2025哈佛大學(xué)人工智能如何限制人類體驗

HowArtificialIntelligenceConstrainstheHumanExperienceANAVALENZUELA,STEFANOPUNTONI,DONNAHOFFMAN,NOAHCASTELO,JULIANDEFREITAS,BERKELEYDIETVORST,CHRISTIANHILDEBRAND,YOUNGEUNHUH,ROBERTMEYER,MIRIAME.SWEENEY,SANAZTALAIFAR,GEOFFTOMAINO,ANDKLAUSWERTENBROCHABSTRACTArtificialintelligence(AI)andrelatedtechnologiesaretransformingmanyconsumptionactivities,pow-eringbreakthroughsthatexpandthehumanexperiencebyenhancinghumancapabilities,performance,andcreativity.Whilethisexplainstheconsumerenthusiasmandrapidadoptionofthesetechnologies,AIsystemscanalsohavetheoppositeeffect:reducingandconstrainingtherangeofexperiencesthatareavailabletoconsumers.Thisarticleexam-inesthemechanismsthroughwhichAIcanconstrainthehumanexperience,consideringindividual,interpersonal,andsocietalprocesses.OuranalysisuncoversacomplexinterplaybetweentheadvantagesofAIanditsinadvertentnega-tiverepercussions,whichpotentiallyrestricthumanautonomy,self-identity,relationaldynamics,andsocialbehavior.Inthisarticle,weproposethreedifferentmechanismsatthecoreoftheseconstrainingforces:parametricreduction-ism,agencytransference,andregulatedexpression.Ourexplorationofthesemechanismshighlightstherisksconnectedtosystemdesignandpointstoquestionsandimplicationsforfutureresearchersandpolicymakers.

ANAVALENZUELA，STEFANOPUNTONI，DONNAHOFFMAN，NOAHCAJULIANDEFREITAS,BERKELEYDIETVORST,CHRISTIANHILDEBRAND,YOUNGEUNHUH,ROBERTMEYER,MIRIAME.SWEENEY,SANAZTALAIFAR,GEOFFTOMAINO,ANDKLAUSWERTENBROCH摘要人工智能（AI）AIAIAI人類自主性、自我認(rèn)同、關(guān)系動態(tài)和社會行為。在本文中，我們提出了三種構(gòu)成這些限制力量的核心機制：參數(shù)化簡約主義、代理轉(zhuǎn)移和規(guī)范表達(dá)。我們對這些機制的研究突出了系統(tǒng)設(shè)計相關(guān)的風(fēng)險，并為未來的研究人員和政策制定者指出了問題和影響。anyconsumptiondecisionsandexperiencesaredigitallymediated.Asaconsequence,consumerbehaviorisincreasinglythejointproductofmanpsychologyandubiquitousalgorithms(e.g.,Melumadetal.2020;Sangersetal.2024).Thecomingofageoflargelanguagemodels(LLMs)isfurtheracceleratingthedissem-inationandimpactofartificialintelligence(AI).AIholdsthepromiseofimprovingthelifeofconsumerseverywhere,inwayssmallandlarge.Atthesametime,thedeploymentofthistechnologyisnotwithoutrisks.Thesocietal

courseonthepotentialrisksofAItendstofocusonissuesofdiscriminationandprivacy,orondistant“existential”risks(e.g.,thepossibilityofhumanextinctionoranirre-versibleglobalcatastrophe).Ourfocusisdifferent.Follow-ingtheworkofconsumerresearcherswhohavestartedtoidentifypsychologicaltensionsintheconsumerexperienceofAI(Puntonietal.2021),wecontributetothisnascentliteraturebyexploringhowAIcanreducetherangeofpeo-ple’sexpressionandchoices,aswellastheopportunities

在的算法（例如，Melumad等人，2020年；Sangers等人，2024年）的聯(lián)合產(chǎn)物。大型語言模型（LLMs）的成熟進一步加速了人工智能（AI）的傳播和影響。AI承諾以大小不同的方于AI潛在風(fēng)險的課程往往側(cè)重于歧視和隱私問題，或者關(guān)注遙遠(yuǎn)的“存在性”風(fēng)險（例如，人類滅絕或不可逆轉(zhuǎn)的全球災(zāi)難AI消費者體驗中的心理緊張之后（Puntoni等人，2021年），我們通過探討AI如何減少人們的表達(dá)和選擇范圍，以及他們個

對AI潛在風(fēng)險的課程往往側(cè)重于歧視和隱私問題，或者關(guān)注遙遠(yuǎn)的存在性風(fēng)險（例如，人類滅絕或在消費者研究人員開始識別AI消費者體驗中的心理緊張之后（Puntoni等人，2021年），我們通過探討AI如何減少人們的表達(dá)和選擇范圍，以及他們個人發(fā)AnaValenzuela(correspondingauthor:ana.valenzuela@)isaprofessoratESADE-RamonLlul,Barcelona,Spain,andBaruchCollege,CityUniversityofNewYork(CUNY),NewYork,NY10010,USA.StefanoPuntoni(puntoni@)isaprofessoratWhartonSchool,UniversityofPennsylvania,Philadelphia,PA19104,USA.DonnaHoffman(dlhoffman@)isaprofessoratGeorgeWashingtonUniversity,Washington,DC20052,USA.NoahCastelo(ncastelo@ualberta.ca)isaprofessoratUniversityofAlberta,EdmontonABT6G2R3,Canada.JulianDeFreitas(jdefreias@)isaprofessoratHarvardBusinessSchool,Boston,MA02163,USA.BerkeleyDietvorst(berkeley.dietvorst@)isaprofessoratBoothSchoolofBusiness,UniversityofChicago,Chicago,IL60637,USA.ChristianHildebrand(christian.hildebrand@unisg.ch)isaprofessoratUniversityofSt.Gallen,St.Gallen,Switzerland.YoungEunHuh(younghuh@kaist.ac.kr)isaprofessoratKoreaAdvancedInstituteofScienceandTechnology,Daejeon,SouthKorea.RobertMeyer(meyerr@)isaprofessoratWhartonSchool,UniversityofPennsylvania,Philadelphia,PA19104,USA.MiriamE.Sweeney(mesweeney1@)isaprofessoratUniversityofAlabama,Tuscaloosa,AL35487,USA.SanazTalaifar(s.talaifar@imperial.ac.uk)isaprofessoratImperialCollegeLondon,SouthKensington,LondonSW72AZ,UK.GeoffTomaino(geoffrey.tomaino@)isaprofessoratUniversityofFlorida,Gainesville,FL32611,USA.KlausWertenbroch(klaus.wertenbroch@)isaprofessoratINSEAD,Fontainebleau,France.TheauthorsthankJohnDeightonforhisveryclearguidanceasGuestEditorandthefullreviewteamfortheirhelpfulcomments.TheywouldalsoliketothankZivCarmonandPauloAlbuquerque,andthewholeteamatINSEAD,fororganizingthe2023ChoiceSymposium,whichrepresentedthegenesisofthisarticle.JACRisgratefultoJohnDeighton,whograciouslyagreedtoserveasGuestEditorforthisarticle.IssueEditors:StefanoPuntoniandKlausPublishedonlineMay15,TheUniversityofChicagoPressfortheAssociationforConsumerResearch./10.1086/730709

·（通訊作者：ana.valenzuela@）ESADE?RamonLlulCUNY）·（puntoni@）是美國賓夕法尼亞大學(xué)沃頓商學(xué)院的教授。唐·（dlhoffman@）·（ncastelo@ualberta.ca）是加拿大阿爾伯塔大·（jdefreias@）·（berkeley.dietvorst@）·（christian.hildebrand@unisg.ch）是瑞士圣加·（younghuh@kaist.ac.kr）·（meyerr@）·E·（mesweeney1@）·（s.talaifar@imperial.ac.uk）·（geoffrey.tomaino@）是美國佛羅里達(dá)大學(xué)的教·（klaus.wertenbroch@）INSEAD···INSEAD2023·本期編輯：StefanoPuntoni和KlausWertenbroch在線發(fā)表于2024年5月15日。TheUniversityofChicagoPressfortheAssociationforConsumerResearch./10.1086/730709 HowArtificialIntelligenceConstrainstheHuman Valenzuelaet

H人工智能如何限制人類經(jīng) 瓦倫齊亞等ourintentistodelineatehowconsumers’possibilityspacesareincreasinglyshapedbyalgorithms.Thisapproachisalignedwithindustrycallstounderstandhow“customersfaceanarrayofnewdeviceswithwhichtointeractwithfirms,fundamentallyalteringthepurchaseexperience”(MarketingScienceInstitute2018,4).Theterm“artificialintelligence”wascoinedbyJohnMc-Carthyandhiscolleaguesin1955as“thescienceandengi-neeringofmakingintelligentmachines,”withtheemphasisonmachines’capabilitiestolearn,atleast,inpart,ashu-mansdo(McCarthyetal.2006,1).Today,AIrepresentsawiththepotentialtoaffecteveryindustry(Agrawal,Gans,andGoldfarb2018)andtransformtheeconomy(FurmaningscholarlyliteraturedocumentsthebusinesspotentialofAIandBigData(e.g.,Agrawaletal.2018).Forconsumers,AIholdsthepromiseofhelpingthemmakemoreefficientandeffectivedecisions,savetime,andenjoybetterproductsandexperiencesthatmoreaccuratelymatchtheirpreferences.Toillustrate,recommendersystemslikeNetflix’sandper-sonalizedproductslikeSpotify’s“madeforyou”playlistsimprovethequalityofourconsumption.GenerativeAIap-plicationslikeChatGPT,whichcanhelppeoplewritete-diousreportsinminutes,helpusbecomemoreproductiveandsavetime(NoyandZhang2023).RoboadvisersandotherdecisionsupportsystemscanhelpincreasethequalityequallyconsiderthepotentiallyperniciousoutcomesthatcouldemergefromourincreasingrelianceonAIformakingtheirpotentialrisks,whichseemintrinsicinapowerfultech-nologythatremains,inmostcases,a“blackbox”toconsum-ersandresearchersalike(DeFreitasetal.2023).Inothersystemsraisesseriousconcernsaboutitspotentialforalter-ingcoreelementsofhumanagencyandmotivationinunin-tendedways.Todate,fewinvestigationshavefocusedonthereducehumanexperience(HoffmanandNovak2018).OurfocusinthisarticleistoexaminethepotentialofAItoconstrainhumanexperience.By“humanexperience,”werefertopeople’soverallperceptions,feelings,andbehaviorsastheyengageininteractionswithtechnology(AI),particu-journey(followingBrakus,Schmitt,andZarantonello

andLemonandVerhoef2016).Weintendtounderstandunderstoodasreductionsorlimitationsofpeople’sexperiencedirectly,asbyfeedingusersalimitedarrayofoptionsthatpreventsthemfromexploringnewoptionsandperspectives.Theselimitationsreflecttheactive,agenticroleofAIsys-temsinshapingthenatureandextentofhuman-technologyinteractions,whichmayinvolveremovingcomponents,lim-interactionsamongthem.Intandemwithdirectlyconstraininghumanexperience,AIcanconstrainhumanexperienceindirectly,aswhenpeo-pleuseAIinwaysthatlimittheirownengagementwiththeexperience(HoffmanandNovak2018).Themotivationforsuchself-restrictioncanvary,butcanincludeconcernsoverprivacy,autonomy,ethicalconsiderations,oraccommodat-inganAI’slimitations(e.g.,theinabilityofavoiceassistanttounderstandlong,complexexpressions).Throughself-restriction,peopledirectlylimitthescopeandfunctionalityoftheirAIinteractions,effectivelyreducingitspotentialim-pactandcapabilities.Forexample,aconsumermightdecideprefertomanuallyoverridesettingsondeviceslikethermo-statsratherthanrelyonAIoptimizationorpersonalizationbutself-restrictionsofthehumanexperiencearecharacter-izedbyconsumersmodifyingtheirbehavior,language,orinteractionpatternstodefendagainstanAIsystemorcon-formtoitsoperationalparameters.Thisadaptationmayresultintheconsumerusingareducedsetoftheirownca-pacitiesoralteringtheirinteractionstylestoaccommodatebeingdiminishedor“l(fā)essthan”becausetheconsumermustinteractinwaysthatdonotfullyrepresenttheirpotentialorpreferredmodesofinteraction.Forexample,consumersmightexposuretodifferentoptions,reducingtheiragencyindecisioncoreoftheseconstrainingforces:agencytransference,para-withtheintentofmappingconsumers’interactionswith

地受到算法的影響。這種方法與行業(yè)呼吁理解客戶驗（營銷科學(xué)研究所2018年，第4頁）相一致。在1955年提出，定義為“制造智能機?的科學(xué)和工程”，強調(diào)機?至少部分地具有人類的學(xué)習(xí)能力（麥一種通用技術(shù)（Brynjolfsson和McAfee，2014年），具有影響每個行業(yè)（Agrawal、Gans和Goldfarb，2018年）和改變經(jīng)濟（Furman和Seamans，2019了人工智能和大數(shù)據(jù)的商業(yè)潛力（例如，Agrawal等，2018年）。對于消費者來說，人工智能有望幫助他們合他們偏好的更好產(chǎn)品和服務(wù)。例如，Netflix的推薦了我們消費的質(zhì)量。像ChatGPT這樣的生成式人工智我們更有效率，節(jié)省時間（Noy和Zhang，2023年）并改善結(jié)果（Hildebrand和Bergner，2021年）。雖然這些收益是真實且顯著的，但同樣重要的是要考慮我們?nèi)找嬉蕾嚾斯ぶ悄軄碜龀鲇绊懴M者和公民的決策可能產(chǎn)生的潛在有害后果，并評估其潛在風(fēng)險，這些風(fēng)險似乎與這種強大的技術(shù)固有相關(guān)，而在大多數(shù)情況下，這種技術(shù)對消費者和研究人員來說仍“”（DeFreitas2023）句話說，將越來越多的任務(wù)委托給人工智能系統(tǒng)，可能會以意想不到的方式嚴(yán)重改變?nèi)祟惸軇有院蛣訖C的核心要素。迄今為止，很少有研究關(guān)注智能技術(shù)如人工智能限制、限制或減少人類經(jīng)驗的潛力（HoffmanNovak，2018）。所謂“人類經(jīng)驗”，是指人們在與技術(shù)（人工智能）費者和在整個客戶旅程（參見Brakus、Schmitt和Zarantonello2009年）

驗（HoffmanNovak2018）”，因為他們必須以不完全代表他們潛在或首選Volume9Number3

第9卷第3期2024 throughouttheirdecisionjourney:handingoveragencytoalgorithms,whichreduceconsumerstoalimitedsetofpa-rameters,andpotentiallyconstrainhowconsumersexpressthemselvesandcommunicate.Below,wediscussthekeymechanismsthathelpexplainandpredictAI’spotentialtoconstraintheconsumerexperience.Thearticleconcludeswithasetoffutureresearchquestionsandimplicationsforpractitioners,includingAIdevelopersandpolicymakers.Asidefromconsumerexperiences,weshowhowAImayim-poseconstraintsonagency,skills,equality,dignity,anddi-versity.Itthereforedemandsurgentandmulti-disciplinarymanexperience.AGENCYTRANSFERENCEAgencytransferencerelatestoAI’sabilitytolimitone’sper-sonalagency,asagencyistransferredfromhumanstoalgo-rithms(alsoseeDeFreitasetal.2023).Personalagencyhasbeendefinedas“thesensethatIamtheonewhoiscausingorgeneratinganaction”(Gallagher2000,15)orashavingthepowertoinfluenceone’sownactionsandcircumstancesratecontent(Andréetal.2018;Wertenbrochetal.2020).Asaresult,consumersbecomelesslikelytobeexposedtoop-tionsandcontentthatdoesnotcorrespondtotheirrevealedchangethesepreferencesandchoosesomethingelse(enact-cansubtlymanipulatedecision-makingtrajectoriesinwaysthatultimatelyconstrainself-determination(Bhattacharjeeetal.2014).Withinthislargerthemeofagencybeing“trans-cusshowAIfavorsalossofserendipityinaccessingoptionsandenablesbothcognitiveandemotionalde-skilling.Oneofthejoysoflifeisserendipity,makingfortunatediscov-eriesbyaccident.Serendipitousdiscoveriesareonesthatarerelevanttous,despitebeingunexpected(Kotkov,Wang,andVeijalainen2016).Forinstance,youmightpopintoabook-storeanddiscoveranobscurebookyetfindithighlyrelevanttoapaperyouhavebeenwriting.Oryoumighttalktoacol-leagueandfindthat,surprisingly,theyareafountainofknowledgeonthelocalartscenejustwhenyouwereinsearchofagreatshow.

However,recommendationalgorithmsoftenlimitourexperienceofserendipity.Theydosobecausemostofthesesystemsaresetuptofeedcontentbasedonourpastbehav-iors(Andréetal.2018;Wertenbrochetal.2020).Conse-quently,theyreinforcethosebehaviorsandcreateinertiathatlimitsexplorationandchange(TalaifarandLowery2023),constraininghumanagency.Forinstance,sequen-tiallyviewingcontentcausesonetoconsumemorecontentfromthesamecategory,explainingthecommonexperienceof“goingdownarabbithole”(WoolleyandSharif2022).Moreover,pastworkfindsthatrecommendationalgo-rithmsmayalsolimitaggregateserendipity.Becausesimilargroupsofconsumerssimultaneouslyreceivesimilarcontentrecommendations,thismayencouragesimilarconsumerstohomogenizeevenmore(FlederandHosanagar2009;Leeinwhichthemarketshareforproductsthatarealreadypop-ularincreasesattheexpenseofothers;whilerecommenda-tionalgorithmsmayincreaseabsolutesalesfornicheitems,morepopularitems(LeeandHosanagar2019).Forthesereasons,somehaveadvocatedcomplementingthecurrent“Skinnerian”approachtoserendipityinrecom-ingofconsumers(Morewedgeetal.2023).Forinstance,justman,Rogers,andBazerman2009),platformscouldemploysusserendipitouscontent(Morewedgeetal.2023).Infact,pastresearchhasshownthatwhenrecommendersystemsommendations(Smetsetal.2022).certaintasksorskillsthatwerepreviouslyperformedbyhu-thoseskillsatrophyovertime,aphenomenonknownasde-broughtabouttheinventionoftextilemachinerylikethespinningjenny,powerloom,andcottongin.These

為代理權(quán)從人類轉(zhuǎn)移到算法（參見DeFreitas等人，2023）我是導(dǎo)致或產(chǎn)生行動的人的感覺（Gallagher2000，15）或擁有影響自己行動和情況的能力（Bandura2006）。通過內(nèi)容（André等人，2018年；Wertenbroch等人，2020年）。因此，消費者不太可能接觸到不符合他們Wertenbroch等人，2020年）。由于人工智能系統(tǒng)可以塑造我們的環(huán)境——我們看到的東西以及我們可用的機會（Grafanaki2017年）——它們可以微妙Bhattacharjee等人，2014年）。在這個更大的主題到人工智能并因此限制人類代理出乎意料的發(fā)現(xiàn)（Kotkov,Wang,和Veijalainen2016）。例如，你可能會走進一家書店，發(fā)現(xiàn)一本冷

提供內(nèi)容的（André等人2018；Wertenbroch等人限制了探索和變化（TalaifarLowery2023），限消費更多同一類別的內(nèi)容，解釋了常見的“陷入兔子洞的經(jīng)歷（Woolley和Sharif2022）。此外，以往的研究發(fā)現(xiàn)，推薦算法還可能限制總體上的偶然發(fā)現(xiàn)。因為相似的消費者群體同時接收到相似的內(nèi)容推薦，這可能會鼓勵相似的消費者進一步同質(zhì)化（FlederHosanagar2009Lee2019）相關(guān)，這導(dǎo)致了流行度偏差，即已經(jīng)流行的產(chǎn)品的市場份額增加，而其他產(chǎn)品的份額則減少；雖然推薦算法可能會增加利基產(chǎn)品的絕對銷量，但這些收益與更受歡迎的產(chǎn)品相比可能微不足道（LeeHosanagar2019）。因此，有些人主張在推薦算法中補充當(dāng)前的“斯偶然發(fā)現(xiàn)方法，以更好地理解消費者心理（Morewedge等人2023）。例如，僅僅因為消費者的愿望和應(yīng)該做的事情（Milkman,Rogers和Bazerman2009），平臺可以采用分析更長的消費時偶然內(nèi)容程度的技術(shù)（Morewedge等人2023）。事實上，以往的研究表明，當(dāng)推薦系統(tǒng)（例如Yelp）建些推薦（Smets等人2022）。技能化（Wood1987）。例如，工業(yè)革命帶來了紡織 HowArtificialIntelligenceConstrainstheHuman Valenzuelaet H人工智能如何限制人類經(jīng) Valenzuela等replacedartisanweaversandspinners,transformingthetex-tileindustryfromaskilledcrafttoafactory-basedsystemwhereworkersoperatedmachinerywithoutneedingtheskillsofthetraditionaltextileartisan(BergandHudsonTheemergenceofChatGPTandotherformsofgenera-tiveAIhaverevivedtheconcernthatAIcouldcontributetoanewwaveofde-skillingamongknowledgeworkers.Forexample,generativeAImodelscannowoutperformmosthumansatcreativeideageneration(Guzik,Byrge,andGilde2023;KoivistoandGrassini2023).Relyingonsuchmodelscanincreasethequantityandqualityofknowl-edgeworkers’output(NoyandZhang2023;Dell’Acquaetal.2024).IfworkersbegintorelyonAItoperformpartsoftheirjobs,suchaswritingorbrainstorming,mighttheybeforegoingopportunitiestopracticeandstrengthenthoseskills?Couldthisinturncausethoseskillstoweaken,rela-tivetoworkerswhodonotrelyonAIassistance?Similaref-pleseemtooffloadtheneedtorememberinformationtotheInternet,suchthatmemorybecomesworsewhenrely-ingontheInternettofindinformation(Ward2021;Fisher,Smiley,andGrillo2022).De-skillingmightalsobeaconcernoutsideofworkortask-orientedcontexts;byimpactingemotionalandsocialskills.Forinstance,recentresearchsuggestedthat,ascon-sumerstradeoffhumancontactwithmorerelianceonAI-poweredtechnology,theremaybeanincreasedvarianceinsocialadjustment,impactingemotionalintelligence,par-ticularlyamongouryouth(e.g.,Beranuyetal.2009;RalphandNunez2021).Thereseemstobeearlysupportforthepossibilitythatrelyingonalgorithmstomanageinterac-tionsonsocialmediaplatformsisconnectedtoadeclineinourabilitytounderstandandnavigatecomplexsocialcues.Furthermore,manypeoplearealreadyinteractingex-tensivelywithAI-basedsyntheticcompanionssuchasReplika(DeFreitasetal.2024).Duetothe“blackbox”na-tureofthealgorithms,itisimpossibletopredictinadvancehowtheseconversationswillunfold.DeFreitasetal.(2024)andtorespondappropriatelytosignsofdistress,whichcallsintoquestionthesafetyofchatbotsforindividualswithmentalhealthissues.Furthermore,thecurrentregulatorystructureisnotsetuptoaddresstheserisks(DeFreitasandCohen2024).Finally,wepositthattheremaybeadditional,societalimplicationsofde-skillingduetoalgorithmicdeci-sionmaking.Whenaskillisentirelylostina

certainexperiencesmaynolongerbeaccessible.Forin-stance,thelossofskilltooperateandrepairolderdevices,suchasmediaplayersforobsoletestandards,canmakecer-taininformationinaccessible,ashappenedafewyearsagoinBritain.Amassiveprojectofhistoricaldocumentationcarriedoutinthe80sbytheBBCrequiredyearsofwork2023).Thedangerofautomation-causedde-skillinghaslongbeenasalientconcerninthecontextsofcomplexsys-temsoroccupations,suchasairlinepilots(e.g.,Carr2014).Inthesamevein,itispossiblethatifwegrowentirelyreli-antonAIforcertaintasks,theresilienceofcommunitiescouldbeunderminedifthoseskillssuddenlybecomevalu-ableagain(e.g.,becauseofsystemicfailureofourdigitalin-frastructure).Forthesereasons,itisessentialtomaintainabalancebetweenusingalgorithmstoaiddecisionmakingandpreservingthehumancapacitytomakedecisionsandcarryouttasksindependently.Insum,AIhasthepotentialtoconstrainhumanagencybyreducingtherangeofpossiblechoicesandactionsthatconsumersmightconsider,thuslimitingopportunitiestoexplore,learn,andchangeestablishedbehavioralpatterns.Moreover,relianceonAImayalsoleadusers—consumersandworkersalike—tounlearnvaluablecognitiveandprac-ticalskills,notonlyasindividualsbutalsoassocieties.PARAMETRICREDUCTIONISMBytheirverynature,algorithmsarereductionists.Theyneedtotranslatehumanbehavior,identity,preferences,andattributesintoasmallersetofindependent,computa-tionallyreadablevariables,parameters,andformulae(Hil-debrand2019).Inthisway,AIsystemstendtoobjectifyin-dividualsandcommunities,reducingorcompressingtheiruniquecharacteristicsandculturalcontexts.Thisprocessmayleadtomisalignment,thatis,themisrepresentationorunder-representationofpeople’sactualpreferencesandinterestswhentheyaretranslatedintoanalgorithmicfor-mula.Objectificationandmisalignmentarethetwomecha-nismswediscussnext.AIfunctionsthroughparameterizationandcategorization,reducingthecomplexitiesofhumanbeingsintoasetofquan-tifiablemetrics,classifications,andriskscorestosort,assess,andpredictbehavior..Thus,thisprocessislimitedinitsabil-itytofullyaccountfortheuniquecharacteristicsandcircum-

（Berg和Hudson1992年）ChatGPT和其他形式生成AI的出現(xiàn)重新引發(fā)了人們對于AI可能引發(fā)知識工作者新一輪技能退化的擔(dān)憂。例如，生成AI模型現(xiàn)在在創(chuàng)意想法生成方面可以超越大多數(shù)人類（Guzik，Byrge和Gilde2023年；Koivisto和Grassini2023年）。依賴這些模型可以提高知識工作者的產(chǎn)出數(shù)量和質(zhì)量（Noy和Zhang2023年；Dell’Acqua等人2024年）。如果工人開始依賴AI來完成他們工作中的一部分，比如寫作或頭這反過來會不會導(dǎo)致這些技能相對于不依賴AI輔助的賴互聯(lián)網(wǎng)查找信息時，記憶力會變差（Ward2021Fisher，Smiley和Grillo2022年）AI年；Ralph和Nunez2021年）。似乎有早期證據(jù)支AI的合成伴侶，Replika（DeFreitas2024）進行互動。將如何展開。DeFreitas等人（2024年）AI來以應(yīng)對這些風(fēng)險（DeFreitas和Cohen2024年）

BBC80獲?。℉arford2023）。自動化導(dǎo)致的去技能化在復(fù)雜系統(tǒng)或職業(yè)（如飛行員，例如，Carr2014）的背景賴AI來完成某些任務(wù)，那么如果這些技能突然變得有總之，AI有可能通過減少消費者可能考慮的選擇和改變既定行為模式的機會。此外，對AI的依賴還可的變量、參數(shù)和公式（Hildebrand2019）。因此，AI系統(tǒng)往往將個人和社區(qū)對象化，減少或壓縮它們的Volume9Number3

卷9第3期2024 dehumanizationas“theactofperceivingortreatingpeopleasiftheyarelessthanfullyhuman,”Fiske(2009)specifiesthatthereisonespecificformofdehumanization,whichmightbetermedobjectification,whichviewspeopleasautomatons(tools,robots,machines).Webelievethisdefinitionbestanismdefinedas“algorithmsdistillingindividualsintodataInthismanner,evenAIdesignedforbeneficialpurposescanpropagateasubtleyetperniciousformofhumanobjecti-ficationthatpersistswithintechnicalsystems.Thesesystemsoftenoverlookhowcharacteristicsusedtojudgeindividualswillsystematicallycorrelatewithotheraspectsofanindivid-ual’sbackground.AnotableexamplewasAmazon’shiringal-gorithm,whichexhibitedabiasagainstwomenduetoitshighpositiveweightingofcharacteristicstraditionallyassoci-atedwithmen(Dastin2018).Similarpatternsofdiscrimina-tionhavebeenuncoveredinAIapplicationswithinjudicial(Larsson2019)andeducationaldomains(Engler2021).Fur-rithmitselfasitaimstorepresentpeople’spreferencestruc-ture(Morewedgeetal.2023).ThisgapbetweentherealhumanindividualandtheAI-codedrepresentationofthemcanresultinunintendedharmtiesandcausingindirectsocialspillovereffects.Objectifica-tioncanamplify(andobscure)systemicinequalitiesbyreduc-ingindividualstogroupcharacteristics.Thatis,individualsofcertainbackgroundsareoftensystematicallygivendifferentopportunitiesasaresultofalgorithmicobjectification,affect-Weill2021)andpricing(Chapdelaine2020).Consumersthus(rightly)questionAI’scapacitytoappreciatetheiruniquetraitsandcircumstances,andshowreluctancetoutilizeitinimportantcontexts,suchashealthcare(e.g.,Longoni,Bonezzi,andMorewedge2019)orfinancialservices(Yalcinetal.2022).Additionally,theobjectificationintroducedbyAIcanhaveinginteractionswithAIscanleadtospillovereffectswherebythemmoreinstrumentally(e.g.,Onuretal.2023).Further-more,consumersviewobjectifyingAIsystemsasbeinglesscapableofassessinginterpersonalskills(Castelo,Bos,andLehmann2019).Asaresult,theobjectifyingnatureofAIscaninfluenceperceptionsofthoseselectedbyAIinhiring

ample,Granuloetal.(2024)findthattheuseofAIinman-agementtaskscanincreasefeelingsofobjectificationwhich,inturn,reducesprosocialmotivationandbehavior.Whileagreatdealofworkhasbeendoneonthesemisperceptions,giventheriseofcommercializedLLMs(e.g.,ChatGPT),itisincreasinglyurgenttofurtherunderstandthewaysinwhichAIsobjectifyhumansandtheconsequencesthereof.aredesignedtorelyonareductionistrepresentationofusertionandthecomplexityofactualhumanpreferencesrisksamisalignmentbetweenAIrecommendationsandwhatuserscuratingcontentonsocialmediaplatformstypicallyaimtomaximizeuserengagement(Kim2017),yetusersoftheseplatformsmightprefertooptimizeforadifferentoutcome.Thisdiscordcanstemfromdivergentincentives,suchasaforareductionistalgorithmtolearnefficiently(e.g.,onlywantingtacodeliveryonrainyTuesdaysafteradrink).Addi-onlybasedontheirpastbehavior,theymaybeshownmoretemptations,atoddswiththeirgoals(Carmonetal.2019;Morewedgeetal.2023).Beyondmisalignmentinoutcomepreference,discrepan-ciescanalsoariseinhowoutcomesareoptimized.Thisistyp-icallygovernedbyanobjectivefunctionthatdefinestherel-ativedesirabilityofvariousoutcomes.Thechoiceofobjectivefunctionisconsequential,asdifferentfunctionsleadtodif-ferentalgorithmoutputs,potentiallymisaligningwithuserinterests.Forexample,inthedomainofprediction,algorithmstypicallyuseobjectivefunctionswithincreasing(e.g.,rootmeansquareerror[RMSE])orconstant(e.g.,meanabsoluteerror[MAE])sensitivitytoerror;however,peopleoftenex-hibitdecreasingsensitivitytopredictionerror(DietvorstandBharti2020).Inpracticalterms,thiscouldmean

雖然Haslam和Stratemeyer(2016)將去人性化定義”，F(xiàn)iske(2009)指出，有一種特定的去人性化形式，可以稱為AI對女性的偏見（Dastin2018）。在司法（Larsson2019）和教育（Engler2021）領(lǐng)域的AI應(yīng)用中也（Morewedge等人2023）這種真實人類個體與AI編碼的他們之間的差距可同的機會，這影響了諸如貸款決策（BertrandWeill2021）和定價（Chapdelaine2020）等關(guān)鍵結(jié)果。因此，消費者（正確地）質(zhì)疑AI欣賞他們獨特特征和情況的能力，并表現(xiàn)出在重要場合（（例如，Longoni、Bonezzi和Morewedge2019）或金融服務(wù)（Yalcin等人2022））此外，AI引入的客觀化可能產(chǎn)生嚴(yán)重的間接社會AI如通過更工具性地對待他們（例如，Onur等人，2023年）。此外，消費者認(rèn)為客觀化的AI系統(tǒng)在評估人際交往能力方面能力較低（Castelo、Bos和Lehmann，2019年）。因此，AI的客觀化性質(zhì)可AIAI

例如，Granulo（2024）發(fā)現(xiàn)，AILLM（ChatGPT）的興起，迫切需要進一步了解AI如何客觀化人類以及其后其次，由于今天AI系統(tǒng)背后的算法旨在依賴于用戶興雜性之間的任何差異都可能導(dǎo)致AI推薦與用戶真正想多決策外包給AI系統(tǒng)，這些系統(tǒng)可能會產(chǎn)生更多不匹通常旨在最大化用戶參與度（Kim，2017年），而這者追求自身福利之間的差異（Castelo等人，2023年）（Carmon等人，2019年；Morewedge等人，2023年）。法通常使用具有遞增（例如，均方根誤差RMSE）恒定（例如，平均絕對誤差MAE）敏感度的目標(biāo)函數(shù)；（Dietvorst和Bharti2020）。從實際意義上講， HowArtificialIntelligenceConstrainstheHuman Valenzuelaet

How人工智能約束人類體 Valenzuela等algorithmsoftenprioritizeavoidinglargeerrorswhenmak-ingpredictions,whileusersmaypreferthemtopursuenear-perfectpredictionsevenattheriskoflargeerrors(Dietvorst2023).Inthedomainofinvesting,forexample,misalignedobjectivefunctionscouldleadrobo-advisorstogoformoreorlessriskthanaclientdesires.strategyisdebatableinsomedomains.Manyalgorithmsarecomeasprescribedbytheirobjectivefunction.However,Forinstance,consumersoftenhaveethicalconcernsabouttrade-offs.Asaresult,theymayobjecttoanyalgorithmim-whendevelopersmayhaveworkedtoalignitwithpeople’spreferencesasmuchaspossible(DietvorstandBartelsversificationorwillingnesstoexplorenewoptions(Talaifarsuggeststhatthiscouldmakeconsumerslesshappywiththeirchoicesandpotentiallyoverall.siblyperpetuatingsubtledehumanization.Suchoversimpli-ficationrisksmisrepresentingpeople’struepreferences,potentiallyleadingtomisguideddecisions.Moreover,thisreductionistapproachmightdiscourageAIuseinsensitiveareasordeterioratehuman-AIinteractions,asindividualsfeelthatAIcannottrulygrasptheiruniqueness.REGULATEDEXPRESSIONAIsystemsrequirelargeamountsofinformationfromuserstooperateand,thus,tendtorequiresignificantself-disclosure.Suchself-disclosurehasthepotentialtobeharmful.Inthisrespect,along-standingresultinthestudyofhuman-computerprovideonline,theyarealsooftenwillingtoopenlysharetheirmostintimatethoughtsandfeelingsinsocialmediaposts,responsestoonlinesurveys,andchatbotswhentheyfeeltheyarereceivingsomethingofvalueinreturn(e.g.,Norbergetal.2007;Joinsonetal.2010;Acquistietal.2015;Tomaino,Wertenbroch,andWalters2023).Addingtot