IMF-人工智能的经济影响与调控——学术文献与政策行动综述（英）-2024.3_市场营销策划_202.docx

1 IntroductionThisreviewpaperinvestigateshowArtificialIntelligence(AI)affectstheeconomyandhowthetechnologyhasbeenregulated,relyingonacademicandpolicysourcesthroughearly2024.Wecoverinsightsonemploymentandwageeffects,productivity,andeconomicgrowthfromtheeconomicliterature.Inthepolicyrealm,wesummarizetheregulatoryactionsundertakenindifferentregions,detailingtheirrationales,approachesandareasofcoverage.GiventherapidevolutionofAItechnologiesandtherelatedliterature,thepaperaimstoprovideastructuretoorganizethelatestcontributionsfortheuseofpolicymakers,economists,researchers,andindustrystakeholders.Beforedelvingdeeperintothecontentofourpaper,weshallclarifyitsscopewithsomekeydefinitions.ProfessorJohnMcCarthy,oneoftheorganizersofthe1956DartmouthresearchprojectthatstartedAIasafield,defined"AI"as"thescienceandengineeringofmakingintelligentmachines(McCarthy2007).TheoriginaldocumentcoiningthetermAIwasthe1955proposalfortheDarthmouthResearchProjectcoauthoredbyMcCarthy,Minsky,RochesterandShannon,seeInthisreview,weconcernourselveswiththeeconomicimpactandregulationofrecentadvancesinAI,suchasmachinelearning(M1.),anditssub-fieldsofdeeplearning,generative-Al(gen-AI)and1.arge1.anguageModels(1.1.Ms).M1.studieshowmputeragentscanimprovetheirperception,knowledge,thinking,oractionsbasedonexperienceordata.1.1.MsarebroadlydefinedasneuralnetworksthatIeamsntextandmeaningbytrackingrelationshipsinsequentialdata.SeeBroadly,wecoverempiricalstudiesfocusingoneitherM1.applicationsexcludinggen-AI(44pre-gen-A,untillate2022)oronthelatestgen-AIand1.1.Ms("post-gen-AI,”mostlysince2023).Thisdistinctionismotivatedbythefactthat1.1.Msandgen-AIhavecometotheforeoftheeconomicandpolicydebateafterthereleaseofDall-E2andChatGPTbyOpenAIinlate2022,andrelateddata,applicationsandresearchremainrelativelyscarce.Accordingly,wereferto"M1."asmachinelearningordeeplearningforprediction,imageandpatternrecognition,textanalysisanddataanalysis.Whenusingtheterm"M1.,"weexcludegen-AI,whichistreatedseparatelyduetothewideavailabilityoftext-andimage-generating1.1.Mtools.WewillclarifythetypeofAIcoveredbyeachpaperwhenthecontextdoesnotmakeitclear.Consistentwiththisdiscussion,wewilluse“M1."todenoteresearchthatappliestomachinelearningandexcludesgen-AI,usuallyforlackofavailabledate.Weinsteaddenotestudiesas“gen-AI“iftheyfocusexclusivelyongen-AI.Whenweusetheterm"AI,"werefertocontextsencompassingbothM1.andgen-AItechnologies.Thisdistinctionismostlyrelevanttoempiricalorexperimentalpapers,sincetheoreticalstudiesoftendonotdistinguishgen-AIfromM1.Inthecaseofregulation,thereisnoglobalconsensusonadefinitionof“AI"orthetechnologiescoveredby"AI"regulation.SeeO'Shaughnessy(2022)andalsoAppendixA.lWestartbydescribingIheimpactofAIonlabormarkets,wheretheliteraturehasfocusedonemploymentandwagesofvariousoccupationalgroups.MostofthestudieswereviewadoptimplicitlyorexplicitlythetaskframeworksofZeira(1998),AcemogluandAutor(2011),andAcemogluandRestrepo(2018)thathavebeenappliedtostudyhuman-robotsubstitutionintheautomationliterature.Aswebrieflydiscuss,theseframeworksmodeloutputasabundleoftaskscarriedoutbyeitherworkersorcapital,andobtainthatemploymentandwagesofdifferentoccupationalordemographicgroupsaredirectlyrelatedtothequantityoftasksassignedtoeachgroup.Accordingly,theearlierpre-gen-AIliteratureandmuchofthepost-gen-AIliteraturefocusedoncomputing“taskexposures',-theshareoftasksthatcanpotentiallybereplacedbyAI-togiveestimatesofthepotentialimpactofAIongroupsofworkers.Followingtheearlierautomationliterature,theseworksgenerallyassociateahighertaskexposurewithlargerpotentialdisplacementforaffectedgroupsofworkers.Moststudiesagreethatwhite-collar,higher-skilledoccupationshavehighertaskexposuresandthereforefacestrongeremploymentriskfromAIadoption(e.g.,ongen-AI,Eloundouetal.(2023).Otherresearchersinsteadseparatedtaskexposureandemploymentrisk,highlightingtheaugmentationpotentialofAItechnologiesorother“shieldingfactors”(e.g.,Cazzanigaetal.(2024).EmpiricalstudiesonM1.showthatemploymenteffectsmightoverallbenilorpositive(Acemoglu,Autor,Hazell,etal.2022;Albanesietal.2023),whileexperimentalpapersongen-AIhighlightproductivitygainsforlow-skilledworkersthatcouldharborinwagecompression(NoyandW.Zhang2023;Brynjolfsson,1.i,andRaymond2023).Ultimately,theorysuggeststhatthefinalverdictonlabormarketimpactsofAIwilldependontheracebetweenjobdisplacementandproductivityincreases,resultingfromdirectworkercomplementationoreconomy-widegainsfromAI.Thefindingsthatwepresentprovidesomeelementstoevaluatethistrade-off,butleaveustarfromadefinitiveunderstanding.Inparticular,allstudiesseemtoagreethatexposureispervasive,butremaininconclusiveonhowsuchexposuremaytranslateinsubstitutionofcomplementationofworkers.WecloseourdiscussionoflabormarketimpactsofAIwithadiscussionofpotentialpolicytotacklethepotentiallabordisplacementbroughtaboutbythistechnology.Next,weproceedtosurveythemorelimitedstudiesconcerningproductivityandgrowtheffectsofAI,whichencompasstheory,firm-levelstudiesonM1.,andtheexperimentalevidenceongenAIcitedabove.Theworkinth