The Atlas of AI_ Power, Politics, and the Planetary Costs of Artificial Inte.pdf
《The Atlas of AI_ Power, Politics, and the Planetary Costs of Artificial Inte.pdf》由会员分享,可在线阅读,更多相关《The Atlas of AI_ Power, Politics, and the Planetary Costs of Artificial Inte.pdf(336页珍藏版)》请在冰豆网上搜索。
AtlasofAIThispageintentionallyleftblankAtlasofAIPower,Politics,andthePlanetaryCostsofArtificialIntelligenceKATECRAWFORDNewHavenandLondonCopyright2021byKateCrawford.Allrightsreserved.Thisbookmaynotbereproduced,inwholeorinpart,includingillustrations,inanyform(beyondthatcopyingpermittedbySections107and108oftheU.S.CopyrightLawandexceptbyreviewersforthepublicpress),withoutwrittenpermissionfromthepublishers.YaleUniversityPressbooksmaybepurchasedinquantityforeducational,business,orpromotionaluse.Forinformation,pleasee-mailsales.pressyale.edu(U.S.office)orsalesyaleup.co.uk(U.K.office).CoverdesignandchapteropeningillustrationsbyVladanJoler.SetinMinionbyTsengInformationSystems,Inc.PrintedintheUnitedStatesofAmerica.LibraryofCongressControlNumber:
2020947842ISBN978-0-300-20957-0(hardcover:
alk.paper)AcataloguerecordforthisbookisavailablefromtheBritishLibrary.ThispapermeetstherequirementsofANSI/NISOZ39.48-1992(PermanenceofPaper).10987654321ForElliottandMargaretThispageintentionallyleftblankContentsIntroduction1ONe.Earth23twO.Labor53three.Data89fOur.Classification123fIve.Affect151SIx.State181CONCluSION.Power211COdA.Space229Acknowledgments239Notes245Bibliography269Index315IntroductionTheSmartestHorseintheWorldAttheendofthenineteenthcentury,EuropewascaptivatedbyahorsecalledHans.“CleverHans”wasnothinglessthanamarvel:
hecouldsolvemathproblems,telltime,identifydaysonacalendar,dif-ferentiatemusicaltones,andspelloutwordsandsentences.PeopleflockedtowatchtheGermanstalliontapoutanswerstocomplexproblemswithhishoofandconsistentlyarriveattherightanswer.“Whatistwoplusthree?
”Hanswoulddili-gentlytaphishoofonthegroundfivetimes.“Whatdayoftheweekisit?
”Thehorsewouldthentaphishooftoindicateeachletteronapurpose-builtletterboardandspelloutthecorrectanswer.Hansevenmasteredmorecomplexquestions,suchas,“Ihaveanumberinmind.Isubtractnineandhavethreeasaremainder.Whatisthenumber?
”By1904,CleverHanswasaninternationalcelebrity,withtheNewYorkTimeschampioninghimas“BerlinsWonderfulHorse;HeCanDoAlmostEvery-thingbutTalk.”1Hansstrainer,aretiredmathteachernamedWilhelmvonOsten,hadlongbeenfascinatedbyanimalintelligence.2IntroductionVonOstenhadtriedandfailedtoteachkittensandbearcubscardinalnumbers,butitwasntuntilhestartedworkingwithhisownhorsethathehadsuccess.HefirsttaughtHanstocountbyholdingtheanimalsleg,showinghimanumber,andthentappingonthehoofthecorrectnumberoftimes.SoonHansrespondedbyaccuratelytappingoutsimplesums.NextvonOstenintroducedachalkboardwiththealphabetspelledout,soHanscouldtapanumberforeachletterontheboard.Aftertwoyearsoftraining,vonOstenwasastoundedbytheanimalsstronggraspofadvancedintellectualconcepts.SohetookHansontheroadasproofthatanimalscouldreason.Hansbecametheviralsensationofthebellepoque.Butmanypeoplewereskeptical,andtheGermanboardofeducationlaunchedaninvestigativecommissiontotestVonOstensscientificclaims.TheHansCommissionwasledbythepsychologistandphilosopherCarlStumpfandhisassis-tantOskarPfungst,anditincludedacircusmanager,aretiredschoolteacher,azoologist,aveterinarian,andacavalryofficer.YetafterextensivequestioningofHans,bothwithhistrainerpresentandwithout,thehorsemaintainedhisrecordofcor-rectanswers,andthecommissioncouldfindnoevidenceofdeception.AsPfungstlaterwrote,Hansperformedinfrontof“thousandsofspectators,horse-fanciers,trick-trainersoffirstrank,andnotoneofthemduringthecourseofmanymonthsobservationsareabletodiscoveranykindofregularsignal”betweenthequestionerandthehorse.2ThecommissionfoundthatthemethodsHanshadbeentaughtweremorelike“teachingchildreninelementaryschools”thananimaltrainingandwere“worthyofscientificexamination.”3ButStrumpfandPfungststillhaddoubts.Onefindinginparticulartroubledthem:
whenthequestionerdidnotknowtheanswerorwasstandingfaraway,Hansrarelygavethecorrectanswer.ThisledPfungstandStrumpftocon-Introduction3siderwhethersomesortofunintentionalsignalhadbeenpro-vidingHanswiththeanswers.AsPfungstwoulddescribeinhis1911book,theirintu-itionwasright:
thequestionersposture,breathing,andfacialexpressionwouldsubtlychangearoundthemomentHansreachedtherightanswer,promptingHanstostopthere.4Pfungstlatertestedthishypothesisonhumansubjectsandconfirmedhisresult.Whatfascinatedhimmostaboutthisdiscoverywasthatquestionersweregenerallyunawarethattheywereprovidingpointerstothehorse.ThesolutiontotheCleverHansriddle,Pfungstwrote,wastheunconsciousdi-rectionfromthehorsesquestioners.5Thehorsewastrainedtoproducetheresultshisownerwantedtosee,butaudiencesfeltthatthiswasnottheextraordinaryintelligencetheyhadimagined.ThestoryofCleverHansiscompellingfrommanyangles:
therelationshipbetweendesire,illusion,andaction,thebusi-nessofspectacles,howweanthropomorphizethenonhuman,WilhelmvonOstenandCleverHans4Introductionhowbiasesemerge,andthepoliticsofintelligence.Hansin-spiredaterminpsychologyforaparticulartypeofconceptualtrap,theCleverHansEffectorobserver-expectancyeffect,todescribetheinfluenceofexperimentersunintentionalcuesontheirsubjects.TherelationshipbetweenHansandvonOstenpointstothecomplexmechanismsbywhichbiasesfindtheirwaysintosystemsandhowpeoplebecomeentangledwiththephenomenatheystudy.ThestoryofHansisnowusedinma-chinelearningasacautionaryreminderthatyoucantalwaysbesureofwhatamodelhaslearnedfromthedataithasbeengiven.6Evenasystemthatappearstoperformspectacularlyintrainingcanmaketerriblepredictionswhenpresentedwithnoveldataintheworld.Thisopensacentralquestionofthisbook:
Howisintel-ligence“made,”andwhattrapscanthatcreate?
Atfirstglance,thestoryofCleverHansisastoryofhowonemanconstructedintelligencebytrainingahorsetofollowcuesandemulatehumanlikecognition.Butatanotherlevel,weseethattheprac-ticeofmakingintelligencewasconsiderablybroader.Theen-deavorrequiredvalidationfrommultipleinstitutions,includ-ingacademia,schools,science,thepublic,andthemilitary.ThentherewasthemarketforvonOstenandhisremarkablehorseemotionalandeconomicinvestmentsthatdrovethetours,thenewspaperstories,andthelectures.Bureaucraticau-thoritieswereassembledtomeasureandtestthehorsesabili-ties.Aconstellationoffinancial,cultural,andscientificinter-estshadaparttoplayintheconstructionofHanssintelligenceandastakeinwhetheritwastrulyremarkable.Wecanseetwodistinctmythologiesatwork.Thefirstmythisthatnonhumansystems(beitcomputersorhorses)areanaloguesforhumanminds.Thisperspectiveassumesthatwithsufficienttraining,orenoughresources,humanlikeintel-ligencecanbecreatedfromscratch,withoutaddressingtheIntroduction5fundamentalwaysinwhichhumansareembodied,relational,andsetwithinwiderecologies.Thesecondmythisthatintelli-genceissomethingthatexistsindependently,asthoughitwerenaturalanddistinctfromsocial,cultural,historical,andpoliti-calforces.Infact,theconceptofintelligencehasdoneinordi-nateharmovercenturiesandhasbeenusedtojustifyrelationsofdominationfromslaverytoeugenics.7Thesemythologiesareparticularlystronginthefieldofartificialintelligence,wherethebeliefthathumanintelligencecanbeformalizedandreproducedbymachineshasbeenaxi-omaticsincethemid-twentiethcentury.JustasHanssintel-ligencewasconsideredtobelikethatofahuman,fosteredcarefullylikeachildinelementaryschool,soAIsystemshaverepeatedlybeendescribedassimplebuthumanlikeformsofintelligence.In1950,AlanTuringpredictedthat“attheendofthecenturytheuseofwordsandgeneraleducatedopinionwillhavealteredsomuchthatonewillbeabletospeakofmachinesthinkingwithoutexpectingtobecontradicted.”8Themathe-maticianJohnvonNeumannclaimedin1958thatthehumannervoussystemis“primafaciedigital.”9MITprofessorMarvinMinskyoncerespondedtothequestionofwhethermachinescouldthinkbysaying,“Ofcoursemachinescanthink;wecanthinkandwearemeatmachines.”10Butnoteveryonewasconvinced.JosephWeizenbaum,earlyAIinventorandcreatorofthefirstchatbotprogram,knownaselIZA,believedthattheideaofhumansasmereinformationprocessingsystemsisfartoosimplisticanotionofintelligenceandthatitdrovethe“perversegrandfantasy”thatAIscientistscouldcreateama-chinethatlearns“asachilddoes.”11Thishasbeenoneofthecoredisputesinthehistoryofartificialintelligence.In1961,MIThostedalandmarklectureseriestitled“ManagementandtheComputeroftheFuture.”Astellarlineupofcomputerscientistsparticipated,including6IntroductionGraceHopper,J.C.R.Licklider,MarvinMinsky,AllenNewell,HerbertSimon,andNorbertWiener,todiscusstherapidad-vancesbeingmadeindigitalcomputing.Atitsconclusion,JohnMcCarthyboldlyarguedthatthedifferencesbetweenhumanandmachinetaskswereillusory.Thereweresimplysomecomplicatedhumantasksthatwouldtakemoretimetobeformalizedandsolvedbymachines.12ButphilosophyprofessorHubertDreyfusarguedback,concernedthattheassembledengineers“donotevenconsiderthepossibilitythatthebrainmightprocessinformationinanentirelydifferentwaythanacomputer.”13InhislaterworkWhatComputersCantDo,Dreyfuspointedoutthathumanintelligenceandexpertiserelyheavilyon