新工科大學(xué)英語CollegeEnglishfor

NewEngineeringTechwithaconscience6Unit新工科大學(xué)英語CollegeEnglishfor

NewEngineeringAfterstudyingthisunit,youwillbeableto:explaintheethicalconcernsaboutAI;analyzethepotentialofgeneeditingindiversefieldsandthekeyethicalconsiderationsrelatedtoit;discusstechnology-inducedproblemsandtheirsolutions;talkaboutwaystodevelopandapplynewtechnologiesresponsiblythroughapaneldiscussion.Learning

objectivesCONTENTSGettingonthestage新工科大學(xué)英語CollegeEnglishfor

NewEngineeringUnlockingthetopicViewingthroughthelensExploringthefrontierSettingthesceneNewtechnologiesarerapidlytransformingourlives,bringingunprecedentedconvenienceandefficiencythatweoncecouldonlyimagine.However,alongsidetheseadvancementscomepotentialchallengesthatwehaveneverencounteredbefore.Addressingthesecomplexissuesdemandsacoordinatedeffortacrossallsectorsofsociety.Tofacilitateadialogamongdiversestakeholders,apaneldiscussionwillbeheldattheWorldInternetConference,focusingontheresponsibledevelopmentandapplicationofemergingtechnologies.Asaparticipant,youwillcollaboratewithotherstodiscusshowemergingtechnologiescanbedevelopedandappliedinwaysthatbenefitsocietywhiletheirpotentialharmcanbeminimized.

ScanthecodeandcompletetheknowledgeactivationexerciseonUcampus.Workingroups.LookatthepicturesanddiscusswhattechnologiesyouuseinyourdailylifemightinvolveAI.ThenshareyourexperiencesofinteractingwithAI.

ReferenceanswersInmydailylife,IinteractwithAItechnologymorethanIinitiallyrealized.MysmartphoneusesAIinmultipleways–fromthepredictivetextthatsuggestswordsorphrasesasItypetothefacialrecognitionthatunlocksmydevice.WhenIaskmyvirtualassistantquestionsortosetreminders,I’messentiallyconversingwithanAIsystemthatlearnsmypreferencesovertime.

ReferenceanswersStreamingservicesconstantlyemployAIalgorithmstoanalyzemyviewingandlisteninghabits,recommendingcontentthatmatchesmytasteswithsurprisingaccuracy.EvenmyemailinboxusesAItofilterspamandcategorizemessagesautomatically.Recently,I’vebeenusingAI-poweredphotoeditingappsthatcanenhanceimagequalityandevenremoveunwantedobjectsfrompictures.I’venoticedthesetoolsbecomingincreasinglysophisticated–whatoncerequiredprofessionaleditingskillsnowhappenswithasingletap.WhiletheseAIinteractionsgenerallyimprovemydailyefficiency,Isometimesfeeluneasyabouthowmuchthesesystemsknowaboutmypreferencesandbehaviors.I’mparticularlyinterestedinhowwebalanceconveniencewithprivacyasthesetechnologiesbecomemoreintegratedintoourlives.

ReferenceanswersViewingandsynthesizingmimic

vt.模仿spamn.垃圾電子郵件ubiquitousa.普遍存在的salienta.

顯著的swapvt.

換用precedevt.先于…（發(fā)生或存在）Wordbank

AIhasbecomealmostomnipresentinourdailylives,shapingourinteractions,decisions,andexperiencesincountlessways.Yet,thereisasignificant“awarenessgap”–manyofusinteractwithAI-poweredtechnologieswithoutrealizingit.Watchthevideoclipandcompletetheoutlinewithwhatyouhear.(Withsubtitles)(Withoutsubtitles)IntroductiontotheawarenessgapAIisubiquitousindailyactivities,likeexercisingwithasmartwatch,puttingona(n)1)__________________,orringingacameradoorbell.However,onlyabout2)______ofthesurveyedAmericansaccuratelyidentifiedwhethertheseactivitiesusedAI,despitethefactthatabout3)_______ofU.S.adultsareonlinedaily.(suggested)playlist60%90%AIhasbecomeagiant4)_____________,makingitdifficultforpeopletoclearlyidentifywhatfallsunderitsscope.Enabledbyadvancementsin5)__________________andthemassiveamountsofInternetdata,newtechnologiesleveragedmachinelearning,whileexistingtechnologiesreplacedtheir6)____________withitwithoutusers’awareness.CausesfortheawarenessgapumbrellatermcomputingpoweralgorithmsCallforclosingtheawarenessgapRaisingawarenessaboutAIisessentialfortalkingaboutthe8)__________________________ofitsusage.PotentialproblemfortheawarenessgapWithoutournotice,AImaymakedecisionsforusthatcarrythesame7)________asthehumandecisionsthatshapedtheAIsystem.biasesmoralandethicalboundaries

Scripts(We’realreadyusingAImorethanwerealize)N=Narrator;A=AlecTysonN:Ofalltheinteractionsyouhavewithtechnologyinaday,interactingwithartificialintelligence,ornot,feelslikeachoice.Butinsomeways,itisn’t.Overthepastdecade,we’vebecomesurroundedbyAIsystemsthatperceiveourworlds,thatsupportourdecisions,andthatmimicourabilitytocreate.Whetherwe’reawareofitornotisanotherstory.

Scripts(We’realreadyusingAImorethanwerealize)Imagineadaylikethis:Youdosomeexercisewithasmartwatch,putonasuggestedplaylist,gotoafriend’shouseandringtheircameradoorbell,browserecommendedshows,checkyourspamfolderforanemailyou’vebeenwaitingfor,andwhenyoucan’tfindit,talktoacustomersupportchatbot.Eachofthosethingsismadepossiblebytechnologiesthatfallundertheumbrellaofartificialintelligence.ButwhenasurveyaskedAmericanstoidentifywhethereachof

Scripts(We’realreadyusingAImorethanwerealize)thoseusedAIornot,theiraccuracyrateforeachquestionwasonly60%.A:SomeoftheseapplicationsofAIhavebecomefairlyubiquitous.Theyalmostexistinthebackground,andit’snotterriblyapparenttofolksthatwhetherthetoolsorservicesthey’reusingarebeingpoweredbythistechnology.

Scripts(We’realreadyusingAImorethanwerealize)N:That’sAlecTyson,oneoftheresearchersbehindthatstudy.WhenTysonandhisteamaskedrespondentshowoftentheythinktheyuseAI,almosthalfdidn’tthinktheyregularlyinteractwithitatall.Someofthemmightberight,butmostprobablyjustdon’tknowit.A:Weknowabout90%ofU.S.adultsareonlineeveryday.Somefolksareonlinealmostallthetime.Thissuggestsabitofagap

Scripts(We’realreadyusingAImorethanwerealize)wherethereseemtobesomefolkswhoreallymustbeinteractingwithAI,butit’snotverysalienttothemiftheydon’tperceiveit.N:Sowhydoesthatgapexist?Partoftheproblemisthatthetermartificialintelligencehasbeenusedtorefertoalotofdifferentthings.Artificialintelligenceistotallythisgiantumbrellatermthatnowhasbecomeakitchensinkofeverything.Inthepast,therewere

Scripts(We’realreadyusingAImorethanwerealize)distinctdisciplinesaboutwhichaspectofthehumanbrainwewantedtorecreate,like“Dowewanttorecreatethevisionpart?”,“Dowewanttorecreateourabilitytohear,ourabilitytowriteandspeak?”Givingamachinetheabilitytoseebecamethefieldofcomputervision.Givingamachinetheabilitytowriteandspeakbecamethefieldofnaturallanguageprocessing.Butontheirown,thesetasks

Scripts(We’realreadyusingAImorethanwerealize)stillrequiredamachinetobeprogrammed.Ifwewantedmachinestorecognizespamemails,wehadtoexplicitlyprogramthemtolookoutforspecificthingslikepoorspellingandurgentphrasing.Thatmeantthetoolsweren’tveryadaptabletocomplexsituations.Butthatallchangedwhenwestartedrecreatingthebrain’sabilitytolearn.Thisbecamethesubfieldofmachinelearning,wherecomputersaretrainedonmassiveamountsofdatasothat

Scripts(We’realreadyusingAImorethanwerealize)insteadofneedingtohand-coderulesaboutwhattosee,orspeak,orwrite,thecomputerscandeveloprulesontheirown.Withmachinelearning,acomputercouldlearntorecognizenewspamemailsbyreviewingthousandsofexistingemailsthathumanshavelabeledasspam.Themachinerecognizespatternsinthisstructureddataandcreatesitsownrulestohelpidentifythosepatterns.Whenthattrainingdatahasn’tbeenstructuredandlabeledbyhumans,

Scripts(We’realreadyusingAImorethanwerealize)thatmethodiscalleddeeplearning.MostofthetimepeopletalkaboutAInow,they’renottalkingaboutthewholefieldbutspecificallythesetwomethods.Improvementsincomputingpower,togetherwiththemassiveamountsofdatageneratedontheInternet,madepossibleawholenewgenerationoftechnologiesthatleveragedmachinelearning,andexistingonesswappedtheiralgorithmsformachinelearningtoo.

Scripts(We’realreadyusingAImorethanwerealize)Nowmachinelearninganddeeplearningmodelspowerrecommendationsforshows,music,videos,products,andadvertisements.Theydeterminetherankingofitemseverytimewebrowsesearchresultsorsocialmediafeeds.Theyrecognizeimageslikefacestounlockphonesorusefilters,andthehandwritingonremotedepositchecks.Theyrecognizespeechintranscription,voiceassistants,andvoice-enabledTVremotes,andtheypredicttextinautocompleteandautocorrect.

Scripts(We’realreadyusingAImorethanwerealize)ButAIisseepingintomorethanthat.Machinelearningalgorithmsarealreadybeingusedtodecidewhichadswesee,whichjobswequalifyfor,andwhetherwequalifyforloansorgovernmentbenefits,andoftencarrythesamebiasesasthehumandecisionsthatprecededthem.Andthat’spartofthereasonwhyitmatterstoclosethatgapbetweenthosewhoknowinglyinteractwithAIeverydayandthosewhodon’tquiteknowityet.

Scripts(We’realreadyusingAImorethanwerealize)A:Awarenessneedstogrowforfolkstobeabletoparticipateinsomeoftheseconversationsaboutthemoralandethicalboundaries(of)whatAIshouldbeusedforandwhatitshouldn’tbeusedfor.Workingroupsanddiscussthequestions.Basedonthevideoclipandyourpersonalexperiences,identifyspecificsituationswhereindividualsinteractwithAIwithoutrecognizingit.HowmightalackofawarenessaboutAIaffectindividuals,andwhatarethebroaderimplicationsforsociety?WhatapproachescanbeadoptedtohelpraisepublicawarenessaboutAIandencouragemoreinformedinteractionswithAI-relatedtechnologies?ReferenceanswersInboththevideoexamplesandmypersonalexperience,therearenumeroussituationswherepeopleinteractwithAIwithoutconsciouslyrecognizingit.Whentheyscrollthroughsocialmediafeeds,sophisticatedalgorithmscuratecontenttailoredtotheirinterestsandpastinteractions.Similarly,whentheyusenavigationapps,AIsilentlyoptimizestheirroutebasedonreal-timetrafficdata,roadconditions,andtheirdrivingpatterns.EvenroutineactivitieslikeonlineshoppinginvolveAI-drivenproductrecommendationsanddynamicpricingthatoftengounnoticed.ReferenceanswersThelackofawarenessabouttheseAIinteractionshassignificantimplications.Atanindividuallevel,itcreatesanillusionofindependentdecision-makingwhilesubtlyinfluencingourchoices.Forinstance,ImightbelieveIindependentlydiscoveredaproduct,unawarethatAIalgorithmsstrategicallybroughtittomyattentionbasedonmydigitalprofile.Thisinvisibilitymakesitdifficulttodistinguishbetweenourauthenticpreferencesandalgorithmically-shapedones.ReferenceanswersThebroadersocietalimplicationsareevenmoreconcerning.WhenpopulationsunknowinglyconsumenewsandinformationfilteredthroughAIsystems,itcanleadtoinformationbubblesandpolarization.Thealgorithmicamplificationofengagingbutpotentiallymisleadingcontentthreatenssharedunderstandingoffacts.ViewingthroughthelensReferenceanswersThiswidespreadyetunrecognizedAIinfluenceraisesfundamentalquestionsaboutautonomyandinformedconsentinthedigitalsociety.ReferenceanswersRaisingpublicawarenessaboutAIrequiresamulti-facetedapproachthataccommodatesdiverselearningstylesandreachesallsegmentsofsociety.SchoolsshouldintegrateAIliteracyintocurriculafromelementaryschoolstouniversities,teachingnotonlyhowtouseAItoolsbutalsohowtocriticallyassesstheirimpacts.TheselessonsshoulddemystifyAIbyexplainingcoreconceptsinaccessiblelanguage,emphasizingbothitscapabilitiesandlimitations.ReferenceanswersThemediaplaysakeyroleinshapingpublicperceptions.Documentaries,podcasts,andnewsfeaturescantranslatecomplextechnologiesintorelatablestories.Community-basedinitiatives,suchasworkshops,techfairs,andpublicforums,cancreateopportunitiesforhands-onlearninganddiscussionaboutAI.PubliclibrariescouldprovideAIliteracyprogramsalongsidetraditionaldigitaltraining,helpingbridgethegapforunderservedpopulations,includingtheelderly.ViewingthroughthelensReferenceanswersFinally,policymakersshouldcreateclearregulatoryframeworksthatpromotetransparency.ThiscouldincludestandardizedAIimpactlabelsthatdisclosehowdataisused,howdecisionsaremade,andwhatpotentialbiasesexist–encouragingmoreinformedandresponsibleengagementwithAItechnologies.v新工科大學(xué)英語CollegeEnglishfor

NewEngineeringExploringthefrontierReading

1Reading

2HowbadalgorithmskeepsusfromgoodAI1

Sayacomputerandahumanwerepittedagainsteachotherinabattleforneutrality.Whodoyouthinkwouldwin?Plentyofpeoplewouldbetonthemachine.Butthisisthewrongquestion.2

Humanscreatedcomputers,andtheydesignandtrainthesystemsthatmakemoderntechnologywork.Asthesesystemsarebuilt,theyinevitablyembodythebiasesoftheirhumancreators.WhenpeoplerefertoAIbias,thisis,inessence,whattheyaretalkingabout.Likehumanbias,AIbias,whentranslatedintodecisionsoractions,becomesdiscrimination.3

AI-poweredsystemsaretrainedonsetsofexistingdata,likephotos,videos,audiorecordings,ortext.Thisinvolvesexposingacomputertodatasoitlearnstomakejudgmentsorpredictionsbasedonthepatternsitnotices.Forinstance,toteachacomputertorecognizeadesk,youprovidedataonmetricslikematerial,weight,anddimensions.Youshowthecomputerthesemetricsrepeatedly,tellingitwhichobjectsaredesks.Aftercontinuousrefinement,thesystemshouldbeabletopredictdesk-likeobjectsindependently.However,thissimplicityfadesinmorecomplexsituationswheremachinelearningsystemsoftenfaceissueslikeincompleteorimbalancedtrainingdata,whichleadstoalgorithmicbias.4Forinstance,facialrecognitionsoftwareneedsphotostolearnhowtospotfaces,butifthedatasetitistrainedoncontainsphotosthatdepictmostlywhitepeople,thesystemmightnotworkaswellonnon-whitefaces.AnAI-poweredcaptioningprogrammightnotbeabletoaccuratelytranscribeEnglishspokenwithaslightforeignaccentifthataccentisn’trepresentedintheaudioclipsinitstrainingdatabase.AIcanonlylearnfromwhatithasbeengiven.Infact,speechrecognitiontechnologieshavealong-standinghistoryoffailingincertainscenarios.TheymightnotrecognizerequestsfrompeoplewhodonotspeakEnglishastheirfirstlanguage.Whilesomepeoplemaychoosetoavoidtheseproblemsbynotusingthesetechnologies,thesefailurescanbeparticularlydevastatingforthosewithdisabilitieswhomayrelyonvoice-activatedtechnologies.5Amiddiscussionsofalgorithmicbias,AIsystemcreatorsoftenclaimtheyaretryingtoremovebiasintroducedbydatasets.However,certainmethodsfallshortoftheirintendedobjectives.Makinganalgorithm“blind”toanattributelikeraceorgenderdoesn’tmeanthattheAIwon’tfindotherwaystointroducebiasesintoitsdecision-makingprocess–andperhapstoidentifythesameattributesitwassupposedtoignore.Forinstance,asystemthatisdesignedtoassessapplicationsforajobmightberendered“blind”toanapplicant’sgenderbutlearntodistinguishmale-soundingandfemale-soundingnames,orlookforotherindicatorsintheirCV,likeadegreefromanall-women’scollege,ifthedatasetitistrainedonpredominantlyfavorsmaleapplicants.6Whilecreatingmorerepresentativedatasetsmightbepartofthesolution,notalleffortstobuildbetterdatasetsareethical.Forinstance,somefacialrecognitioncompanieshavecontroversiallyharvestedpubliclyavailableimagesfromsocialmediawithoutuserconsenttoimprovethediversityoftheirdatasets,whichraisessignificantprivacyconcerns.Andit’snotjustaboutthedata.AIcanalsobedesignedtoframeproblemsinfundamentallyproblematicways.Forexample,analgorithmdesignedtodetermine“creditworthiness”thatprioritizesmaximizingprofitcouldultimatelydecidetoissuepredatory,subprimeloans,targetingvulnerablepopulationswithunfairlendingpractices.7WewilllikelyneednewregulationsandpoliciestoregulateAI.Theencouragingnewsisthatanincreasingnumberofcountriesaretakingproactivemeasurestoeliminate,oratleastmitigate,algorithmicbiasinAI.InChina,aninterimregulationonthemanagementofgenerativeAIservicescameintoeffectin2023,mandatingthatserviceprovidersshouldpreventdiscriminationbasedonfactorssuchasethnicity,gender,andagethroughoutallstages,includingalgorithmdesign,trainingdataselection,modeldevelopmentandoptimization,andservicedelivery.InNovember2023,theworld’sfirstAISafetySummitwashostedintheU.K.TheEuropeanUnionand28countries,includingChina,theU.S.,andtheU.K.,signedtheBletchleyDeclaration,alandmarkinternationalaccordthatrecognizestheneedforAIdevelopmentandusagetobe“human-centric,trustworthy,andresponsible.”Itemphasizestheimportanceoftransparencyandexplainability,fairness,accountability,regulation,biasmitigation,andmore.8WillAIeverbeunbiased?Yesandno.Whileachievingalgorithmicneutralityremainstheoreticallyconceivable,itisunlikelythatanentirelyimpartialAIcanexistifanentirelyimpartialhumanminddoesnot.Ultimately,algorithmicbiasisahumanproblemandtheonlysolutionistobegineliminatingbiasinallaspectsofourpersonalandsociallives.Thisentailsfosteringgenuinedemographicdiversityandinclusiverepresentationwithinprofessionalenvironments,educationalinstitutions,politicalsystems,andotherdomainsofourlives.Ifwewanttoimproveouralgorithms,wemustfirstimproveourselves.BackgroundinformationAlgorithmsaresetsofinstructionsorrulesthatcomputersfollowtosolveproblemsormakedecisions.InAI,algorithmsareusedtoprocesslargeamountsofdata,recognizepatterns,andmakepredictionsorchoices–oftenexecutingtasksindependently.Fromsortingemailsandrecommendingvideostoassistinginmedicaldiagnosesanddeterminingloaneligibility,algorithmsincreasinglyinfluencecrucialaspectsofourlives.Whilealgorithmscangreatlyenhanceefficiencyandaccuracy,theyultimatelyreflectthedatatheyaretrainedonandtheprioritiesprogrammedbytheirhumancreators–arealitythatraisesimportantquestionsaboutpotentialbias.Sayacomputerandahumanwerepittedagainsteachotherinabattleforneutrality.[Notes]:Inthissentence,“say”functionsasamarkerforintroducingahypotheticalscenarioorthoughtexperiment.Here,“say”workssimilarlyto“suppose,”“imagine,”or“l(fā)et’sconsider”toinvitethereadertoentertainatheoreticalsituation.Itisaconversationalwaytoframeanabstractconcept,makingcomplexideasmoreaccessiblebygroundingtheminaconcrete,albeithypothetical,situation.Butthisisthewrongquestion.[Notes]:Thisisa“wrongquestion”becauseitoversimplifiestheconceptofneutrality.Itassumesthatneutralityisaninherentqualityofeitherhumansorcomputers,butfailstoacknowledgethatcomputersarecreated,designed,andtrainedbyhumans.Asaresult,biasespresentinhumansinevitablytransfertothesystemstheybuild.Insteadofasking“whowouldwin,”thefocusshouldbeonunderstandingandaddressingthesourcesofbias.translatedintoThephrase“translate(sth.)into(sth.)”means“render,convert,transform,orexpresssth.fromoneformorstateintoanother(將……從一種形式或狀態(tài)呈現(xiàn)、轉(zhuǎn)換、轉(zhuǎn)化或表達(dá)為另一種形式或狀態(tài)).”Thisphraseisparticularlyusefulfordescribinghowideas,efforts,orconditionstransformintotangibleoutcomesorconsequences.e.g.Herhardworkanddedicationtranslatedintoexceptionalresultsinthefinalproject.metric

n.[C]asystemorstandardofmeasurement

計(jì)量體系；衡量標(biāo)準(zhǔn)a.usingorconnectedwiththemetricsystemofweightsandmeasures公制的；米制的fallshortof

Thephrase“fallshortofsth.”means“belessthanwhatyouneed,expected,orhopedfor,orfailtoreachasatisfactorystandard(達(dá)不到[目的、期望、標(biāo)準(zhǔn)]).”e.g.Thefilmadaptationfellshortofcapturingtheemotionaldepthpresentintheoriginalnovel.Makinganalgorithm“blind”toanattributelikeraceorgenderdoesn’tmeanthattheAIwon’tfindotherwaystointroducebiasesintoitsdecision-makingprocess–andperhapstoidentifythesameattributesitwassupposedtoignore.[Notes]:Theword“blind”isusedmetaphoricallyhere,whichmeansintentionallypreventinganalgorithmfrom“seeing”or“considering”specificdataattributes(likeraceorgender).Metaphorsareoftenemployedintechnicalwritingtomakeabstractconceptsmoreconcreteandrelatable.creditworthiness

n.[U]thedegreetowhichaperson,organization,orcountryisconsideredlikelytopaybackmoneythattheyborrow信譽(yù)度；信用等級(jí)

subprime

a.(ofaloan)madetoaborrowerwithapoorcreditrating,usu.atahighrateofinterest（貸款）次級(jí)的

Forexample,analgorithmdesignedtodetermine“creditworthiness”thatprioritizesmaximizingprofitcouldultimatelydecidetoissuepredatory,subprimeloans,targetingvulnerablepopulationswithunfairlendingpractices.[Notes]:Asubprimeloan(次級(jí)貸款)isaloanmadetoaborrowerwhoisnoteligibleforthebestmarketrates(knownasprimerates),butratheratahigherrateofinterestbecauseofincreasedriskfactors.Subprimeborrowersusuallyhavepoororlimitedcredithistoriesandaretypicallyperceivedasriskierthanprimeborrowersduetotheirlackofcollateralorlowincome.Inordertocompensateforincreasedrisk,lenderschargesubprimeborrowersapremium.generativeAI

n.[U]theuseorstudyofAIthatisabletoproducetext,images,etc.生成式人工智能Itemphasizestheimportanceoftransparencyandexplainability,fairness,accountability,regulation,biasmitigation,andmore.[Notes]:Here,“andmore”functionsasanopen-endedphrasethatindicatesthelistisnotexhaustive.Thisisparticularlyusefulinacademicorpolicydiscussionswhereprovidingacompletelistmightbeimpractical,orwheretheauthorwantstofocusonkeyexampleswhileacknowledgingthebroaderscopeofthetopic.Similarexpressionsinclude“amongothers,”“tonameafew,”“andsoon,”and“andthelike.”demographic

a.

relatingtothepopulationandgroupsofpeople

人口的；人口統(tǒng)計(jì)的

Whileachievingalgorithmicneutralityremainstheoreticallyconceivable,itisunlikelythatanentirelyimpartialAIcanexistifanentirelyimpartialhumanminddoesnot.[Notes]:Therepetitionof“entirelyimpartial”beforeboth“AI”and“humanmind”createsapowerfulparallelstructurethatemphasizesthedirectconnectionbetweenthelimitationsofAIsystemsandthoseoftheirhumancreators.ItmeansthatAIsystemscannotexceedtheirdesignersinimpartiality.Thisrhetoricalstrategyisreinforcedinthepassage’sconcludingsentence,wheretheparallelstructures“improveouralgorithms”and“improveourselves”highlighttheinterdependencebetweenadvancingAIsystemsandaddressinghumanbiases.糟糕的算法如何阻礙我們獲得優(yōu)秀的人工智能1

如果讓計(jì)算機(jī)和人類在中立性上一較高下，你覺得誰會(huì)贏？很多人可能會(huì)把賭注押在機(jī)器身上。但其實(shí)這個(gè)問題本身就是錯(cuò)的。2

人類創(chuàng)造了計(jì)算機(jī)，也設(shè)計(jì)并訓(xùn)練了現(xiàn)代科技系統(tǒng)的運(yùn)行機(jī)制。在這些系統(tǒng)被構(gòu)建的過程中，它們不可避免地承載了人類研發(fā)者的偏見。人們所說的人工智能偏見，本質(zhì)上指的就是這種現(xiàn)象。就像人類偏見一樣，這種偏見一旦被轉(zhuǎn)化為決策或行動(dòng)，就會(huì)演變?yōu)槠缫暋?

基于人工智能的系統(tǒng)通常通過大量已有數(shù)據(jù)進(jìn)行訓(xùn)練，例如照片、視頻、音頻或文本。訓(xùn)練過程會(huì)讓計(jì)算機(jī)接觸數(shù)據(jù)，使其學(xué)會(huì)根據(jù)識(shí)別出的模式做出判斷或預(yù)測(cè)。例如，為了教會(huì)計(jì)算機(jī)識(shí)別“書桌”，我們會(huì)提供關(guān)于材質(zhì)、重量、尺寸等參數(shù)數(shù)據(jù)。我們會(huì)向計(jì)算機(jī)反復(fù)呈現(xiàn)這些參數(shù)，告訴它哪些是書桌。經(jīng)過持續(xù)優(yōu)化，這個(gè)系統(tǒng)最終可以獨(dú)立識(shí)別書桌類物體。然而，更復(fù)雜的情境下，這種簡(jiǎn)單的訓(xùn)練模式就會(huì)失效，機(jī)器學(xué)習(xí)系統(tǒng)常常會(huì)遇到訓(xùn)練數(shù)據(jù)不完整或不均衡等問題，從而導(dǎo)致算法偏見。4

例如，面部識(shí)別軟件需要通過照片學(xué)習(xí)如何識(shí)別人臉，但如果訓(xùn)練數(shù)據(jù)集中大多數(shù)是白人照片，那么系統(tǒng)在識(shí)別非白人面孔時(shí)的準(zhǔn)確率就可能大打折扣。人工智能驅(qū)動(dòng)的字幕程序如果未接受含有外國口音的英語音頻片段訓(xùn)練，就可能無法準(zhǔn)確轉(zhuǎn)錄稍微帶有該口音的英語。人工智能只能從它所獲得的內(nèi)容中進(jìn)行學(xué)習(xí)。事實(shí)上，語音識(shí)別技術(shù)在某些場(chǎng)景下長期無法取得技術(shù)突破。它們可能無法識(shí)別非英語母語人士的請(qǐng)求。雖然有些人可以選擇不使用此類技術(shù)來避免這些問題，但對(duì)那些有可能依賴語音控制設(shè)備的殘障人士而言，這些技術(shù)的失效可能會(huì)帶來尤其嚴(yán)重的后果。5

在討論算法偏見時(shí)，人工智能系統(tǒng)的研發(fā)者常常聲稱他們正在努力消除由數(shù)據(jù)集帶來的偏見。然而，一些方法并未達(dá)到其預(yù)期效果。讓算法對(duì)種族或性別等屬性“視而不見”，并不意味著系統(tǒng)不會(huì)通過其他方式引入偏見——甚至可能識(shí)別出原本應(yīng)該忽略的特征。例如，一個(gè)用于評(píng)估工作申請(qǐng)的系統(tǒng)可能被設(shè)計(jì)成對(duì)申請(qǐng)人的性別“視而不見”，但如果其訓(xùn)練數(shù)據(jù)集主要偏向男性申請(qǐng)人，它可能會(huì)學(xué)會(huì)區(qū)分男性化和女性化的名字，或者通過簡(jiǎn)歷中的其他指標(biāo)（比如從女子學(xué)院取得的學(xué)位）來判斷性別。6

雖然構(gòu)建更具代表性的數(shù)據(jù)集可以解決部分問題，但并非所有改進(jìn)數(shù)據(jù)集的手段都符合道德規(guī)范。例如，一些面部識(shí)別公司為提升數(shù)據(jù)集的多樣性，未經(jīng)用戶同意就從社交媒體中抓取公開照片，這引發(fā)了嚴(yán)重的隱私擔(dān)憂。而且問題也并不僅關(guān)乎數(shù)據(jù)本身。人工智能在構(gòu)建問題方式上存在根本性設(shè)計(jì)缺陷。比如，一個(gè)旨在評(píng)估“信譽(yù)度”的算法，如果以利潤最大化為首要目標(biāo)，最終可能會(huì)決定發(fā)放掠奪性的次級(jí)貸款，助長針對(duì)弱勢(shì)群體的不公平借貸。7

我們很可能需要新的法規(guī)和政策來規(guī)范人工智能。令人鼓舞的是，越來越多的國家正積極采取措施，試圖消除或至少緩解人工智能中的算法偏見。在中國，2023年生效的生成式人工智能服務(wù)管理暫行辦法規(guī)定服務(wù)提供者應(yīng)在算法設(shè)計(jì)、訓(xùn)練數(shù)據(jù)選擇、模型生成和優(yōu)化、服務(wù)提供等過程中，防止產(chǎn)生民族、性別、年齡等歧視。2023年11月，全球首屆人工智能安全峰會(huì)在英國舉辦。歐盟及包括中國、美國和英國在內(nèi)的28個(gè)國家共同簽署了《布萊奇利宣言》。這一具有里程碑意義的國際協(xié)議認(rèn)為人工智能的發(fā)展和使用需要“以人為本、可信并且負(fù)責(zé)任”，同時(shí)強(qiáng)調(diào)透明度和可解釋性、公平性、問責(zé)制、監(jiān)管、偏見減緩等方面的重要性。8

人工智能是否會(huì)變得完全沒有偏見？是，也不是。雖然實(shí)現(xiàn)算法中立在理論上是可能的，但如果不存在完全公正的人類大腦，那么也不太可能存在完全公正的人工智能。歸根結(jié)底，算法偏見是人類的問題，唯一的解決方案是我們從個(gè)人和社會(huì)生活的各個(gè)方面開始消除偏見。這意味著我們需要在職場(chǎng)、教育機(jī)構(gòu)、政治體系及其他各個(gè)生活領(lǐng)域中，真正實(shí)現(xiàn)人口多樣性和包容性代表。若想改進(jìn)算法，必先改善自己。GlobalunderstandingReadthepassageandcompletethesummarywithinformationfromthepassage.IntroductiontoAIbiasAIbiasreflectsthebiasesofAIsystems’1)__________________.Whentranslatedinto2)___________________,AIbiasbecomesdiscrimination.(human)creatorsdecisionsoractions

OriginofAIbiasAIsystemsaretrainedonsetsof3)_____________.Whentrainingdatais4)_______________________,itleadstoalgorithmicbias.AIcanbedesignedto5)_______________inwaysthatarefundamentallyflawed.existingdataincompleteorimbalanced

frameproblemsPotentialsolutionsforAIbiasCreating6)____________________datasetsinethicalwaysImplementingnew7)______________________toregulateAIEliminatingbiasinallaspectsofour

8)_________________