,cost-effectivenessandimprovedsocialskillsoftheAIwillspeeduptheprocess,yet,asforphysicallabour,weexpectlongertimescales.

Weconstructameasureoftheshareofthetimespentoncomputerinteractionsbasedonabout19,000detailedtaskdescriptionsavailableintheO*NETdatabase.Basedonthede-scriptionsofeachoccupation,weinstructedGPT-4toestimatethetimespentwithi)computerinteraction,ii)socialinteraction,andiii)physicallabour.Theexactpromptisshowninthe

BoxA1,andoneexampleoftaskdescriptionisprovidedtotheChatGPT-4in

TableA1.

Notethatcomputerinteractionrepresentsworkingonacomputerthatcommonlydoesnotincludecommunicationviae-meetingsorothersimilarsocialinteraction.

Ingeneral,ChatGPT-4provesveryhighcomparabilitywithconventionalhuman-basedpro-ceduresforcategorisationpurposes.

Eloundouetal.

(2023)usesbothapproaches(human

-andGPT4-based)todirectlyidentifyoccupationalAIexposure,findingaveryhighcorrelationbe-tweenhumanassessmentsandGPT4-basedself-assessments

.11

Gmyreketal.

(2023)follows

theirapproachemp

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论