【国际货币基金组织IMF】2025欧洲人工智能与生产力研究报告_第1页
【国际货币基金组织IMF】2025欧洲人工智能与生产力研究报告_第2页
【国际货币基金组织IMF】2025欧洲人工智能与生产力研究报告_第3页
【国际货币基金组织IMF】2025欧洲人工智能与生产力研究报告_第4页
【国际货币基金组织IMF】2025欧洲人工智能与生产力研究报告_第5页
已阅读5页,还剩64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

ArtificialIntelligenceandProductivityinEurope

FlorianMisch,BenPark,CarloPizzinelliandGalenSher

WP/25/67

IMFWorkingPapersdescriberesearchin

progressbytheauthor(s)andarepublishedtoelicitcommentsandtoencouragedebate.

TheviewsexpressedinIMFWorkingPapersare

thoseoftheauthor(s)anddonotnecessarily

representtheviewsoftheIMF,itsExecutiveBoard,orIMFmanagement.

2025

APR

NATn

·

ARY

*Correspondingauthor.

**WethankSimonBunel,EraDabla-Norris,RomainDuvalandDavidKollforexcellentcomments.MahikaGandhiprovidedexcellentresearchassistance.

©2025InternationalMonetaryFundWP/25/67

IMFWorkingPaper

EuropeanDepartment

ArtificialIntelligenceandProductivityinEurope

PreparedbyFlorianMisch*,BenPark,CarloPizzinelliandGalenSher**

AuthorizedfordistributionbyStephanDanningerandKristinaKostial

Month2025

IMFWorkingPapersdescriberesearchinprogressbytheauthor(s)andarepublishedtoelicit

commentsandtoencouragedebate.TheviewsexpressedinIMFWorkingPapersarethoseofthe

author(s)anddonotnecessarilyrepresenttheviewsoftheIMF,itsExecutiveBoard,orIMFmanagement.

ABSTRACT:ThediscussiononArtificialIntelligence(AI)oftencentersarounditsimpactonproductivity,butmacroeconomicevidenceforEuroperemainsscarce.UsingtheAcemoglu(2024)approachwesimulatethemedium-termimpactofAIadoptionontotalfactorproductivityfor31Europeancountries.Wecompilemanyscenariosbypoolingevidenceonwhichtaskswillbeautomatableinthenearterm,usingreduced-form

regressionstopredictAIadoptionacrossEurope,andconsideringrelevantregulationthatrestrictsAIuse

heterogeneouslyacrosstasks,occupationsandsectors.Wefindthatthemedium-termproductivitygainsforEuropeasawholearelikelytobemodest,ataround1percentcumulativelyoverfiveyears.While

economciallystillmoderate,thesegainsarestilllargerthanestimatesbyAcemoglu(2024)fortheUS.Theyvarywidelyacrossscenariosandcountriesandaresustantiallylargerincountrieswithhigherincomes.

Furthermore,weshowthatnationalandEUregulationsaroundoccupation-levelrequirements,AIsafety,anddataprivacycombinedcouldreduceEurope’sproductivitygainsbyover30percentifAIexposurewere50

percentlowerintasks,occupationsandsectorsaffectedbyregulation.

RECOMMENDEDCITATION:[Misch,F.,Park,B.,Pizzinelli,C.andSher,G.(2024).ArtificialIntelligenceandProductivityinEurope.IMFWorkingPaperWP/25/67.

JELClassificationNumbers:

E24,J24,O30,O47

Keywords:

ArtificialIntelligence;Productivity;Technology;Regulation

Author’sE-MailAddress:

FMisch@;

HPark2@;

CPizzinelli@;

GSher@

WORKINGPAPERS

ArtificialIntelligenceandProductivityinEurope

PreparedbyFlorianMisch

1,

BenPark,CarloPizzinelliandGalenShe

r2

1Correspondingauthor.

2MahikaGandhiprovidedexcellentresearchassistance.

IMFWORKINGPAPERSArtificialIntelligenceandProductivityinEurope

INTERNATIONALMONETARYFUND2

Contents

1.Introduction 3

2.StylizedFacts 6

3.Methodology 6

4.Results 8

4.1VariationinMedium-termProductivityGains 8

4.2PreferredScenario 12

4.3TheRoleofRegulation 15

5.ConclusionsandPolicyImplications 16

Appendix1:SampleandCountry-SpecificData 18

Appendix2:AIExposureofTasksandOccupations 19

Appendix3:AIAdoptionRate 22

Appendix4:LaborCostSavingsfromAI 28

Appendix5:NationalOccupationRegulation 29

Appendix6:EUAIAct 30

Appendix7:DataPrivacyLaws 31

References 32

IMFWORKINGPAPERSArtificialIntelligenceandProductivityinEurope

INTERNATIONALMONETARYFUND3

1.Introduction

Artificialintelligence(AI)isoftenseenasageneral-purposetechnologythathasthepotentialtotransformtheeconomyandspurbroad-basedeconomicgrowth,akintothearrivalofelectricityandpersonalcomputers.

Muchofthedebateonitslikelyimpactthusfocusesonitseffectonproductivity.InEurope,thisquestionis

especiallytopicalgivenlacklusterproductivitygrowthoverrecentdecades,whichresultedinalarge

productivitygapvis-à-vistheUS(IMF,2024).Moreover,thereisawidespreadviewthattheregionisfallingbehindtheUSandChinainAIdevelopmentandadoption,notleastbecauseofitsmorestringentregulatoryenvironment(e.g.,TheEconomist,2023).Theobjectiveofthispaperistoprovideestimatesofthesizeof

effectsofAIontotalfactorproductivity(TFP)acrossEuropeancountriesoverthemediumtermandexamineanyimpedingeffectsofregulationinEurope.

MicroeconomicstudiessuggestlargeproductivitygainsfromdifferenttypesofAIforspecificoccupations.

Theseestimates,whichinmostcasesarebasedonrandomizedtrialswherethetreatmentgroupisgiven

accesstoAItools,rangefrom14%forlow-skilledtaxidriverstoover50%forsoftwareengineers;see

Appendix4forasummary.Firm-levelstudiesontheeffectsofadoptionofAItechnologiesotherthan

generativeAIfindsmallerbutstillsignificantproductivitygains(seeComunaleandManera,2024,and

Filippuccietal.,2024aforsurveys).AseparatestrandoftheliteraturelooksattheimpactofAIonemployment.Cazzanigaetal.(2024),forinstance,notethatalargeshareofjobsgloballyislikelytobeaffectedbyAI,

particularlyinadvancedeconomies,andthatAIwillsubstituteratherthancomplementhumanlaborinmanyjobs.UsingdatafromtheUS,Bonfigliolietal.(2025)andHuang(2024)showthathigherAIadoptionis

associatedwithfallsintheemployment-to-populationratios.AseparatestrandoftheliteratureexaminesmacroeconomicpoliciestobroadenthegainsfromAI;seeBrolloetal.(2024).

Theextenttowhichthesemicro-levelproductivitygainsandpossibleemploymenteffectsareassociatedwithaggregateproductivitygainsandgrowth,however,remainsunclear.Studiesexaminingthemedium-term

macroeconomiceffectofAIshowasubstantiallywiderrangeofestimates.McKinsey(2023)andGoldman

Sachs(Hatziusetal.,2023)envisioncumulativeGDPgainsofabove35percentforadvancedeconomiesand7percentgloballyovera10-yearperiod,respectively.Commissiondel’IntelligenceArtificielle(2024)infers

potentialgrowthimpactsfromAIofupto1.3annuallybydrawingparallelstotheeffectsofelectricityand

InformationandCommunicationtechnologies.Similarly,IMF(2024)andCazzanigaetal.(2024)estimate

annualgrowthimpactsofupto0.8percentagepointsbasedonlaborreallocationandchangesinthecapitalsharecomparabletothoseobservedforautomationinthepast.

Bycontrast,Acemoglu(2024)doesnottakeintoaccountanypotentiallonger-termtransformationaleffectsofAIandthereforeestimatesmuchmoremodestTFPgainsoflessthan0.7percentcumulativelyover10yearswhichhereferstoas‘mediumterm’.Heusesarigorousframeworkthatquantifiesthegainsbottom-upusingmeasuresoftheAIexposureofindividualtasksforeachoccupation.AghionandBunel(2024)showthatusingalternativeassumptionswithinAcemoglu’s(2024)frameworkcan10-foldtheestimatedproductivitygainsfor

theUS.

EvidenceconsideringEurope-specificfactorsandcross-countryheterogeneitywithintheregionremains

relativelyscarce.Bergeaud(2024)simulatesproductivitygainsfromAIfortheeuroareausingtheAcemoglu(2024)framework,combiningsomeoftheoriginalpaper’sparametervalueswithownassumptionsand

estimatestogenerateafewalternativescenarios.Hefindscumulativeproductivitygainsof2.9percentforthe

IMFWORKINGPAPERSArtificialIntelligenceandProductivityinEurope

INTERNATIONALMONETARYFUND4

euroareainhiscentralscenario,whilehiscountry-specificresultsrangefromaround1.5inIrelandto3.3in

Belgium.Filippuccietal.(2024b)extendtheAcemoglu(2024)frameworkbymoreexplicitlymodellingsectoralspilloversandpresenttheproductivitygainsforallG7countrieswhileassumingdifferentadoptionrates.TheestimatedgainsarealsosignificantlyhigherthantheAcemoglu(2024)resultsformostcountriesand

scenarios.

Ourstudyisbroaderinscopeintermsofcountrycoverage,numberofscenariosandconsiderationofthe

effectsofregulation,andituseseconometrically-groundedparameterestimates.Weinvestigatecross-countryvariationfor31EuropeancountriesbothintermsofthemagnitudeandtheuncertaintyoftheproductivitygainsmoresystematicallybyallowingtheratesofAIadoptiontovaryacrosscountriesaccordingtotheireconomic

characteristics.Tothisend,wealsouseAcemoglu’s(2024)frameworktoestimatethemedium-term

productivitygainsfromAIinEurope(whichweinterprettobe5years,giventhemodelcharacteristicsandourassumptions)

.1

First,wequantifytheuncertaintyaroundtheimpactofAIonTFP.Ratherthanmakingourownassumptions,wecombineacomprehensivesetoftheavailableestimatesofAIexposureofindividualtasks,delivering44

scenarios.Foreachscenario,wecalibrateestimatesofAIadoptionbySvanbergetal.(2024),asusedin

Acemoglu(2024),tospecificcountriesandsectorsbasedonregressionevidenceofthedriversofAIadoptioninEurope.WagelevelsturnouttobethemaindriverofAIadoption,ratherthancapitalcosts,industry

concentration,digitalization,orhumancapital.Thisallowsustotakeintoaccounthowdifferencesinwage

levelsandothersectoralandcountry-levelfactorsaffectAIadoptionandshapecross-countryvariationoftheproductivitygains.

Second,weexaminetheroleofregulationinreducingproductivitygainsofAIwhichisoneareaofpolicythatisoftendiscussedwhenitcomestoAI;seeforexampleBradford(2024)forabroaderdiscussion.Survey

evidencefromGermanysuggeststhatwhilefirmsseemanybarrierstoAIadoption,regulationisseenasmostimportant(Wintergerst,2024).Tothisend,weconsiderregulationwhichcouldplausiblyhavelargeeffectson

AIadoption:licensingandtrainingrequirementsforspecificoccupationsatthenationallevel,dataprivacylaws,andtheEUAIAct

.2

Wethenidentifythetasks,occupationsandsectorswherethisregulationcouldundermineAIadoptionandassumethatregulationhalvesAIcapabilities,strikingareasonablemiddlegroundbetween

assumingthatregulationcompletelypreventsAIuseandthatregulationhasnoimpact.

Weshowthatinourpreferredscenario,whichisbasedonwhatwethinkarethemostplausibleassumptionsonAIoccupationalexposureandadoptionratesinEuropeancountries,theEurope-wideeffectsaremodestataround1.1percentcumulativelyoverthemediumterm,whichexceedsAcemoglu’sestimatesfortheUSby

almost60percent.ThesedifferencesaredrivenalmostentirelybymoreoptimisticassumptionsofAI

capabilitieswhichwethinkarejustifiedinlightofarecentrefereedpublicationwhichweexplainbelow.

However,thereissignificantheterogeneityacrosscountries.EstimatedTFPgainsinhigher-incomecountriestendtobemuchlargerthanthoseinlower-incomeeconomiesinlinewithfindingsbyCeruttietal.(2025),due

1IncontrasttoAcemoglu(2024)whoconsidersa10-yearperiod,weassumethatthesimulatedproductivitygainsrefertoa5-yearhorizonfortworeasons.First,theframeworkbyAcemoglu(2024)doesnotcaptureanylarge-scaletransformationaleffects

whichwethinkcouldindeedariseoverperiodsexceeding5years.Second,weassumethatAIcapabilitiesandtheAIadoptionrate(i.e.,theproportionoftasksforwhichitwillbeprofitabletouseAI)willremainconstantovertime.Over10years,botharemorelikelytochange.

2TheEUAIActisakeyAIsafetylawwhichcapsthecapacityofAIsystemsandincreasesthecostofusingAIinadefinedlistofhigh-risksystems.

IMFWORKINGPAPERSArtificialIntelligenceandProductivityinEurope

INTERNATIONALMONETARYFUND5

toboththeirlargershareofvalueaddedinindustriesthathavegreaterAIexposure(e.g.,financialservices)andtheirhigherwagelevels—whichprovidegreaterincentivesforlabor-savingorlabor-augmentingAI

adoption.

Moreover,inhigherincome-countries,therearemuchlargerupsiderisksfromAI.Forexample,inourpreferredscenario,thegainsinLuxembourgcouldbe2percentcumulatively,almosttwicetheEuropeanaverage,and

morethan4timeslargerthanthoseinRomania.ThisisduetothecompositionoftheLuxembourgish

economy,withmorevalueaddedinsectorslikefinancialserviceswithhigherAIexposure,anddueto

Luxembourg’shigherwages,whichgiveemployersthereagreaterincentivetoadoptAI.Inaddition,

productivitygainsinLuxembourgcouldbemorethantwiceashighifAIturnsouttobemorecapablethaninthe‘preferred’scenario,pointingtolargerupsiderisksaswell.

Wealsofindthatthecombinedadverseeffectsofnationaloccupation-levelregulation,theEUAIActanddataprivacylawsonproductivitycouldbesignificant,withtheformertwohavingthelargesteffects.Occupational

restrictionsareoftenoverlookedindebatesonAIbutcouldsubstantiallyreducetheproductivitygainsfromAI.Bycontrast,dataprivacylawswidelyaffectsomeindustrieswithhighAIexposure,suchastheITandfinancialservicessectors,buttheirproductivity-inhibitingeffectissomewhatsmaller.

Ourresultscaninformongoingpolicydebates.Ontheonehand,theysuggestthatoverthemediumterm,AIisnotasilverbullettosignificantlyboostsluggishproductivitygrowthinEurope,ortoclosetheproductivitygaptotheUnitedStates.Moreover,AIcouldslowtherateofincomeconvergenceamongEuropeancountries

becausethegainsfromAItendtobelargerinmoreadvancedeconomies.Ontheotherhand,ourresultshelpinformthepotentialtrade-offsbetweenthebenefitsofregulationsincludingrelatedtoprivacyandsafety,andmaximizingtheproductivitygainsfromAI,whilenotingthatourstudyisnotacomprehensivecost-benefit

analysisoftheseregulations.

Thispaperisorganizedasfollows.Section2presentsstylizedfactsonthediffusionandadoptionofAIto

motivatetheanalysisandthefocusonmedium-termeffects.Section3providesasummaryoftheAcemoglu

(2024)modelanddiscusseshowwecalibratesomeofitskeyparameterstotheEuropeancontext,consideringarangeofalternativescenarios.Section4presentstheresultsandtheanalysisofregulations.Section5

concludes.

INTERNATIONALMONETARYFUND6

2.StylizedFacts

Figure1.SpreadofPastInnovations

(Yearstoreach100millionusers/units)

100

M

M

M

M

M

M

M

M

M

M

M

0135791113151719212325···80YearstoReach100MillionUsers/Unit

NumberofUsers

90

80

70

60

50

40

30

20

10

0

GenAI

TV

Telephone

InternetCableTV

ComputersCellphone

Source:CHATdatabase;IMFstaffcomputation.

Note:TheGenAIandInternetlinereflectstheusercount.ThefiguresforComputers,TV,andCellphonerepresentthenumberofunits.CableTVandTelephoneindicatethenumberof

connections.

ThefastspeedofdiffusionofgenerativeAI(genAI),asubsetofthebroadsetoftechnologiesthatfall

underthedefinitionofAI,overthelasttwoyears

explainsinparttherecentinterestinandpublic

debatesontheeffectsofAImorebroadly.

Comparedtopastinnovations,ittookonlyafew

monthsforgenAItoreach100millionusers

(measuredbytheuserbaseofOpenAI’sChatGPTwhichwasthefirstwidelyavailableandaccessiblegenAIapplication).Bycontrast,ittookyears(and

sometimesdecades)forothergeneral-purpose

technologiestoreachthesamenumberofusers.

Eventhoughthesedifferencesmaybemainly

drivenbytheverylowaccesscostofgenAI

technologiesthroughsmartphonesandpersonal

computers,thishistoricallyunprecedentedspeedofadoptionpointstothepotentialforAItobe

applicabletoawiderangeoftasks.However,aswediscussbelow,thesheernumberofindividualusersneitherimpliesthatAIisbeingemployedtoawide

rangeoftasksnorisitsynonymousforbroaderAIadoptionbyfirms.

3.Methodology

TosimulatetheeffectsofAIonproductivity,thispaperusesthemodelfromAcemoglu(2024),whichinturnisbasedonAcemogluandRestrepo(2018,2019,2022).Inthemodel,theproductionofauniquefinalgood

requiresafixedsetoftaskstobeperformed,andinturn,thesetaskscanbeproducedwitheithercapitalor

labor.Thisframeworkserveswelltoexaminethemedium-termeffectsofAI,interpretedasthegrowthinoutputandTFPoriginatingfromsmallchangesinproductivityandinthemixoflaborandcapitalinputs,butwithout

fundamentallong-runchangesinthestructureoftheeconomy(i.e.,itssectoralcompositionandtasks).

Acemoglu(2024)includestwochannelsforAI-basedproductivitygains:automationandtaskcomplementarity.Theformerentailsasubstitutionofworkersintheperformanceofindividualtaskswithinanoccupation,

decreasingtheoverallneedforhumanlabor.Thelatterreferstopartialautomationoftaskssothatthemarginallaborproductivityincreasesincomplementarytasksperformedbyhumans.

Acemoglu(2024)thenappliesHulten’stheorem(Hulten,1978),whichshowshowmicro-level,non-

transformationalproductivityimprovementsinindividualoccupationstranslateintomacrochangesand

aggregatedproductivitygrowthincompetitiveeconomieswithconstantreturnstoscale.SinceAcemoglu’s

focusisonsmallchangesintechnologyandthecompetitiveequilibriumisefficient,theimpactofall

reallocationsoffactorsacrosstasksandindirecteffectsviapricesareofsecondaryorderandthereforesmallenoughtobeignoredincomputingtheproductivityandGDPgainsduetoAI.Thisalsoimpliesthatthemodel

IMFWORKINGPAPERSArtificialIntelligenceandProductivityinEurope

INTERNATIONALMONETARYFUND7

isnotsuitableforestimatinganypotentiallylargetransformationalandlonger-termeffects,suchasthecreationofnewindustriesoranaccelerationintherateofscientificdiscoveries.

EstimatingtheproductivitygainsofAIadoptionintheAcemoglu(2024)modelamountstocalibratingthreekeyparameters.ThefirstrequiredinputconsistsofameasureoftheexposuretoAIofdifferentoccupations,

essentiallyreflectingassumptionsaboutAIcapabilitiesandthepotentialscopeofAIapplications.Importantly,forthepurposeofestimatingproductivitygains,exposurecanrefertobothautomationandtask

complementarities

.3

Usingaversionofthetask-basedoccupationalexposuremeasureconstructedby

Eloundouetal.(2024),Acemoglu(2024)calculatesthatthewagebill-weightedshareofexposedtasksinthe

U.S.economyis19.9percent,meaningthatAIhasthecapabilityofperformingaround20percentoftasksin

theU.S.economy(whenwagebillweightsareusedtoproxytherelativeimportanceoftasksandoccupations).

ThesecondparameterofinterestistheAIadoptionrate,i.e.,theshareofAI-exposedtaskswherethebenefitsofusingAIexceedthecosts,thusmakingAIadoptionprofitable.Intuitively,theremaybetasksthatAIcan

performbutwhereitsapplicationmaybetoocostlyrelativetothepriceoflaborinthesamejob.Hence,therawexposuremeasureofanoccupationshouldbecombinedwithanestimateoftheeconomicfeasibilityof

adoption.Acemoglu(2024)drawsfromthecostingestimatesofSvanbergetal.(2024)toassumethatthebenefitsofusingAIexceedthecostsfor23percentofAI-exposedtasks.

Finally,translatingtheapplicationofAIintoproductivitygainsrequiresanestimateofthesavingsintermsof

laborcoststhatAIprovideswhenproducingaunitofoutput.Acemoglu(2024)usesanaverageofthree

microeconomicproductivityestimatestocalibratetheassumptionthattasksautomatedbyAIreducelabor

costsby27%.Theselaborcostsavingstranslateinto15%total(laborandcapital)costsavings,giventhat

laborcostsaccountfor53percentofoutput.Multiplyingthesethreekeyparameters(19.9%,23%,and15%),

Acemoglu(2024)estimatesa0.71%cumulativemedium-termincreaseintotalfactorproductivityfortheUS.IncontrasttoAcemoglu(2024)whoconsidersthemediumtermtobe10years,weassumethatthecumulative

medium-termgainsariseover5years.

Weapplythesamemethodologyto31Europeancountries(seeAppendix1forlistofcountries).Foreach

country,weobtainthefirstkeyparameter—theshareofAI-exposedtasksinindustry-specificvalueadded—byweightingoccupation-levelexposuresbyeachoccupation’swagebillwithinanindustry(seeAppendix2fortheAIexposureestimatesfromtheliteratureweuse).Toobtainthesecondkeyparameter—theAIadoptionrate—weestimatethehistoricalrelationshipbetweenAIadoptionandcountries’andsectors’economic

characteristics(e.g.,laborandcapitalcosts)inEurope.Thisgivestheshareoftasksineachcountry-industry

pairforwhichitisprofitabletoapplyAI(seeAppendix3).Wecalibratethethirdkeyparameter(costsavings)inthesamewayasAcemoglu(2024).Finally,wecalculateproductivitygainastheproductofthesethreekey

parameters(seeAppendix1fordetailsondatausedforweightingandlaborshares).

3Thedistinctionbetweenautomationandcomplementarity,however,wouldberelevantforexaminingtheimpactofadoptiononemploymentandlaborproductivity,whichweleaveforfutureresearch.

IMFWORKINGPAPERSArtificialIntelligenceandProductivityinEurope

INTERNATIONALMONETARYFUND8

4.Results

4.1VariationinMedium-termProductivityGains

Wepresenttheresultsinseveralstages.First,wecompareproductivitygainsinEuropetotheestimatesby

Acemoglu(2024)fortheUS.Second,weshowhowalternativeassumptionsofAIcapabilitiesandAIadoptionaffecttheproductivitygainsacrossthecountriesinoursample.Finally,wecombineallassumptionstoarrivetoaplausiblerangeofestimatesforeachcountry.

First,weapplythesamemethodologyasinAcemoglu(2024)toEuropeinordertocompareourestimateswiththeonesforAcemoglu(2024)fortheUS.Inparticular,wecalculatetheproductivitygainsforEuropean

countriesandEuropeasawhole,usingthesamemeasureofoccupation-levelAIexposure,AIadoptionrate,

andlaborcostsavingsfromAIasintheoriginalAcemoglu(2024)paper.Thisparametrization,henceforth

referredtoasthe‘Acemoglu(2024)baseline’,allowsustoexaminehowdifferencesinthesectoralcompositionandwagestructuretranslateintodifferencesintheproductivitygainsfromAIbetweentheUSandEuropean

countries,giventhattherearenootherdifferencesbetweenAcemoglu’sresultandourresultsforEurope.

Figure2showsthat,basedondifferencesinthesectoralcomposition,theaverageproductivitygainsinEuropearesomewhatlargerthanintheUS,butnotbymuch:theyamounttocloseto0.8percentinEurope,

cumulativelyoverthemediumterm,comparedtoaround0.7inAcemoglu(2024)fortheUS.However,thereissubstantialcross-countryvariation,rangingfromaround0.5percentinRomaniatocloseto1percentin

Luxembourg.Broadly,higher-incomecountrieshavelargergains,aresultdrivenbythehigherprevalenceofwhite-collarservicesincludingforinstancefinancialservices,whichtendtobemoreexposedtoAI.

IMFWORKINGPAPERSArtificialIntelligenceandProductivityinEurope

INTERNATIONALMONETARYFUND9

Figure2.ProductivitygainsinEuropeandtheUScompared

Sources:Eurostat,EULFS,EUSILC,andIMFstaffcalculations.

Note:TheblackandredlinesrepresenttheaverageTFPgainforthe31EuropeancountriesinoursampleandfortheUSasestimatedbyAcemoglu(2024),respectively.

Second,wequantifytheuncertaintyaroundtheAcemoglu(2024)baselineresultspresentedinFigure2usingacomprehensivesetofalternativescenariosaboutAIcapabilities(AItask-levelexposuremeasures)

.4

WethusrepeattheexerciseconsideringvariousalternativeestimatesonAIexposure.Appendix2providesalistofAI

exposureestimatesandbriefdescriptionsoftheirmainmethodologicaldifferences.Forinstance,some

computeexposureforspecifictasksandthenconsidereachoccupationasabundleoftasks(Eloundouetal.,2024,Gmyreketal.,2023,Webb,2019),whileothersfocusontheoverlapbetweenAIapplicationandhumanskills(Feltenetal.,2021).Beyondthemechanicalmeaningoftestingthesensitivityoftheresultstothese

assumptions,exploringalternativemeasuresofexposurethusalsoreflectstheconsiderableuncertaintyaroundhowindividualtasksandoccupationswillbeaffectedbyAI.Figure3showsthatwhilemoreconservative

assumptionsmute

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