THEFUTUREOFEMPLOYMENT:HOWSUSCEPTIBLEAREJOBSTOCOMPUTERISATIONCarlBenediktFreyandMichaelA.OsborneSeptember17,2013.AbstractWeexaminehowsusceptiblejobsaretocomputerisation.Toas-sessthis,webeginbyimplementinganovelmethodologytoestimatetheprobabilityofcomputerisationfor702detailedoccupations,usingaGaussianprocessclassier.Basedontheseestimates,weexamineex-pectedimpactsoffuturecomputerisationonUSlabourmarketoutcomes,withtheprimaryobjectiveofanalysingthenumberofjobsatriskandtherelationshipbetweenanoccupation’sprobabilityofcomputerisation,wagesandeducationalattainment.Accordingtoourestimates,about47percentoftotalUSemploymentisatrisk.Wefurtherprovideevidencethatwagesandeducationalattainmentexhibitastrongnegativerelation-shipwithanoccupation’sprobabilityofcomputerisation.Keywords:OccupationalChoice,TechnologicalChange,WageInequal-ity,Employment,SkillDemandJELClassication:E24,J24,J31,J62,O33.tinProgrammeontheImpactsofFutureTechnologyforhostingthe“MachinesandEmploy-ment”Workshop.WeareindebtedtoStuartArmstrong,NickBostrom,ErisChinellato,MarkCummins,DanielDewey,DavidDorn,AlexFlint,ClaudiaGoldin,JohnMuellbauer,VincentMueller,PaulNewman,Seánhigeartaigh,AndersSandberg,MurrayShanahan,andKeithWoolcockfortheirexcellentsuggestions.carl.frey@oxfordmartin.ox.ac.uk.dom,mosb@robots.ox.ac.uk.I.INTRODUCTIONInthispaper,weaddressthequestion:howsusceptiblearejobstocomputerisa-tionDoingso,webuildontheexistingliteratureintwoways.First,drawinguponrecentadvancesinMachineLearning(ML)andMobileRobotics(MR),wedevelopanovelmethodologytocategoriseoccupationsaccordingtotheirsusceptibilitytocomputerisation.1Second,weimplementthismethodologytoestimatetheprobabilityofcomputerisationfor702detailedoccupations,andexamineexpectedimpactsoffuturecomputerisationonUSlabourmarketout-comes.OurpaperismotivatedbyJohnMaynardKeynes’sfrequentlycitedpre-dictionofwidespreadtechnologicalunemployment“duetoourdiscoveryofmeansofeconomisingtheuseoflabouroutrunningthepaceatwhichwecanndnewusesforlabour”(Keynes,1933,p.3).Indeed,overthepastdecades,computershavesubstitutedforanumberofjobs,includingthefunc-tionsofbookkeepers,cashiersandtelephoneoperators(Bresnahan,1999;MGI,2013).Morerecently,thepoorperformanceoflabourmarketsacrossadvancedeconomieshasintensiedthedebateabouttechnologicalunemploymentamongeconomists.Whilethereisongoingdisagreementaboutthedrivingforcesbehindthepersistentlyhighunemploymentrates,anumberofscholarshavepointedatcomputer-controlledequipmentasapossibleexplanationforrecentjoblessgrowth(see,forexample,BrynjolfssonandMcAfee,2011).2Theimpactofcomputerisationonlabourmarketoutcomesiswell-establishedintheliterature,documentingthedeclineofemploymentinroutineintensiveoccupations–i.e.occupationsmainlyconsistingoftasksfollowingwell-denedproceduresthatcaneasilybeperformedbysophisticatedalgorithms.Forexam-ple,studiesbyCharles,etal.(2013)andJaimovichandSiu(2012)emphasisethattheongoingdeclineinmanufacturingemploymentandthedisappearanceofotherroutinejobsiscausingthecurrentlowratesofemployment.3Inad-1Werefertocomputerisationasjobautomationbymeansofcomputer-controlledequip-ment.2ThisviewndssupportinarecentsurveybytheMcKinseyGlobalInstitute(MGI),showingthat44percentofrmswhichreducedtheirheadcountsincethenancialcrisisof2008haddonesobymeansofautomation(MGI,2011).3Becausethecorejobtasksofmanufacturingoccupationsfollowwell-denedrepetitiveprocedures,theycaneasilybecodiedincomputersoftwareandthusperformedbycomputers(AcemogluandAutor,2011).ditiontothecomputerisationofroutinemanufacturingtasks,AutorandDorn(2013)documentastructuralshiftinthelabourmarket,withworkersreallo-catingtheirlaboursupplyfrommiddle-incomemanufacturingtolow-incomeserviceoccupations.Arguably,thisisbecausethemanualtasksofserviceoccu-pationsarelesssusceptibletocomputerisation,astheyrequireahigherdegreeofexibilityandphysicaladaptability(Autor,etal.,2003;GoosandManning,2007;AutorandDorn,2013).Atthesametime,withfallingpricesofcomputing,problem-solvingskillsarebecomingrelativelyproductive,explainingthesubstantialemploymentgrowthinoccupationsinvolvingcognitivetaskswhereskilledlabourhasacomparativeadvantage,aswellasthepersistentincreaseinreturnstoeducation(KatzandMurphy,1992;Acemoglu,2002;AutorandDorn,2013).Thetitle“LousyandLovelyJobs”,ofrecentworkbyGoosandManning(2007),thuscapturestheessenceofthecurrenttrendtowardslabourmarketpolarization,withgrowingemploymentinhigh-incomecognitivejobsandlow-incomemanualoccupa-tions,accompaniedbyahollowing-outofmiddle-incomeroutinejobs.AccordingtoBrynjolfssonandMcAfee(2011),thepaceoftechnologi-calinnovationisstillincreasing,withmoresophisticatedsoftwaretechnolo-giesdisruptinglabourmarketsbymakingworkersredundant.Whatisstrikingabouttheexamplesintheirbookisthatcomputerisationisnolongerconnedtoroutinemanufacturingtasks.Theautonomousdriverlesscars,developedbyGoogle,provideoneexampleofhowmanualtasksintransportandlogisticsmaysoonbeautomated.Inthesection“InDomainAfterDomain,Comput-ersRaceAhead”,theyemphasisehowfastmovingthesedevelopmentshavebeen.Lessthantenyearsago,inthechapter“WhyPeopleStillMatter”,LevyandMurnane(2004)pointedatthedifcultiesofreplicatinghumanperception,assertingthatdrivingintrafcisinsusceptibletoautomation:“Butexecut-ingaleftturnagainstoncomingtrafcinvolvessomanyfactorsthatitishardtoimaginediscoveringthesetofrulesthatcanreplicateadriver’sbehaviour[...]”.Sixyearslater,inOctober2010,Googleannouncedthatithadmodi-2011).Toourknowledge,nostudyhasyetquantiedwhatrecenttechnologicalprogressislikelytomeanforthefutureofemployment.Thepresentstudyintendstobridgethisgapintheliterature.Althoughthereareindeedexistingusefulframeworksforexaminingtheimpactofcomputersontheoccupationalemploymentcomposition,theyseeminadequateinexplainingtheimpactoftechnologicaltrendsgoingbeyondthecomputerisationofroutinetasks.Semi-nalworkbyAutor,etal.(2003),forexample,distinguishesbetweencognitiveandmanualtasksontheonehand,androutineandnon-routinetasksontheother.Whilethecomputersubstitutionforbothcognitiveandmanualroutinetasksisevident,non-routinetasksinvolveeverythingfromlegalwriting,truckdrivingandmedicaldiagnoses,topersuadingandselling.Inthepresentstudy,wewillarguethatlegalwritingandtruckdrivingwillsoonbeautomated,whilepersuading,forinstance,willnot.DrawinguponrecentdevelopmentsinEn-gineeringSciences,andinparticularadvancesintheeldsofML,includingDataMining,MachineVision,ComputationalStatisticsandothersub-eldsofArticialIntelligence,aswellasMR,wederiveadditionaldimensionsrequiredtounderstandthesusceptibilityofjobstocomputerisation.Needlesstosay,anumberoffactorsaredrivingdecisionstoautomateandwecannotcapturetheseinfull.Ratherweaim,fromatechnologicalcapabilitiespointofview,todeterminewhichproblemsengineersneedtosolveforspecicoccupationstobeautomated.Byhighlightingtheseproblems,theirdifcultyandtowhichoccupationstheyrelate,wecategorisejobsaccordingtotheirsusceptibilitytocomputerisation.Thecharacteristicsoftheseproblemswerematchedtodif-ferentoccupationalcharacteristics,usingONETdata,allowingustoexaminethefuturedirectionoftechnologicalchangeintermsofitsimpactontheoccu-pationalcompositionofthelabourmarket,butalsothenumberofjobsatriskshouldthesetechnologiesmaterialise.Thepresentstudyrelatestotwoliteratures.First,ouranalysisbuildsonthelaboureconomicsliteratureonthetaskcontentofemployment(Autor,etal.,2003;GoosandManning,2007;AutorandDorn,2013).Basedondenedpremisesaboutwhatcomputersdo,thisliteratureexaminesthehistoricalim-pactofcomputerisationontheoccupationalcompositionofthelabourmar-ket.However,thescopeofwhatcomputersdohasrecentlyexpanded,andwillinevitablycontinuetodoso(BrynjolfssonandMcAfee,2011;MGI,2013).DrawinguponrecentprogressinML,weexpandthepremisesaboutthetaskscomputersareandwillbesuitedtoaccomplish.Doingso,webuildonthetaskcontentliteratureinaforward-lookingmanner.Furthermore,whereasthisliter-aturehaslargelyfocusedontaskmeasuresfromtheDictionaryofOccupational4
。。。以上简介无排版格式，详细内容请下载查看