2025哈佛大學(xué)生成式 AI 與工作的本質(zhì)_第1頁(yè)
2025哈佛大學(xué)生成式 AI 與工作的本質(zhì)_第2頁(yè)
2025哈佛大學(xué)生成式 AI 與工作的本質(zhì)_第3頁(yè)
2025哈佛大學(xué)生成式 AI 與工作的本質(zhì)_第4頁(yè)
2025哈佛大學(xué)生成式 AI 與工作的本質(zhì)_第5頁(yè)
已閱讀5頁(yè),還剩69頁(yè)未讀, 繼續(xù)免費(fèi)閱讀

下載本文檔

版權(quán)說(shuō)明:本文檔由用戶提供并上傳,收益歸屬內(nèi)容提供方,若內(nèi)容存在侵權(quán),請(qǐng)進(jìn)行舉報(bào)或認(rèn)領(lǐng)

文檔簡(jiǎn)介

WorkingPaper25-GenerativeAIandtheNatureof

生成式AI與工作的本質(zhì)GenerativeAIandtheNatureofWorkingPaper25-Copyright?2024byManuelHoffmann,SamBoysel,FrankNagle,SidaPeng,andKevinWorkingpapersareindraftform.Thisworkingpaperisdistributedforpurposesofcommentanddiscussiononly.Itmaynotbereproducedwithoutpermissionofthecopyrightholder.Copiesofworkingpapersareavailablefromtheauthor.FundingforthisresearchwasprovidedinpartbyHarvardBusiness

生成式AI與工作本質(zhì)Copyright?2024byManuelHoffmann,SamBoysel,FrankNagle,SidaPeng,andKevin處獲得。FundingforthisresearchwasprovidedinpartbyHarvardBusinessGenerativeAIandTheNatureofAbstract:Recentadvancesinartificialintelligence(AI)technologydemonstrateconsiderablepotentialtocomplementhumancapitalintensiveactivities.Whileanemergingliteraturedocumentswide-rangingpro-ductivityeffectsofAI,relativelylittleattentionhasbeenpaidtohowAImightchangethenatureofworkitself.Howdoindividuals,especiallythoseintheknowledgeeconomy,adjusthowtheyworkwhentheystartusingAI?Usingthesettingofopensourcesoftware,westudyindividualleveleffectsthatAIhasontaskallocation.WeexploitanaturalexperimentarisingfromthedeploymentofGitHubCopilot,agener-ativeAIcodecompletiontoolforsoftwaredevelopers.Leveragingmillionsofworkactivitiesoveratwoyearperiod,weuseaprogrameligibilitythresholdtoinvestigatetheimpactofAItechnologyonthetaskallocationofsoftwaredeveloperswithinaquasi-experimentalregressiondiscontinuitydesign.WefindthathavingaccesstoCopilotinducessuchindividualstoshifttaskallocationtowardstheircoreworkofcodingactivitiesandawayfromnon-coreprojectmanagementactivities.Weidentifytwounderlyingmechanismsdrivingthisshift-anincreaseinautonomousratherthancollaborativework,andanincreaseinexplorationactivitiesratherthanexploitation.Themaineffectsaregreaterforindividualswithrelativelylowerability.Overall,ourestimatespointtowardsalargepotentialforAItotransformworkprocessesandtopotentiallyflattenorganizationalhierarchiesintheknowledgeeconomy.JEL-Classification:H4,O3,Acknowledgement:TheauthorsaregratefulforfinancialandadministrativesupportfromGitHuband,inparticular,forgenerousadvicefromPeterCihon.WethankShaneGreenstein,TimSimcoe,DavidAutor,andSamRansbothamfortheirfeedback.TheauthorsarealsoindebtedforcommentsbyseminarparticipantsattheresearchseminarsfromtheHarvardLaboratoryforInnovationScience,BostonUniversity,theMassachusettsInstituteofTechnology,andtheUniversityofPassau.Wearefurthergratefulforfeedbackfromparticipantsatthe“LaborintheAgeofGenerativeAI”conferenceattheUniversityofChicago,theNBERSI2024DigitalEconomicsandArtificialIntelligenceinCam-bridge,MA,the2024NBERProductivitySeminarinCambridge,MA,the2024AcademyofManagementScienceinChicago,IL,the22ndZEWEconomicsofICTconference,inMannheim,Germany,the20thSymposiumonStatis-ticalChallengesinElectronicCommerceResearchinLisbon,Portugal,theACMCollectiveIntelligenceConferenceinBoston,MA,theMITCodeConferenceinCambridgeMAandtheCESifoAreaConferenceonEconomicsofDigitization2024inMunich,Germany.

生成式AI與工作本質(zhì)有越來(lái)越多的文獻(xiàn)記錄了AI的廣泛生產(chǎn)力影響,但相對(duì)較少關(guān)注AI如何改變工作的本質(zhì)本身。當(dāng)個(gè)人,尤其是知識(shí)經(jīng)濟(jì)中的個(gè)人開(kāi)始使用AI軟件的設(shè)置,研究AI對(duì)任務(wù)分配的個(gè)體層面影響。我們利用GitHubCopilot的部署這一自然實(shí)驗(yàn),GitHubCopilot是一個(gè)面向軟件開(kāi)發(fā)者的生成式AI代碼補(bǔ)全工具。利用兩年期間數(shù)百萬(wàn)個(gè)工作活動(dòng),我們使用項(xiàng)目資格門檻來(lái)調(diào)查AI不連續(xù)設(shè)計(jì)。我們發(fā)現(xiàn),能夠訪問(wèn)Copilot的個(gè)人會(huì)促使他們將任務(wù)分配轉(zhuǎn)向核心編碼活動(dòng),并言,我們的估計(jì)表明,AIJEL分類:致謝:作者感謝GitHub在財(cái)務(wù)和管理方面的支持,特別是感謝PeterCihonShaneGreensteinTimSimcoeDavidAutorSamRansbotham“生成式人工智能時(shí)代的勞動(dòng)”會(huì)議、NBERSI2024數(shù)字經(jīng)濟(jì)學(xué)和人工智能會(huì)議(馬薩諸塞州劍橋)、2024年NBER生產(chǎn)力研討會(huì)(馬薩諸塞州劍橋)、2024年管理科學(xué)學(xué)會(huì)會(huì)議(伊利諾伊州芝加哥)、22ZEW20會(huì)、ACM集體智能會(huì)議(馬薩諸塞州波士頓)、麻省理工學(xué)院代碼會(huì)議(馬薩諸塞州劍橋)以及CESifo數(shù)字化經(jīng)濟(jì)領(lǐng)域會(huì)議2024(德國(guó)慕尼黑)的參與者提供的反饋。Throughouthumanhistory,therehavebeenahandfuloftechnologicalinnovationsthatfundamen-tallyshifthowtheeconomyworks.Theprintingpress,internalcombustionengine,andcomputersareoft-citedexamplesofsuchgeneralpurposetechnologies.Althoughartificialintelligence(AI)hasexistedforsometime,manyhavearguedthatrecentadvancesmaypushitintothiselitecate-goryoftechnologiesthatalterthecourseofhistory(Crafts,2021;Goldfarb,Taska,andTeodoridis,2023;Eloundouetal.,2024).IfAI—broadlydefinedastheuseofcomputersandmachinestomimichumanintelligence––isdestinedtohavesuchasubstantialimpact,wearelikelystillatthebeginningofthistechnologicalrevolutionthatisslowlyandsteadilyreachingallsectorsoftheeconomy(Acemogluetal.,2022).Importantly,thehighesteconomicimpactofAIispredictedtobeonproductivitygrowththroughthelabormarket,especiallyinknowledgeintensiveindus-tries(BughinandManyika,2018;Sachs,2023).However,duetothenoveltyandbreadthofAI,researchisonlystartingtoelucidateitsimpactonthenatureofworkandtaskallocationinproduc-tionsettings.ThisisparticularlytrueofgenerativeAI(generativeAI)—asubsetofAIbuiltonlarge-languagemachine-learningmodels(LLMs)—whichexplodedontothescenein2022andcurrentlyrepresentsthecutting-edgeofAI.Thesemodels,includingOpenAI’sGPT4,Google’sGemini,1Meta’sLLaMa,andnumerousothers,aretrainedonmassive,Internet-scaledatabasesandusebillionsofparameterstoconstructaprobabilisticmodelthatpredictswhatthenextwordinananswertoapromptfromausershouldbe.Thesemodelscanalsobetrainedondatasetsthataremorefocusedonspecificcontexts—e.g.,health,finance,customerservice,softwaredevelop-ment,etc.Whetherandhowthesenewtechnologieswillshapethenatureofworkremainopenquestions.Further,whetherAIcanbeacomplementtoskilledworkers(Autor,2024)andhelpaddresscriticalaspectsofteamproduction,especiallyinthecontextofdistributedwork,hasgoneAlthoughsomeearlystudiesongenerativeAIhaveshownpositivehigh-levelproductivityimpacts(Brynjolfsson,Li,andRaymond,2023;Dohmke,Iansiti,andRichards,2023;NoyandZhang,2023;Pengetal.,2023),itislessclearwhatthemechanismsbehindtheseimprovements1FormerlyknownasBard.

這類通用技術(shù)中經(jīng)常被引用的例子。盡管人工智能(AI)已經(jīng)存在了一段時(shí)間,但許多人認(rèn)為最近的進(jìn)步可能將其推向改變歷史進(jìn)程的精英技術(shù)類別(Crafts2021Goldfarb,?來(lái)模擬人類智能——注定要產(chǎn)生如此重大的影響,那么我們可能還處于這場(chǎng)正在緩慢AI(生成AI)(LLMs)AI2022AIOpenAIGPT4Gemini1Meta的LLaMa以及許多其他模型,它們?cè)邶嫶蟮?、互?lián)網(wǎng)規(guī)模的數(shù)據(jù)庫(kù)上進(jìn)行訓(xùn)練,什么。這些模型還可以在更專注于特定上下文的數(shù)據(jù)庫(kù)上進(jìn)行訓(xùn)練——例如,健康、金Pengetal.2023),但這些改進(jìn)背后的機(jī)制尚不清楚。之前被稱為Bardare.DoestheuseofgenerativeAIshiftuserstofocusonparticulartypesoftasksthatleadtothoseproductivityimprovements?Ifso,whichtasks?HowexactlydoestheworkprocesschangewhenusinggenerativeAI?Toanswerthesequestions,wedevelopatheoreticalmodelthatleadstotestablehypothesesthatofferinsightsintowhereandwhythemostsalientimpactsarelikelytooccur.Understandingtheseimpactsinformslaborstrategyinamannerrelevanttobothfirms(Tamayoetal.,2023)andpolicymakers(U.S.DepartmentofLabor,2024),includinghiringpoli-cies,worktrainingprograms,andupskillingorreskillingeffortsforcurrentemployees.ThekeychallengeintestingourhypothesesandassessinghowAIchangesthenatureofworkforworkershasbeenintroducedinaquasi-exogenousmanner.Oursetting—theintroductionofGitHubCopilot,asoftwaredevelopmentgenerativeAItool,forkeydevelopers(knownasmain-tainers)inopensourcesoftware(OSS)projects—addressesbothofthesecriteria.OSSsourcecodeispubliclyavailableandpermissivelylicensedforuse,modification,andredistribution.Fre-quentlydevelopedbydistributedteamsofdevelopers,OSSisaclassicexampleofaproductthatisproducedthroughthedistributedworkofteamsandisgenerallyfree(MoonandSproull,2002).AlthoughOSScreatessocietalvalueontheorderoftrillionsofdollars(Hoffmann,Nagle,andZhou,2024)andisthereforeimportantinitsownright,weargueandprovidesuggestiveevidencethatthefindingsinthissettinggeneralizetothebroadersetofworkactivitiesthatoccurintheknowledgeeconomy.Further,aswithmanyteamproductionsettings,OSSalsosuffersfromthe“l(fā)inchpin”problem(Ballester,Calv′-Armengol,andZenou,2006;Godin,2010)asasmallsetofdevelopersarethedrivingforcebehindthewidelyusedandincrediblyvaluabledigitalinfras-tructurethathascometounderliesoftwaredevelopmentandthemoderneconomyasawhole(Eghbal,2020;Geiger,Howard,andIrani,2021;Hoffmann,Nagle,andZhou,2024).Inpractice,aninfluxofnon-expertsenabledbydecreasingcommunicationcosts(Altman,Nagle,andTush-man,2015)createsanadditionalburdenondevelopers,whomusttriagesupportrequests,reviewcontributions,andotherwisemanagetheirproject’sgrowingcommunity.Indeed,surveyevidence

AI是哪些任務(wù)?使用生成式AI時(shí),工作流程究竟是如何變化的?為了回答這些問(wèn)題,我們的地方和原因提供了見(jiàn)解。了解這些影響將有助于企業(yè)(Tamayo,2023)定者(美國(guó)勞工部,2024)的勞動(dòng)力戰(zhàn)略,包括招聘政策、工作培訓(xùn)計(jì)劃和為現(xiàn)有員工進(jìn)AI(1)工作模式是可觀察的,并且(2)一種專門為工人定制的AI工具以準(zhǔn)外生的方式引入。我們的環(huán)境——GitHubCopilot軟件開(kāi)發(fā)生成式AI工具的引入,針對(duì)開(kāi)源軟件(OSS)項(xiàng)目中的主要開(kāi)發(fā)者(稱為維護(hù)者)OSS并且具有許可使用、修改和重新分發(fā)的許可。通常由分布式開(kāi)發(fā)團(tuán)隊(duì)開(kāi)發(fā),OSS是產(chǎn)品通過(guò)團(tuán)隊(duì)分布式工作生產(chǎn)的經(jīng)典例子,通常是免費(fèi)的(Moon和Sproull,2002)。盡管OSS創(chuàng)造了數(shù)萬(wàn)億美元的社會(huì)價(jià)值(Hoffmann,Nagle和Zhou,4),因此本身很重要,但我們認(rèn)為并提供了有說(shuō)服力的證據(jù),證明在這個(gè)環(huán)境中得出的結(jié)論可以推廣到知識(shí)經(jīng)濟(jì)中發(fā)生的更廣泛的工作活動(dòng)。此外,與許多團(tuán)隊(duì)生產(chǎn)環(huán)境一樣,OSS也面臨著“關(guān)鍵人物”問(wèn)題(Ballester,Calvo?Armengol和Zenou,2006;,0),因?yàn)橐恍〔糠珠_(kāi)發(fā)者是廣泛使用且價(jià)值巨大的數(shù)字基礎(chǔ)設(shè)施背后的驅(qū)動(dòng)力,該基礎(chǔ)設(shè)施已成為軟件開(kāi)發(fā)和整個(gè)現(xiàn)代經(jīng)濟(jì)的基礎(chǔ)(,0;Geiger,Howard和Irani,1;Hoffmann,Nagle和Zhou,4)。實(shí)際上,由于通信成本的降低,非專家的大量涌入(Altman,Nagle和Tushman,5)給開(kāi)發(fā)者帶來(lái)了額外的負(fù)擔(dān),他們必須對(duì)支corework(coding)andtoomuchonmanagerial(projectmanagement)tasks(Nagleetal.,2020a).Withthesefactorsinmind,interventionswiththepotentialtorelaxconstraintsonkeyindividualsareofgreatinteresttothedistributedproductionsettingofOSSandarelikelytogeneralizetonumerousothersettingsasdistributedworkhasbecomeincreasinglycommon.WeexploitaspectsofthegeneralaccesslaunchofGithubCopilottothebroaderpublicinJune2022toestablishcausaleffectsofgenerativeAIwheresomedevelopersbelowacertainthresholdofaninternalrankingreceivedfreeaccesstothecodingAIandothersdidnot.Westartwithapanelof187,489distinctdevelopersobservedweeklyfromJuly2022throughJuly2024,whichresultsinmillionsofdeveloper-weekobservationsforCopilotusageandactivitylevelsinpublicGitHubrepositories.2Withinthedatasetoftopdevelopers,wefindthatthosewhoreceivefreeaccesstoCopilotduringthegeneralaccessperiodincreasetheirrelativeshareofcodingtaskswhilereducingtheirrelativeshareofprojectmanagementactivities.Thedynamicsofthetreatmenteffectsarestableforourtwoyearperiod.Wedigfurtherintothemechanismsunderlyingtheseeffectsandfindthattheyaredrivenbyanincreaseinautonomousbehavior(andarelateddecreaseincollaborativebehavior)andanincreaseinexplorationbehavior(ratherthanexploitation).Further,wefindlowerabilitydeveloperswhoreceiveaccesstoAIincreasecodingandreduceprojectmanagementtoagreaterextentcomparedwiththeirhigherabilitypeers.Theresultsarerobusttothestandardregressiondiscontinuitydesigntestsandtodifferentestimationproceduressuchasdifference-in-differenceandmatching.Further,theresultsareconsistentwhenconsideringwhetherdevelopersareworkingonbehalfoftheiremployersorasvolunteers,addingsupporttothelikelihoodthatthesefindingsgeneralizebeyondtheOSSsettingtoabroadersetofOurresultscontributetoagrowingliteratureontheproductivityimpactsofAIinimportantways.Earlyworkinthisareapositsgeneralproductivitygains(Agrawal,Gans,andGoldfarb,2019;Corrado,Haskel,andJona-Lasinio,2021;RajandSeamans,2018),butthatthegainsmaynotbeevenlydistributed(Brynjolfsson,Rock,andSyverson,2018;FurmanandSeamans,2AGitHubrepositoryisalocationwhereallaspectsofaprojectarestoredincludingitssourcecode,andrevision

核心工作(編碼)以及過(guò)多的管理任務(wù)(Nagle等人,2020a)??紤]到這些因素,旨我們利用2022年6月GitHubCopilot向更廣泛公眾開(kāi)放的一般訪問(wèn)功能,來(lái)建立生成式AI的因果效應(yīng),其中一些低于內(nèi)部排名一定閾值的開(kāi)發(fā)者獲得了免費(fèi)訪問(wèn)編碼AI2022720247187,489不同的開(kāi)發(fā)者面板開(kāi)始,這導(dǎo)致了數(shù)百萬(wàn)個(gè)開(kāi)發(fā)者周觀察結(jié)果,用于Copilot在公共GitHub存儲(chǔ)庫(kù)中的使用和活動(dòng)水平。在頂級(jí)開(kāi)發(fā)者的數(shù)據(jù)集中,我們發(fā)現(xiàn)那些在一般訪問(wèn)期間獲得Copilot免費(fèi)訪問(wèn)的人增加了他們的編碼任務(wù)相對(duì)份額,同時(shí)減少了他們的探索行為的增加(而不是利用)驅(qū)動(dòng)的。此外,我們發(fā)現(xiàn)能力較低的開(kāi)發(fā)者在使用AI我們的結(jié)果以重要方式為關(guān)于AI生產(chǎn)力影響的日益增長(zhǎng)的文獻(xiàn)做出了貢獻(xiàn)。該領(lǐng)域早期的工作提出了總體生產(chǎn)力收益(Agrawal,Gans和Goldfarb,2019;Corrado,Haskel和Jona?Lasinio,2021;Raj和Seamans,2018),但收益可能并不均勻(Brynjolfsson,Rock和Syverson,2018;Furman和Seamans,2019)。GitHub存儲(chǔ)庫(kù)是一個(gè)項(xiàng)目所有方面都存儲(chǔ)的地方,包括其源代碼、文檔和修訂歷Subsequentempiricalworkhaslargelyconfirmedthesepredictionsandfoundwide-rangingpro-ductivitybenefitstousingAI,atboththefirmlevel(Czarnitzki,Fern′ndez,andRammer,2023)andtheindividuallevel(F¨generetal.,2022).Particularlyrelatedtothisstudy,researchfo-cusedonCopilotspecificallyhaseitherbeenconductedusingamuchsmallersampleofworkerswithinfirms(Cuietal.,2024)orrelyingonobservationaldatawithoutthebenefitofknowingpreciselywhichcontributorstoOSSweregivenfreeaccesstoCopilot(Yeverechyahu,Mayya,andOestreicher-Singer,2024).Ourworkisconsistentwiththispriorresearchbutaddsadditionalnuancetothelaboraugmentingtechnicalchangeliterature(Acemoglu,2003).Bygoingbeyondproductivitytoexplorehowtechnologychangesthenatureofwork,weprovideoneofthelargestnaturalexperimentsofgenerativeAIandit’simpactonhighlydisaggregatedmeasuresofworkprocesses“inthewild”overatwoyeartimehorizon.OurmainfindingsidentifychangesinthenatureofworkofAIadoptersintheirknowledgeworkprocesses.WeshowthatwhensoftwaredevelopersleverageAImore,theyreallocatetheireffortstowardstechnicalcodingactivitiesandawayfromauxiliaryprojectmanagementactivitiesthatinvolvesocialinteractionswithotherdevelopers.Thisisasignthattheworkerslikelywillintensifytheircorecontributionstopublicgoods,suchasopensourcesoftware,whenleverag-ingskillaugmentingtechnologylikegenerativeAI.Itisalsoconsistentwithreducedcollaborativefrictionsduringtheproblemsolvingprocessofworkandachangeinthewayworkersinteractwitheachotherontheplatform.WecomplementthecurrentliteraturethatleveragesITandconsultancychatsupportAIsandfocusesonhigh-levelproductivityimpactsthroughexperimentation(Bryn-jolfsson,Li,andRaymond,2023;Dell’Acquaetal.,2023)byinvestigatingthenatureofworkthroughchangesinworkactivitiesandhumaninteractionprocessesoverthetwoyearsfollowingtheintroductionoftheprogrammingLLM.BeyondtheidentificationofcausaleffectsthatgenerativeAIhasondecentralizedwork,ourresultssuggestimportantimplicationsforthefutureofOSS.OSShasreceivedgrowingattention(LernerandTirole,2002)asithasbecomeanincreasinglycriticalpartofthemoderneconomy,tothepointwhere96%ofcorporatecodebasescontainsomeopensourcecode(Synopsys,

帶來(lái)了廣泛的生產(chǎn)力效益。(Czarnitzki,F(xiàn)ernández和Rammer,2023年)特別是與本研究相關(guān),針對(duì)Copilot的具體研究要么是在企業(yè)內(nèi)部使用較小的工人樣本進(jìn)行的年費(fèi)訪問(wèn)Copilot的貢獻(xiàn)者(Yeverechyahu,Mayya和Oestreicher?Singer,2024年)。(Acemoglu,2003年)。通過(guò)超越生產(chǎn)力,探討技術(shù)如何改變工作的本質(zhì),我們提供了一項(xiàng)關(guān)于生成式人工智能及其對(duì)高度細(xì)分的“野外”工作流程影響的兩項(xiàng)最大自然實(shí)我們的主要發(fā)現(xiàn)確定了AI采用者在他們的知識(shí)工作流程中工作本質(zhì)的變化。我們表明,當(dāng)軟件開(kāi)發(fā)人員更多地利用AI時(shí),他們將他們的努力重新分配到技術(shù)編碼活動(dòng),而不是涉及與其他開(kāi)發(fā)人員社交互動(dòng)的輔助項(xiàng)目管理活動(dòng)。這表明,當(dāng)利用像生成式AI這編程LLM引入后的兩年內(nèi)工作活動(dòng)和工作互動(dòng)過(guò)程的變化,補(bǔ)充了當(dāng)前文獻(xiàn),該文獻(xiàn)利ITAI,并通過(guò)實(shí)驗(yàn)關(guān)注高級(jí)生產(chǎn)力影響(Brynjolfsson,Li和Raymond,2023年;Dell’Acqua等人,2023年)除了識(shí)別生成式AI對(duì)去中心化工作產(chǎn)生的因果效應(yīng)之外,我們的研究結(jié)果對(duì)開(kāi)源軟的關(guān)注(Lerner和Tirole,2002),以至于96%的企業(yè)代碼庫(kù)中包含一些開(kāi)源代碼Further,recentstudiesestimatethevalueofOSStobeontheorderofbillionsofdollarsforthesupplyside(Blindetal.,2021;Robbinsetal.,2021)andtrillionsofdollarswhenaccount-ingforusage(Hoffmann,Nagle,andZhou,2024).Additionally,firmusageof,andcontributionto,OSShasimportantimplicationsforfirmproductivity(Nagle,2018,2019),firmcompetition(Boysel,Hoffmann,andNagle,2024)andentrepreneurialactivity(Wright,Nagle,andGreenstein,2023).However,despitetheimportanceofOSS,manycriticalprojectsareunder-resourced(Egh-bal,2020;Nagleetal.,2020b)asnumerousfirmsfree-rideontheeffortsofotherswithoutgivingback(Lifshitz-AssafandNagle,2021)leavingvolunteerdevelopersburntoutandoverwhelmed(Ramanetal.,2020).Asourresultsshow,generativeAImayofferasolutiontohelpaddresstheseconcernsandallowtopdeveloperstomoreeasilycontributetothecommongoodbysolvingmoreissues.PriorresearchhasshownthatOSSdevelopersgenerallycontributetoOSSbecauseitgivesthemacreativeoutletandtheydonotwanttospendtheirtimeonmanagerialtaskslikesecurityanddocumentation(Nagleetal.,2020a).AI-poweredtoolsmaymakeiteasiertoquicklyaddresssuchmanagerialtasks,sodeveloperscanspendtimeinamannertheyprefer,whilestillensuringthesecurity,stability,andusabilityofOSS.Theremainderofthispaperproceedsasfollows.Section1developsamodeloftheimpactofgenerativeAIonindividualworkersleadingtotestablehypotheses.InSection2,wediscusstheenvironmentwithinwhichthestudyoccurs.InSection3wecharacterizeourdatasetanddiscusstheconstructionofoursample.WehoneintothesetofdevelopersthatobtainCopiloteligibilityforfreeviaaninternalrankingfromGitHubandpresentourestimationstrategyinSection4.Wethenpresentourresultsusingaregressiondiscontinuitydesign(Section5)whilealsoexploringthemechanismsatplay,andofferingempiricalsupportforourhypotheses.Wediscussthelimitations,implications,andaback-of-the-envelopecalculationtounderstandhowtheresultsarelikelytogeneralizebeyondourempiricalsettinginSection6.Section7concludes.

此外,最近的研究估計(jì)開(kāi)源軟件(OSS)對(duì)供應(yīng)方的價(jià)值約為數(shù)十億美元(Blind2021年;Robbins等人,2021年),而當(dāng)考慮到使用時(shí),其價(jià)值則達(dá)到數(shù)萬(wàn)億美元(Hoffmann、Nagle和Zhou,2024年)。此外,對(duì)OSS的使用和貢獻(xiàn)對(duì)企業(yè)的生產(chǎn)力(Nagle,2018年,2019年)、企業(yè)競(jìng)爭(zhēng)(Boysel、Hoffmann和Nagle,2024年)和創(chuàng)業(yè)活動(dòng)(Wright、Nagle和Greenstein,2023年)具有重要意義。然而,盡管OSS的重要性不言而喻,許多關(guān)鍵項(xiàng)目卻資源不足(Eghbal,2020年;Nagle等人,2020年b),因?yàn)樵S多企業(yè)免費(fèi)搭乘他人的努力,而不給予回報(bào)(Lifshitz?AssafNagle,2021),(Raman等人,2020年)。正如我們的研究結(jié)果所示,生成式AI可能為解決這些問(wèn)題獻(xiàn)。先前的研究表明,開(kāi)源軟件(OSS)的開(kāi)發(fā)者通常因?yàn)殚_(kāi)源軟件為他們提供了一個(gè)創(chuàng)意出口,他們不想把時(shí)間花在管理任務(wù)上,如安全和文檔(Nagle2020a)。AI驅(qū)動(dòng)的工具可能使快速解決此類管理任務(wù)變得更加容易,這樣開(kāi)發(fā)者就可以以他們喜本文的其余部分如下進(jìn)行。第1節(jié)1構(gòu)建了一個(gè)生成式AI對(duì)個(gè)體工人影響的模型,2233中,我44GitHubCopilot不連續(xù)設(shè)計(jì)(55)來(lái)展示我們的結(jié)果,同時(shí)探討其中的機(jī)制,并為我們的假設(shè)提供66中,我們討論了局限性、影響以及一個(gè)估算,以了解結(jié)果如何可能推廣到我們的實(shí)證環(huán)境之外。第7節(jié)7得出結(jié)論。TheoreticalIntheknowledgeeconomy-whichisanincreasinglylargesectoroftheoveralleconomy-,highlyproductiveindividualscanoftenbecomevictimsoftheirownsuccess.Acommonpatternrelevanttoourstudyoccurswhenadeveloperdoesexceptionalcorework,theyareoftenassignedmoremanagerialworkasaresult.Forexample,inthecontextofacademia,whereresearchandteachingarecorework,theresultofdoingagoodjobontheseistogetpromotedandthentobegivenmoremanagerialtasksincludingdepartmentandschoolcommitteeassignments.Thiscanbesummedupbytweakingthewell-knownphrase“Therewardforgoodworkismorework.”tobe“Therewardforgoodcoreworkismoremanagerialwork.”Thisisparticularlytrueinthecontextofpublicgoodswhich,aspublicgoodprojectsbecomemoresuccessfulandmorewidelyused,newusersrequestmorefromthosethatarecreatingthegood.3Thus,theintroductionofanAItoolthatcanhelpreducesomeofthisburdenmayplayanimportantroleinthecreationofpublicInthefollowingsection,wedeveloptheexpositionofourempiricalsettingbyusingasimpleeconomicframeworkwhereindividualworkerschoosebetweentwoactivitiestomaximizeutility:coreworkcandprojectmanagementm.Lettheworker’spreferencesuθ(·)beindexedtheparametervectorθ.Ineachperiod,eachworkerchoosescandmtosolvethefollowingstaticutilitymaximizationproblem:

Theoretical稍作修改來(lái)概括,即出色核心工作的回報(bào)是更多的管理工作。這在公共物品的背景下尤其如的人提出更多要求。因此,引入一個(gè)可以幫助減輕一些這種負(fù)擔(dān)的AI工具可能在公共物好(·)由參數(shù)向量(θ)索引。在每個(gè)時(shí)期,每個(gè)工人選擇cm來(lái)解決以下靜態(tài)效c,

uθ(c,

c,

uθ(c,

subject pcc+pmm≤wherec,m≥0andpc,pm>0.Thechoiceisconstrainedbyrelativecostsofeach

subject pcc+pmm≤cm≥0pcpm>0。選擇受限于每種活動(dòng)的相對(duì)成本,p=(pcpmp=(pc,

),andunitsofanendowmentresource,ω.4Inlinewithsimpleeconomicmodels,

assumethatpreferencesaretime-invariantandthatthereareno3InourempiricalcontextofOSS,this“burden”ofbeinganopensourcedeveloper(Geerling,2022)hasbeencitedassignificantdriverofburnoutandabandonmentofopensourcedevelopment(Nagleetal.,2020a;Ramanetal.,2020).Thus,alleviatingthisburdenisofcriticalimportance.4Inoursetting,theresourceendowmentωcanbeinterpretedastheagent’s“taskbandwidth”theyareabletoacrossvariouswork

倦怠和放棄開(kāi)源開(kāi)發(fā)的顯著驅(qū)動(dòng)因素(Nagle等人,2020a;Raman等人,2020)。因此,減輕這種負(fù)擔(dān)至關(guān)重要。在我們的設(shè)置中,資源稟賦4σ+β1/σ uθ(c,m)β1/σσ+β1/σσ+β1/σ uθ(c,m)β1/σσ+β1/σ

uθ(c,uθ(c,m)β1/σ

σ?1

σ?1 whereforθ=(σ,βc,βm),σistheelasticityofsubstitutionbetweencandm,andβc,βmareCESshareparameters.Withoutlossofgenerality,afternormalizingpm=1,pcbecomestherelativecostofdoingcorework.Undertheoptimalchoiceofthesetwoactivities,theMarshallianmandsforcoreworkandprojectmanagementcanbeexpressedasfunctionsoftheseproductivity,preference,andendowmentparameters:

其中對(duì)于θ(σ,βc,βmσ是c和m之間的替代彈性,而βc,βm是CES分享參數(shù)。不失一般性,在標(biāo)準(zhǔn)化pm=1后,pc成為核心工作的相對(duì)成本。在這些兩種活動(dòng)的最優(yōu)選 cp1?σ+

c?

p1?σ+βm

m?

m?

βcp1?σ+ βcp1?σ+βm βmConsistentwithpriorliterature(Acemoglu,Kong,andRestrepo,2024),wechoosetotheinterventionofgenerativeAIasareductioninthecostofcorework,pc.Assuch,thecompar-ativestaticswithrespecttopcareofinterest.DetailsonthecomparativestaticsforachangeinpccanbefoundinAppendixD.AconsequenceoftheCESdemandsystemisthatareductioninpcincreasestheoptimallevelofcoreworkunderanyvalueoftheelasticityofsubstitutionσ>0.Furtherempiricalsupportforthisrelationshipcomesfrompriorliteratureinthefield.BeyondAI,automationandinformationsystemstechnologieshavebeenshowntocomplementskilledlaborandleadtoareshapingoforganizationalpracticesthatallowsworkerstoengageinmorecom-plexandstrategicactivities(Autor,Levy,andMurnane,2003;Orlikowski,2007;Zammutoetal.,2007).Further,whentechnologyreducesthecostoreffortassociatedwithcertaintasks,economicandmanagementtheorysuggeststhatworkerswillincreasetheamountofthattasktheyperform(AcemogluandRestrepo,2018;Bloometal.,2014).Assuch,wearriveatthefollowingprimary

建模為降低核心工作的成本,pcpc的比較靜態(tài)分析是有趣的。關(guān)于pc變化的比較靜態(tài)分析細(xì)節(jié)可以在附錄D中找到。CES需求系統(tǒng)的結(jié)果是,降低pc會(huì)增加在任何替代彈性σ>0值下的核心工作的最優(yōu)水平。這一關(guān)系的進(jìn)一步經(jīng)驗(yàn)支持來(lái)自該領(lǐng)域的先前文獻(xiàn)。除了AI之外,自動(dòng)化和信息系統(tǒng)技術(shù)已被證明可以補(bǔ)充熟練勞動(dòng)力,并導(dǎo)致組織實(shí)踐的重新塑造,使工人能夠參與更復(fù)雜和戰(zhàn)略性的活動(dòng)(AutorLevyandMurnane,Hypothesis1a(H1a)AftertheadoptionofanAItoolthatassistswithcoreworktasks,aworker’scoreworktasksincreaseasapercentageofalltasks.Incontrasttotheimpactoncoreworktasks,theimpactofgenerativeAIonmanagerialtasksislessclearanddependentontheelasticityofsubstitutionσ.Adoptionofthetoolmayleadtonochangeintheshareofprojectmanagementwhentheelasticityofsubstitutionσ=1.Alternatively,projectmanagementmaydropwhenthepriceofcoreworkdropsgivenaσ>1(projectmanage-mentisasubstitute),ormayincreasewhen0<σ<1(projectmanagementisacomplement).Thisisconsistentwithpriorliteraturethathasshownthatwhileautomationandtechnologytendtoreducetheburdenofroutinetasks,theydonotnecessarilyeliminatemanagerialresponsibilities,whichmayrequirehumanjudgment,creativity,andinterpersonalcoordination(Autor,Levy,andMurnane,2003;Mintzberg,1994).Consequently,evenasAIcanreducethetimespentonroutinetasks,workersmaystillengageinhigh-leveldecision-makingandteamleadership,leavingtheneteffectonmanagerialtasksuncertainandbestdeterminedempirically.Hypothesis1b(H1b)AftertheadoptionofanAItoolthatassistswithcoreworktasks,thechangetoaworker’smanagerialtasksasapercentageofalltasksisambiguous.Wenextseektobetterunderstandthemechanismsthataredrivingtheseeffects.WhatistheeffectofAItechnologyontaskallocationacrossspecifickindsofcoreworkandprojectmanage-ment?Tothisend,weextendthebaseline2-goodCESmodelintoanestedCESmodel,underwhichcoreworkandprojectmanagementareinsteadmodeledascompositesofmoredisaggre-gatedgoods.

1a(H1a)AI工具后,工人的核心工作任務(wù)占所有任與對(duì)核心工作任務(wù)的影響相比,生成式AI對(duì)管理任務(wù)的影響不太明確,并且取決于替代彈性的大小σ。當(dāng)替代彈性σ=1時(shí),采用該工具可能導(dǎo)致項(xiàng)目管理份額不變。另一方面,當(dāng)核心工作價(jià)格下降時(shí),由于σ>1(項(xiàng)目管理是替代品),項(xiàng)目管理可能會(huì)下0<σ<1(項(xiàng)目管理是互補(bǔ)品)時(shí)可能會(huì)增加。這與先前的研究結(jié)果一致,責(zé)任,這可能需要人類的判斷、創(chuàng)造力和人際協(xié)調(diào)Autor,Levy和Murnane,2003;1b(H1b)AI工具后,工人管理任務(wù)占所有任務(wù)的百接下來(lái),我們?cè)噲D更好地理解驅(qū)動(dòng)這些效應(yīng)的機(jī)制。AI技術(shù)對(duì)特定類型的核心工作2CESCESu(c,u(c,c,m,m)β1/σu(c,c

σ?1

σ?1 σ+β1/σu(m,m u(c,c,m,m)βσ+β1/σu(m,m u(c,c,m,m)β1/σu(c,cσ+β1/σu(m,m u(m1,m2)arealsoCESfunctionssimilartoequation2butwiththeirrespectivewithin-nest

u(c1c2)u(m1m2)也是類似于方程2的CES函數(shù),但它們的各自嵌套內(nèi)的替代彈性σc和σm與細(xì)分商品之ticitiesofsubstitution

and

thatcorrespondtorelativesubstitutionbetween

goodsc1,c2andm1,m2respectively.HencethenestedCESextensiontothebaselinemodelper-mitsbothmorerefineddefinitionsofworkpatternsandrichersubstitutionpatternsbetweenthesedisaggregatedgoods.DetailsonthefullnestedCESmodel,aswellasthecomparativestaticsachangeinpccanbefoundinAppendixWeusethismodeltoconsidertwomechanismsthroughwhichtheprimaryrelationshipoper-ates.Inthefirstmechanism,weconsiderwhetherworkersengageinworkthatismoreautonomous(lessinteractionwithothersworkingontheproject)ormorecollaborative(moreinteractionwithothersworkingontheproject).Individualscanengageineitherautonomouscorework,c1or

CESCES模型以及pc變化的比較靜態(tài)分析,可以在附錄D中找到。員互動(dòng)較多)的工作。個(gè)人可以從事自主核心工作,c1或協(xié)作核心工作,c2或管理等效工作,m1和m2。我們發(fā)現(xiàn),通過(guò)AI降低核心工作成本,pc可以提高核心工作的需求laborativecorework,c2orthemanagerialequivalents,m1andm2,Wefindthatareduction

1

<1thecostofcoreworkthroughAI,pccanincreasethedemandofcorework(asinHypothesis

butitdoesnotnecessarilyneedtohappenthroughbothautonomouscoreworkand

<1coreworksimultaneously.Indeed,assumingthattheelasticityofsubstitutionσc>1andthat

1(以及

priceofautonomousworkislowerthanthepriceofcollaborativework,

<1impliesthat

workerwillshifttheireffortstowardsautonomouscoreworkandawayfromcollaborativecoreworksinceautonomouscoreworkislesscostlythancollaborativecorework.Thesame

trueformanagerialworksuchthat

1and

<1.Whiletherearereasonstofindalternativeparameterspaces,

1(and

1)arecredible,wefindthisrestrictedspacewiththepre-existingwedgeofpricesgenerallyplausibleinthecontextofworkersthatarealreadyworkinginahighlycollaborativesettingliketheincreasinglycommonparadigmofdis-tributedwork.Wehypothesizethattheirmainissues—collaborativefrictionssuchasthecostofcoordination,requestsfromotherstosolveproblems,orpersonalconflicts—maybemorecostlythansolvingproblemsbythemselveswhentheyhaveAIasasubstituteavailableatanytime.Thepredictionsofthenestedmodelextensioncansimilarlybederivedfromtheliterature.ThismechanismbuildsontheideathatgenerativeAItoolsreduce(oreveneliminate)muchofthecognitiveandcommunicativefrictioninherentindistributedwork,enablingworkerstotacklecomplextasksautonomously.Priorresearchhasshownthattechnologiesthatstreamlinecommu-nicationanddecision-makingprocessesreducetheoverheadofcollaboration,freeingworkers

嵌套模型擴(kuò)展的預(yù)測(cè)可以從文獻(xiàn)中類似地推導(dǎo)出來(lái)。這種

溫馨提示

  • 1. 本站所有資源如無(wú)特殊說(shuō)明,都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請(qǐng)下載最新的WinRAR軟件解壓。
  • 2. 本站的文檔不包含任何第三方提供的附件圖紙等,如果需要附件,請(qǐng)聯(lián)系上傳者。文件的所有權(quán)益歸上傳用戶所有。
  • 3. 本站RAR壓縮包中若帶圖紙,網(wǎng)頁(yè)內(nèi)容里面會(huì)有圖紙預(yù)覽,若沒(méi)有圖紙預(yù)覽就沒(méi)有圖紙。
  • 4. 未經(jīng)權(quán)益所有人同意不得將文件中的內(nèi)容挪作商業(yè)或盈利用途。
  • 5. 人人文庫(kù)網(wǎng)僅提供信息存儲(chǔ)空間,僅對(duì)用戶上傳內(nèi)容的表現(xiàn)方式做保護(hù)處理,對(duì)用戶上傳分享的文檔內(nèi)容本身不做任何修改或編輯,并不能對(duì)任何下載內(nèi)容負(fù)責(zé)。
  • 6. 下載文件中如有侵權(quán)或不適當(dāng)內(nèi)容,請(qǐng)與我們聯(lián)系,我們立即糾正。
  • 7. 本站不保證下載資源的準(zhǔn)確性、安全性和完整性, 同時(shí)也不承擔(dān)用戶因使用這些下載資源對(duì)自己和他人造成任何形式的傷害或損失。

最新文檔

評(píng)論

0/150

提交評(píng)論