2024哈佛大學(xué)生成式 AI新興技術(shù)風(fēng)險(xiǎn)和初學(xué)者 AI 風(fēng)險(xiǎn)緩解策略

Don’tExpectJuniorstoTeachSeniorProfessionalstoUseGenerativeAI:EmergingTechnologyRisksandNoviceAIRiskMitigationTactics

用生成式AI：新興技術(shù)風(fēng)險(xiǎn)和初學(xué)者AI風(fēng)險(xiǎn)緩解策略Don’tExpectJuniorstoTeachSeniorProfessionalstoUseGenerativeAI:EmergingTechnologyRisksandNoviceAIRiskMitigationTactics

用生成式AI：新興技術(shù)風(fēng)險(xiǎn)及新手AI風(fēng)險(xiǎn)緩解策略Don’tExpectJuniorstoTeachSeniorProfessionalstoUseGenerativeAI:EmergingTechnologyRisksandNoviceAIRiskMitigationTacticsTheliteratureoncommunitiesofpracticedemonstratesthataprovenwayforseniorprofessionalstoupskillthemselvesintheuseofnewtechnologiesthatundermineexistingexpertiseistolearnfromjuniorprofessionals.Itnotesthatjuniorsmaybebetterablethanseniorstoengageinreal-timeexperimentationclosetotheworkitself,andmaybemorewillingtolearninnovativemethodsthatconflictwithtraditionalidentitiesandnorms.However,thisliteraturehasnotexploredemergingtechnologies,whichareseentoposenewriskstovaluedoutcomesbecauseoftheiruncertainandwide-rangingcapabilities,exponentialrateofchange,potentialforoutperforminghumansinawidevarietyofskilledandcognitivetasks,anddependenceonavast,varied,andhighvolumeofdataandotherinputsfromabroadecosystemofactors.Ithasalsonotexploredobstaclestojuniorprofessionalsbeingasourceofexpertiseintheuseofnewtechnologiesformoreseniormembersincontextswherethejuniorsthemselvesarenottechnicalexperts,andwheretechnologyissonewandrapidlychangingthatthejuniorshavehadlittleexperiencewithusingit.However,suchcontextsmaybeincreasinglycommon.InourstudyconductedwithBostonConsultingGroup,aglobalmanagementconsultingfirm,weinterviewed78suchjuniorconsultantsinJuly-August2023whohadrecentlyparticipatedinafieldexperimentthatgavethemaccesstogenerativeAI(GPT-4)forabusinessproblemsolvingtask.Drawingfromjuniorprofessionals’insitureflectionssoonaftertheexperiment,wearguethatsuchjuniorsmayfailtobeasourceofexpertiseintheuseofemergingtechnologiesformoreseniorprofessionals;instead,theymayrecommendthreekindsofnoviceAIriskmitigationtacticsthat:1)aregroundedinalackofdeepunderstandingoftheemergingtechnology’scapabilities,2)focusonchangetohumanroutinesratherthansystemdesign,and3)focusoninterventionsattheproject-levelratherthansystemdeployer-orecosystem-level.

別指望初級(jí)員工教資深專業(yè)人士使用生成式AI：新興技術(shù)風(fēng)險(xiǎn)及新手AIEmergingTechnologyRisksandNoviceAIRiskMitigation在我們與全球管理咨詢公司波士頓咨詢集團(tuán)合作進(jìn)行的研究中，我們于2023年7月至8月采訪了78名這樣的初級(jí)顧問，他們?cè)谧罱鼌⒓拥囊豁?xiàng)實(shí)驗(yàn)中獲得了使用生成式AI（GPT?4）進(jìn)行議三種類型的AI風(fēng)險(xiǎn)緩解策略：1）基于對(duì)新興技術(shù)能力的缺乏深入了解，2）關(guān)注改變?nèi)祟惓Ｒ?guī)而不是系統(tǒng)設(shè)計(jì)，3）關(guān)注項(xiàng)目層面的干預(yù)而不是系統(tǒng)部署或生態(tài)系統(tǒng)層面的干預(yù)。Inthecontemporaryenvironmentofacceleratingtechnologicalchange,seniorprofessionalsfacethedualresponsibilityofquicklyimplementingemergingtechnologiestoday(e.g.,CameronandRahman,2022;Karunakaran,2024;Lebovitz,Levina,andLifshitz,2022;Mazmanian,Orlikowski,andYates,2013;RahmanandValentine,2021;Waardenburg,HuysmanandSergeeva,2022),andanticipatingfutureversionsoftechnologiesandtheirimplicationsforboththeirclientsandtheirownorganizations(e.g.,Barrettetal.,2012;Endacott&Leonardi,2021;Karunakaran,OrlikowskiandScott,2022;Lebovitz,Levina,andLifshitz,2021;PineandMazmanian,2015;Sendaketal.,2020).Toleadtheirteamsandorganizationsingrapplingwitharapidlyexpandingtechnologicalfrontier(Dell’Acquaetal.,2023),seniorsneedtodevelopadeepunderstandingofnewtechnologiesandtheirassociatedcapabilities(e.g.,BeaneandAnthony,2023).Theliteratureoncommunitiesofpracticestellsusthat,whenrapidtechnologicalchangeandshiftsinthenatureofworkundermineexistingexpertise(e.g.,Beane,2019;Bharatan,SwanandOborn,2022;ChanandHedden,2023),thisposesaproblemforseniormembersofcommunitiesofpracticewhohavetraditionallyengagedinsituatedlearningovertimetobecomehighlyskilledintheuseofthetechnologiestheirworkrequires(e.g.,Barley,1986;BrownandDuguid,1991;LaveandWenger,1991;Orr,1990).Seniorprofessionalsareoftenreluctanttotaketimeawayfromdealingwithcomplexproblemstoexperimentwithnewtechnology(Anthony,2018;Zetka,2003).Inaddition,becauseoftheirseniorposition,theyareexpectedtobeexperiencedhandswhoknowhowtoperformthetasksrequiredbyprofessionalsintheirrolebetterthanthosemorejunior,sodemonstratinglackofexpertiseduringlearningcouldthreatentheirimagewithclientsandjuniorprofessionals(Beane,2019).Finally,seniorprofessionalsunderstandthesocialrules,norms,andvaluesthroughwhichaperson’sworthinesstothegroupisjudgedbymembersofthatgroupbetterthanthosemorejuniordo(e.g.,Bechky,2006;Lee,MazmanianandPerlow,2020),soseniorsmaybelesswillingtodeviatefromtraditionalnormstosupportthevaluesconsistentwithnewmethods(Kelloggetal.,2021).Thus,seniorprofessionalsmayattempttolearnfromjuniorprofessionals,whomaybebetterablethantheirseniorcounterpartstolearntoeffectivelyusethenewtechnologiesforseveralreasons(BeaneandAnthony,2024).First,juniorprofessionalsareoftenclosesttotheworkitself,becausetheyaretheonesengaginginconcreteandlesscomplextasks(e.g.,Bharatan,Oborn,andSwan,2024;Karunakaran,2018;PineandMazmanian,2017;RahmanandBarley,2017).Second,juniorprofessionalsmaybemoreabletoengageinreal-timeexperimentationwithnewtechnologies,becausetheydonotrisklosingtheirmandatetoleadifthosearoundthem,includingclients,aswellasthosemorejuniortothem,recognizethattheylackthepracticalexpertisetosupporttheirhierarchicalposition(e.g.,Beane,2019;Endacott&Leonardi,2022;Leonardi,2007).Third,juniorprofessionalsmaybemorewillingtolearnnewmethodsthatconflictwithexistingidentities(e.g.,NelsonandIrwin2014;Lifshitz,2018),practices(e.g.,BeaneandOrlikowski,2015;MazmanianandBeckman,2018)andframes(e.g.,AnthonyandTripsas2016;Mazmanian,2013).Forexample,whenCTscannersarrivedinradiology,seniorradiologistslearnedtoeffectivelyusethemfromjuniortechnicianswhowereclosesttotheworkofinjectingdyesandconstructingimages(Barley,1986).Whenhelp-deskqueuingtechnologyarrivedinIT,seniortechniciansrespondingtohelp-deskcallsbegantoaskforhelpfromjuniortechnicianswhohadengagedinexperimentationwiththenewtechnologyanddevelopedskillinitsuseandfeatures(Leonardi,2007).Andwhenpatientreferralandtrackingtechnologyarrivedinprimarycare,seniormedicalassistantslearnedfromthosemorejuniorhowtofolloweachcolonoscopypatient’sprogressthroughprocedureandfollow-up,becausethejuniorswhoweremorewillingtolearnnewmethodsthatconflictedwithexistingidentities,norms,andframes(Kelloggetal.,2021).Yet,whilejuniorprofessionals’teachingoftheeffectiveuseofnewtechnologytomoreseniorprofessionalscanbesuccessful,itseffectsarefarfrominevitableoruniform.Thus,scholarsoftechnology,work,andoccupationshaveexploredthequestionofhowandwhenjuniorprofessionalsmayfailtobeasourceofexpertiseintheuseofnewtechnologiesformoreseniormembers.Theirstudies

（例如，CameronRahman，2022Karunakaran，2024Lebovitz，LevinaLifshitz，2022；Mazmanian，Orlikowski和Yates，2013；Rahman和Valentine，2021；Waardenburg，Huysman和Sergeeva，2022），二是預(yù)測(cè)未來技術(shù)的版本及其對(duì)客戶和自身組織的影響（例如，Barrett2012EndacottLeonardi，2021Karunakaran，OrlikowskiScott，2022Lebovitz，LevinaLifshitz，2021PineMazmanian，2015Sendak等人，2020）。為了領(lǐng)導(dǎo)他們的團(tuán)隊(duì)和組織應(yīng)對(duì)快速發(fā)展的技術(shù)前沿（Dell’Acqua等人，2023），高級(jí)專業(yè)人士需要深入理解新技術(shù)及其相關(guān)能力（例如，Beane和Anthony，2023）。Beane，2019；Bharatan，Swan和Oborn，2022；Chan和Hedden，2023）時(shí)，這給那Barley，1986；Brown和Duguid，1991；Lave和Wenger，1991；Orr，1990）。高級(jí)專業(yè)人士通常不愿意從處理復(fù)雜問題的任務(wù)中抽出時(shí)間來嘗試新技術(shù)（Anthony，2018；Zetka，2003）。此外，由于他們的高級(jí)職位，人們期望他們比那些職位較低的人更了解如何完成其角人士眼中的形象（Beane，2019）。最后，高級(jí)專業(yè)人士比那些職位較低的人更了解通過哪些社會(huì)規(guī)則、規(guī)范和價(jià)值觀來判斷一個(gè)人對(duì)群體的價(jià)值（例如，Bechky，2006；Lee，Mazmanian和Perlow，2020），因此他們可能不太愿意偏離傳統(tǒng)規(guī)范以支持與新方法一致的價(jià)值（Kellogg等人，2021）。具體且較為簡(jiǎn)單的任務(wù)的人（例如，Bharatan、Oborn和Swan，2024年；Karunakaran，2018年；Pine和Mazmanian，2017年；Rahman和Barley，2017年）。其次，初級(jí)專業(yè)人士Endacott和Leonardi，2022年；Leonardi，2007年）。第三，初級(jí)專業(yè)人士可能更愿意學(xué)習(xí)與現(xiàn)有身份（例如，Nelson和Irwin，2014年；Lifshitz，2018年）、實(shí)踐（例如，Beane和Orlikowski，2015年；Mazmanian和Beckman，2018年）和框架（例如，AnthonyTripsas，2016Mazmanian，2013）相沖突的新方法。例如，當(dāng)CT掃描儀進(jìn)入放射科時(shí)，高級(jí)放射科醫(yī)生從與注射染料和構(gòu)建圖像的工作最為接近的初級(jí)技術(shù)人員那里學(xué)習(xí)如何有效地使用它們（Barley，1986年）。當(dāng)幫助臺(tái)排隊(duì)技術(shù)進(jìn)入IT時(shí)，響應(yīng)幫助臺(tái)電話的高級(jí)技術(shù)人員開始向已經(jīng)對(duì)新技術(shù)進(jìn)行實(shí)驗(yàn)并掌握了其使用和功能的初級(jí)技術(shù)人員尋求幫助（Leonardi，2007年）。當(dāng)患者轉(zhuǎn)診和跟蹤技術(shù)進(jìn)入初級(jí)保健時(shí)，高級(jí)醫(yī)療助理從那些更愿意學(xué)習(xí)與現(xiàn)有身份、規(guī)范和框架相沖突的新方法的初級(jí)人員那里學(xué)習(xí)如何跟蹤每位結(jié)腸鏡檢查患者的手術(shù)和隨訪進(jìn)展（Kellogg等人，2021年）。不要期望初級(jí)專業(yè)人士教給資深專業(yè)人士如何使用生成式AI：新興技術(shù)風(fēng)險(xiǎn)及新手AI風(fēng)險(xiǎn)showthatstatusthreatisoftenthekeyobstacletojuniorprofessionalssuccessfullycoachingmoreseniorprofessionalsintheeffectiveuseofnewtechnology(e.g.,BeaneandAnthony,2024;Barley,1986).Juniorscoachingseniorsinnewtechnologyusechallengesseniormembers’status,whichisbasedonthehistoricaldistinctionsoftheperformanceofhigherlevel,morecomplextasks(e.g.,Leonardi&Bailey,2008;VanMaanenandSchein,1978),thedemonstrationofexpertiseinperformingthesetasks(e.g.,Anthony,2021;Bharatan,Swan&Oborn,2022),andtheenactmentoftraditionalidentities,frames,andtemporalrhythmswhileperformingthetasks(e.g.,Bechky,2003;Lifshitz,2018;ObornandBarrett,Thisliteraturehasbeencriticalinhighlightingthat,whenrapidtechnologicalchangeandshiftsinthenatureofworkundermineexistingexpertise,seniorprofessionalswhoneedtorapidlyupskillthemselvesinnewtechnologiesmayturntojuniorprofessionals.However,thisliteraturehasnotexploredtechnologiessuchasartificialintelligence,dataanalytics,socialmedia,digitalplatforms,blockchain,and3-Dprinting(e.g.,Polykarpou,Barrett,andOborn,2020;Aroraetal.,2023;Rahman,KarunakaranandCameron,2024),whichscholarshavelabeled“emergingtechnologies,”(e.g.,Baileyetal.,2022)becauseoftheiruncertainandwide-rangingcapabilities,exponentialrateofchange,potentialforoutperforminghumansinawidevarietyofskilledandcognitivetasks,anddependenceonavast,varied,andhighvolumeofdataandotherinputsfromabroadecosystemofactors.Theseemergingtechnologiesareseentobeassociatednotonlywiththeriskofnegativeoutcomes,butalsowithrisksthatarenovel,unfamiliar,andinvolveconsiderableuncertainty(Aroraetal.,2023;Barrettetal.,2024).Giventherapidintroductionofandchangeintheseemergingtechnologies,situationsinwhichseniorprofessionalsmayneedtolearntoeffectivelyuseemergingtechnologiesassociatedwithnovelandunfamiliarrisksareincreasinglycommon(Baileyetal.,2022).Forexample,seniorprofessionalsneedtolearntomitigaterisksassociatedwithreducedcontrolthroughthephasesoftheserviceencounterastheylearntoeffectivelyuseplatformtechnologies(RahmanandValentine,2021;Cameron&Rahman,2022).Theyneedtolearntomitigaterisksassociatedwithreliabilitybreakdownsastheylearntoeffectivelyusecloud-basedtechnologies(Karunakaran,2021).TheyneedtolearntomitigaterisksassociatedwithopacityastheylearntoeffectivelyuseAI-basedtechnologies(Berenteetal.,2021;Lebovitz,Levina&Lifshitz,2021;Lebovitz,Lifshitz&Levina,2022).Andtheyneedtolearntomitigaterisksassociatedwithbigdatasurveillance(Brayne,2017;Rahman,KarunakaranandCameron,2024)andopportunisticdatacuratorsastheylearntoeffectivelyusepredictiveanalytictechnologies(Waardenburgetal.,2022).Astheylearntoeffectivelyusethesetechnologies,theyarelikelytoturntothejuniorswithwhomtheyworkforhelp.Thismaybeparticularlyproblematicinthecaseofemergingtechnologies.Thecurrentliteratureonjuniorprofessionalsbeingasourceofexpertiseintheuseofnewtechnologiesformoreseniormembershasnotexploredcontextswherethejuniorsthemselvesarenottechnicalexperts,andwheretechnologyissonewandrapidlychangingthatjuniorshavehadnoformaltrainingonhowtousethetechnology,havehadnoexperiencewithusingitintheworksetting,andhavehadlittleexperiencewithusingitoutsideoftheworksetting.Insuchcontexts,itseemsunreasonabletoexpectthatjuniorsshouldhaveadeepleveloftechnicalunderstandingofthetechnology.Yet,suchcontexts—inwhichseniorsneedtolearntouseemergingtechnologiesassociatedwithnovelandunfamiliarrisks,andinwhichthejuniorstheseseniorsworkwithmaygainaccesstothetechnologyevenwhenthesejuniorsdonothavetechnicalbackgroundsandhavehadlittletrainingorpracticalexperiencewiththetechnology—maybeincreasinglycommon.Emergingtechnologiesassociatedwithnovelandunfamiliarrisksareincreasinglybeingintroducedintoorganizations(e.g.,Baileyetal.,2022).AndemergingtechnologiessuchasgenerativeAIcanbeeasilyaccessedandcustomizedwithoutcodingandwithoutowninginfrastructure,soareavailabletousersofvaryingskillsandtechnicalbackgrounds(e.g.,Schneideretal.,2024).Aswewillshowbelow,thejuniorsinourstudyexpectedthattheywouldneedtoeducateseniorsabouteffectiveuseofthenewtechnology.Thus,wesuggestthattheresearchquestionofhowandwhenjuniorprofessionalsmayfailtobeasourceofexpertiseintheuseofnewtechnologiesformoreseniormembersneedstobeexploredwithinsucha

展示出地位威脅往往是初級(jí)專業(yè)人士在有效使用新技術(shù)方面成功指導(dǎo)更資深專業(yè)人士的關(guān)鍵障礙（例如，Beane和Anthony，2024；Barley，1986）挑戰(zhàn)了資深成員的地位，這種地位基于歷史對(duì)更高層次、更復(fù)雜任務(wù)的績效區(qū)分（例如，Leonardi&Bailey，2008；VanMaanen和Schein，1978），在執(zhí)行這些任務(wù)中展示專業(yè)知識(shí)（例如，Anthony，2021；Bharatan，Swan&Oborn，2022），以及在執(zhí)行任務(wù)時(shí)實(shí)施傳統(tǒng)身份、框架討人工智能、數(shù)據(jù)分析、社交媒體、數(shù)字平臺(tái)、區(qū)塊鏈和3D打印等新興技術(shù)（例如，Polykarpou、Barrett和Oborn，2020；Arora等人，2023；Rahman、Karunakaran和Cameron，2024），學(xué)者們將這些技術(shù)稱為“新興技術(shù)”，（例如，Bailey等人，2022），興技術(shù)不僅與負(fù)面結(jié)果的風(fēng)險(xiǎn)相關(guān)，還與新穎、不熟悉且涉及相當(dāng)不確定性的風(fēng)險(xiǎn)相關(guān)（Arora等人，2023Barrett2024）。知風(fēng)險(xiǎn)相關(guān)的新興技術(shù)的情況越來越普遍（Bailey等人，2022年）。例如，高級(jí)專業(yè)人士需要學(xué)習(xí)在服務(wù)接觸的各個(gè)階段如何通過使用平臺(tái)技術(shù)來減輕控制減少的風(fēng)險(xiǎn)（Rahman和Valentine，2021年；Cameron&Rahman，2022年）。他們需要學(xué)習(xí)在有效使用基于云的技術(shù)時(shí)如何減輕可靠性故障的風(fēng)險(xiǎn)（Karunakaran，2021年）。他們需要學(xué)習(xí)在有效使用基AI（Berente2021Lebovitz，Levina&Lifshitz，2021年；Lebovitz，Lifshitz&Levina，2022年）。他們還需要學(xué)習(xí)在有效使用這些技術(shù)時(shí)如何減輕大數(shù)據(jù)監(jiān)控（Brayne，2017Rahman，KarunakaranCameron，2024年）和機(jī)會(huì)主義數(shù)據(jù)收集者的風(fēng)險(xiǎn)。當(dāng)他們學(xué)習(xí)如何有效使用這些技術(shù)時(shí)，他們很可能會(huì)并且他們一起工作的初級(jí)專業(yè)人士即使沒有技術(shù)背景，也沒有接受過多少培訓(xùn)或?qū)嵺`經(jīng)驗(yàn)，也能接觸到這項(xiàng)技術(shù)——可能越來越普遍。與新穎和未知風(fēng)險(xiǎn)相關(guān)的新興技術(shù)正越來越多地被引入組織（例如，Bailey等人，2022年）。而且，像生成式AI這樣的新興技術(shù)可以無需編碼和無需擁有基礎(chǔ)設(shè)施就能輕松訪問和定制，因此可供不同技能和技術(shù)背景的用戶使用（例如，Schneider等人，2024年）需要教育高級(jí)專業(yè)人士如何有效使用新技術(shù)。因此，我們建議在這樣一個(gè)背景下探討初級(jí)專業(yè)人士可能無法成為更資深成員使用新技術(shù)專家來源的如何和何時(shí)的問題。InourstudyconductedwithBostonConsultingGroup,aglobalmanagementconsultingfirm,weexploredthisquestioninanexperimentalsetting.Weinterviewed78juniorconsultantsinJuly-August2023whohadrecentlyparticipatedinafieldexperimentthatgavethemaccesstogenerativeAI(GPT-4)forabusinessproblemsolvingtask.Thesejuniorprofessionalshadbeenaskedtosolveabusinessproblem(toidentifychannelsandbrandsthatwouldhelpafictionalcompanyoptimizeitsrevenueandprofitability)usingfictitiousinterviewnoteswithcompanyexecutivesandhistoricalbusinessperformancedata.Soonaftertheyhadcompletedtheproblem-solvingtask,weconductedsemi-structured,60-minuteinterviews(Spradley1979)viaZoomwiththem.ThesejuniorswhohadbeengivenaccesstoGPT-4intheexperimentwerenottechnicalexperts,hadhadnoformaltrainingonhowtousethetechnology,hadnoexperiencewithusingitintheworksetting,andhadlittleexperiencewithusingitoutsideoftheworksetting.Inasetofearlybackgroundinterviewswithsimilarjuniorconsultants,wehadlearnedjuniorsthemselveswereexcitedtouseGenAIintheirwork,butthattheyanticipatedthattheywouldneedtoeducatetheirmanagersintheuseofGenAItoalleviatemanagers’concernsabouttherisksthatGenAIposedtooutcomesthatmanagersvalued.Tobetterunderstandthisphenomenon,inthe78interviewsweconductedinJuly-August2023,weaskedthesejuniorconsultants:1)CanyouenvisionyouruseofGenAIcreatinganychallengesinyourcollaborationwithmanagers?(Foreachchallengetheymention,askdirectlyafterdiscussingit):Whataresomewaystodealwiththesechallenges?Howdoyouthinkthesechallengescouldbemitigated?Ourinterviewsrevealedtwofindingsthatruncountertotheexistingliterature.First,thetacticsthatthejuniorsrecommendedtomitigatetheirseniors’concernsrancountertothoserecommendedbyexpertsinGenAItechnologyatthetime,andsorevealedthatthejuniorprofessionalsmightnotbethebestsourceofexpertiseintheeffectiveuseofthisemergingtechnologyformoreseniormembers.Second,whiletheliteraturesuggeststhatthekeyobstacletojuniorseducatingseniorsintheuseofnewtechnologyisstatusthreattoseniors,thejuniorconsultantsweinterviewedreportedthatthekeyobstacletotheirabilitytoeducatetheirmanagerswouldberisksthatGenAIposedtooutcomesthatmanagersvalued(accuracyofoutputs,explainabilityofoutputs,outputsthattakeintoaccountrelevantcontextualdata,andtechnologyusers’activeengagementwithandinterrogationoftheoutputs),ratherthanstatusthreatassociatedwithjuniorscoachingseniors.Inthesectionsthatfollow,wedescribethetwokeytheoreticalargumentsderivedfromourDetailedempiricaldatasupportingthesetheoreticalargumentsisprovidedinAppendix,TablesA-C.Wearguethatjuniorprofessionalsmayfailtobeasourceofexpertiseintheeffectiveuseofanemergingtechnology,becausetheymayrecommendthreekindsofnoviceAIriskmitigationtacticsthat:1)aregroundedinalackofdeepunderstandingoftheemergingtechnology’scapabilities,2)focusonchangetohumanroutinesratherthansystemdesign,and3)focusoninterventionsattheproject-levelratherthansystemdeployer-orecosystem-level.Wefurtherarguethat,toexplaintheconditionsunderwhichjuniorprofessionalsmayfailtobeasourceofexpertiseintheuseofanemergingtechnologyformoreseniormembers,wemusttakeintoaccounttheperceivedrisksthattheemergingtechnologyposestooutcomesthatseniorsvalue.BringingemergingtechnologyrisksintoourunderstandingofbarrierstojuniorsteachingseniorstoeffectivelyusenewtechnologyAstheexistingliteratureoncommunitiesofpracticewouldleadustoexpect,undertheseconditions,thejuniorprofessionalsweinterviewedexpectedthattheywouldlikelybetheonestoeducatetheseniorprofessionalsabouthowtoeffectivelyusethenewtechnology(AppendixTableA).Yet,contrarytowhatthecurrentliteraturewouldexpect,juniorsdidnotfocusonstatusthreattoseniorsasthekeyobstacletotheirabilitytocoachseniorsineffectiveuseofthenewtechnology.Instead,juniorsnotedthatthekeyobstacletotheirabilitytocoachseniorsineffectiveuseofthenewtechnologywouldbethenovelrisksthatthetechnologyposedtooutcomesthatseniorsvalued(Appendix,TableB).JuniorshighlightedthatseniorswouldbeconcernedthatGenAIposedarisktothe

于2023年7月至8月對(duì)78名最近參加過一項(xiàng)實(shí)驗(yàn)的初級(jí)顧問進(jìn)行了訪談，這項(xiàng)實(shí)驗(yàn)使他們能夠訪問生成式AI（GPT?4）來解決商業(yè)問題。這些初級(jí)專業(yè)人士被要求使用虛構(gòu)的與公司高管和品牌）。在他們完成問題解決任務(wù)后不久，我們通過Zoom進(jìn)行了半結(jié)構(gòu)化、60（Spradley1979）。在實(shí)驗(yàn)中獲準(zhǔn)訪問GPT?4用GenAI，但他們預(yù)計(jì)需要教育他們的經(jīng)理使用GenAI，以減輕經(jīng)理對(duì)GenAI對(duì)經(jīng)理重視的結(jié)20237878詢問了這些初級(jí)顧問：1)你能預(yù)見你使用GenAI在你的與經(jīng)理的合作中會(huì)帶來任何挑戰(zhàn)嗎？策略與當(dāng)時(shí)GenAI技術(shù)專家推薦的策略相反，因此表明初級(jí)專業(yè)人士可能不是為更高級(jí)成員有是GenAI對(duì)經(jīng)理重視的結(jié)果（輸出準(zhǔn)確性、輸出可解釋性、考慮相關(guān)背景數(shù)據(jù)的輸出以及技術(shù)用戶對(duì)輸出的積極參與和質(zhì)疑）A?C的專家來源，因?yàn)樗麄兛赡芡扑]三種類型的AI風(fēng)險(xiǎn)緩解策略，這些策略：1)基于對(duì)新興技術(shù)能力的缺乏深入了解；2)專注于改變?nèi)祟惓Ｒ?guī)而不是系統(tǒng)設(shè)計(jì)；3)專注于項(xiàng)目級(jí)干預(yù)而不是將新興技術(shù)風(fēng)險(xiǎn)引入我們對(duì)初級(jí)專業(yè)人士教學(xué)障礙的理解中很可能會(huì)教育資深專業(yè)人士如何有效使用新技術(shù)（A）。然而，與當(dāng)前文獻(xiàn)的預(yù)期相反，的結(jié)果帶來的新穎風(fēng)險(xiǎn)（附錄表B）。初級(jí)專業(yè)人士強(qiáng)調(diào)，資深人士會(huì)擔(dān)心生成式AIaccuracyofoutputs(thedegreetowhichoutputshaveattributesthatcorrectlyreflectthetruevalueoftheintendedattributesofaconceptoreventinaparticularcontextofuse;Steimers&Schneider,2022);totheexplainabilityofoutputs(thedegreetowhichtheoutputsarepresentedinawaythatisunderstandableforhumans;Steimers&Schneider,2022);andtothecontextualizationofoutputs(thedegreetowhichoutputsarecoherent,relevant,orincompliancewithparticularconstraintsorrulesduetoabsorbingandtakingintoaccountpertinentcontextualdatawhenproducingoutputs;Saietal.,2024).JuniorsalsoexpectedthatseniorswouldbeconcernedthatGenAIposedarisktouserengagementbypromotingautomationcomplacency(inwhichusersmaytrusttheoutputsprovidedbygenerativeAIinsituationswheretheyshouldnot;VanDisetal.,2023).WeturnedtotheliteratureontherisksassociatedwithgenerativeAItechnologytobetterunderstandthekindsofrisksthatGenAIwasexpectedtoposetoseniors’valuedoutcomesofaccuracyofoutputs,explainabilityofoutputs,outputsthattookintoaccountrelevantcontextualdata,andGenAIusers’activeengagementwithandinterrogationoftheoutputs.Theliteraturesuggestedthattherewereseveraluniquecharacteristicsofthisemergingtechnologythatposedriskstothesevaluedoutcomes(PaperTable1)thatseemedtobedifferentfromthetechnologiesthathavebeenexaminedbyscholarsintheexistingcommunityofpracticeliteratureonjuniorscoachingseniors.First,GenAIcanbeaccessedandcustomizedbynoviceuserswithoutcodingandwithoutowninginfrastructure.Priorstudiesofjuniorsteachingseniorshaveanalyzedtechnologiesinwhichinteractingwiththetechnologyrequiredinfrastructureandraisedbarriersforordinaryusers.Forexample,theCTscanningtechnologythatBarley(1986)studiedrequiredalargeinvestmentinhardwarebyhospitaladministrators,andintegrationwithexistinghospitalmedicalequipmentandITinfrastructurebyITprofessionals.Incontrast,GenAIallowsnoviceusersaccessthetechnologydirectlyfromtheircomputersandcollaboratewithitinanearlyinstantaneousfashion.Second,GenAIhasuncertainandwide-rangingcapabilitiesandischangingatanexponentialrate(OpenAI,2023a;Webbetal.,2023;Weietal.,2022).Priorstudiesofjuniorsteachingseniorshaveexaminedtechnologiespurpose-builtforaspecificapplication,andchangingataslowerrate.Forexample,theDavincirobotthatBeane(2019)studiedwaspurpose-builttoallowasinglesurgeontoperformbothsupport(retraction)anddirectsurgicalaction(dissection)toanunprecedenteddegree.Incontrast,withGenAI,abroadrangeofapplicationscanbeperformedbyasinglesystem,andGenAI’scapabilitiesareexpandingexponentially.Third,GenAIcarriesthepossibilityofoutperforminghumansinawidevarietyofskilledandcognitivetasks(Bubecketal.,2023).Priorstudiesofjuniorsteachingseniorshaveinvestigatedtechnologiesdesignedtoassistprofessionals.Forexample,theEMRtechnologythatKelloggandcolleagues(2021)studiedwasabletoassistmedicalassistantstodotheirworkofpatientcheck-inandcheck-outprocessesandreferralsandinsuranceauthorizationsmoreaccuratelyandquickly.Incontrast,GenAIholdsthepotentialtosurpasshumanperformanceacrossadiverserangeofskilledandcognitiveFourth,GenAIcombinesdataandotherinputsatanunprecedentedscaleanddetailfromabroadecosystemofactors.Priorstudiesofjuniorsteachingseniorshaveanalyzedtechnologiesthatinvolvedmodeldevelopers,systemdeployers,andsystemusers,butweredependentonalimitedsetofdatafromeach.Forexample,theFactsetandCapIQtechnologythatAnthony(2021)studiedgathereddatadirectlyfrompublicSECfilings,automaticallycalculatedparticularkeymetrics,andthenfedthesecalculationsdirectlyintospreadsheetsintheuser’sorganization.Incontrast,withGenAI,continuouschangeindatasourcesatanunprecedentedscaleanddetailaffectwhatthetechnologyisabletodo.Thus,GenAIemergesthroughasetofexpandedrelationsandcontinuestoemergeinnewwaysasthoserelationsInsum,oneofourfindingsthatrunscountertotheexistingliteratureisthatjuniorsexpectedthatthekeyobstacletotheirabilitytocoachseniorsineffectiveuseofGenAIwouldbethenovelrisksthatthisemergingtechnologyposedtooutcomesthatseniorsvalued.Thus,toexplainhowandwhenjuniorprofessionalsmayfailtobeasourceofexpertiseintheuseofanemergingtechnologyformoreseniormembers,wemusttakeintoaccountnotonlystatusthreat,butalsoriskstovaluedoutcomesthatare

不要期望初級(jí)人員教高級(jí)專業(yè)人員使用生成式AI：新興技術(shù)風(fēng)險(xiǎn)和初級(jí)AI風(fēng)險(xiǎn)緩解策略（Steimers&Schneider，2022年）；到輸出的可解釋性（輸出的呈現(xiàn)方式在多大程度上對(duì)人類來說是可理解的；Steimers&Schneider，2022年）；以及到輸出的情境化（在生成輸出時(shí)Sai等人，2024年）。初級(jí)人員還預(yù)計(jì)，高級(jí)人員會(huì)擔(dān)心生成AI通過促進(jìn)自動(dòng)化懈?。ㄓ脩粼趹?yīng)該不信任生成AI輸出的情況下信任其輸出；VanDis等人，2023年）對(duì)用戶參與構(gòu)成風(fēng)險(xiǎn)。我們轉(zhuǎn)向關(guān)于生成式AI技術(shù)風(fēng)險(xiǎn)的文獻(xiàn)，以更好地了解生成AI預(yù)期會(huì)對(duì)高級(jí)人員重視的輸出準(zhǔn)確性、輸出可解釋性、考慮相關(guān)上下文數(shù)據(jù)的輸出以及生成AI質(zhì)疑構(gòu)成哪些風(fēng)險(xiǎn)。文獻(xiàn)表明，這種新興技術(shù)具有一些獨(dú)特的特征，這些特征對(duì)上述價(jià)值結(jié)果構(gòu)成了風(fēng)險(xiǎn)（論文表1），這些特征似乎與現(xiàn)有實(shí)踐社區(qū)文獻(xiàn)中學(xué)者們所考察的技術(shù)不同。首先，AI可以被新手用戶訪問和定制，無需編碼且無需擁有基礎(chǔ)設(shè)施。先前關(guān)于初級(jí)（1986）CTIT專業(yè)人員將其與現(xiàn)有的醫(yī)院醫(yī)療設(shè)備和IT基礎(chǔ)設(shè)施集成。相比之下，生成AI允許新手用戶直接從他們的電其次，生成式AI具有不確定性和廣泛的能力，并以指數(shù)級(jí)速度變化（OpenAI，2023aWebb2023Wei2022）。先前關(guān)于讓初級(jí)人員教授高級(jí)人員使用技術(shù)的相關(guān)研究，考察的是為特定應(yīng)用而專門構(gòu)建且變化較慢的技術(shù)。例如，Beane（2019）研究的大衛(wèi)·艾文機(jī)比之下，生成式AI可以由單個(gè)系統(tǒng)執(zhí)行廣泛的操作，并且生成式AI的能力正在呈指數(shù)級(jí)增長。第三，生成式AI有可能在各種技能和認(rèn)知任務(wù)中超越人類（Bubeck等人，2023）。先Kellogg及其同事（2021）研究的是電子病歷技術(shù)，該技術(shù)能夠幫助醫(yī)療助理更準(zhǔn)確、更快AI有潛力在廣泛第四，生成式AI入（Anthony，2021）。先前關(guān)于讓初級(jí)人員教授高級(jí)人員使用技術(shù)的相關(guān)研究，分析了涉及模型開發(fā)者、系統(tǒng)部署者和系統(tǒng)用戶的技術(shù)，但這些技術(shù)依賴于每個(gè)參與者有限的數(shù)據(jù)集。例如，Anthony（2021）研究的事實(shí)集和CapIQ技術(shù)直接從公開的SEC文件中收集數(shù)據(jù)，自動(dòng)計(jì)算特定的關(guān)鍵指標(biāo)，然后將這些計(jì)算直接輸入到用戶組織的電子表格中。相比之下，生成式AI在數(shù)據(jù)源方面的連續(xù)變化以前所未有的規(guī)模和細(xì)節(jié)影響著該技術(shù)所能做到的事情。因此，生成式AI通過一系列擴(kuò)展的關(guān)系出現(xiàn)，并且隨著這些關(guān)系的演變而以新的方式不斷出現(xiàn)。使用生成式AI了解釋初級(jí)專業(yè)人員可能無法成為高級(jí)成員使用新興技術(shù)的專家來源的原因，以及何時(shí)可能失敗，我們必須考慮的不僅僅是地位威脅，還有對(duì)重視的結(jié)果帶來的風(fēng)險(xiǎn)。posedbyemergingtechnologiesthat1)haveuncertainandwide-rangingcapabilitiesandarechangingatanexponentialrate,2)havethepossibilityofoutperforminghumansinawidevarietyofskilledandcognitivetasks,and3)combinedataandotherinputsatanunprecedentedscaleanddetailfromabroadecosystemofactors.Suchrisksassociatedwithanemergingtechnologymayposethreatstoseniors’valuedoutcomesofaccuracyofoutputs,explainabilityofoutputs,outputsthattakeintoaccountrelevantcontextualdata,andtechnologyusers’activeengagementwithandinterrogationoftheoutputs.BringingnoviceAIriskmitigationtacticsintoourunderstandingofjuniorsteachingseniorstoeffectivelyusenewtechnologyAsecondfindingthatrunscountertotheexistingliteratureisthatthetacticsthatthejuniorsrecommendedtomitigatetherisksthatGenAIposedtoseniors’valuedoutcomesrancountertotheriskmitigationtacticsrecommendedbyexpertsinGenAItechnologyatthetime,andsorevealedthatthejuniorprofessionalsmightnotbethebestsourceofexpertiseintheeffectiveuseoftheemergingtechnologyformoreseniormembers.Wefoundboththehuman-computerinteraction(HCI)literature’sconceptofnovicetacticsfortechnologyuseandrecentpublicationsbyexpertsinGenAItobehelpfultounderstandingthedynamicsweobserved.TheHCIliteratureonnovicetacticsfortechnologyusehelpstoexplainwhyjuniorprofessionals’time-testedwaysofself-upskillingbyengaginginreal-timeexperimentationmaynotallowthemtoidentifythemostproductivewaystomitigatetheriskthre