2023生成式人工智能的潜在有害影响与未来之路

epic.org ELECTRONICPRIVACYINFORMATIONCENTER GENERATNG HARMS GenerativeAl'sImpact& PathsForward MAY2023 CONTRIBUTIONSBY GrantFergussonTomMcBrien CaitrionaFitzgeraldCalliSchroeder ChrisFrascellaBenWinters MeganlorioEnidZhou EDITEDBY GrantFergusson,CalliSchroeder,BenWinters,andEnidZhou ThankyoutoSarahMyersWestandKatharinaKoppforyourgenerous commentsonanearlierdraftofthepaper. NotesonthisPaper: Thisisversion1ofthispaperandisreflectiveofdocumentedandanticipatedharmsofGenerativeAlasofMay15,2023.Duetothefast- changingpaceofdevelopment,use,andharmsofGenerativeAl,wewantto acknowledgethatthisisaninherentlydynamicpaper,subjecttochangesin thefuture. Throughoutthispaper,weuseastandardformattoexplainthetypologyof harmsthatgenerativeAlcanproduce.EachsectionfirstexplainsrelevantbackgroundinformationandpotentialrisksimposedbygenerativeAl,thenhighlightsspecificsharmsandinterventionsthatscholarsandregulators havepursuedtoremedyeachharm.Thispaperdrawsontwotaxonomiesof A.l.harmstoguideouranalysis: 1.DanielleCitron'sandDanielSolove'sTypologyofPrivacyHarms,comprisingphysical,economic,reputational,psychological,autonomy,discrimination,andrelationshipharms;land 2.JoyBuolamwini'sTaxonomyofAlgorithmicHarms,comprisinglossofopportunity,economicloss,andsocialstigmatization,includinglossofliberty,increasedsurveillance,stereotypereinforcement,andother dignitaryharms.2 ThesetaxonomiesdonotnecessarilycoverallpotentialAlharms,andouruseofthesetaxonomiesismeanttohelpreadersvisualizeandcontextualizeAlharmswithoutlimitingthetypesandvarietyofAlharmsthatreadersconsider. TableofContents Introduction. TurbochargingInformationManipulation. Harassment,Impersonation,andExtortion...9 Spotlight:Section230..19 ProfitsOverPrivacy:IncreasedOpagueDataCollection..24 IncreasingDataSecurityRisk..30 ConfrontingCreativity:ImpactonIntellectualPropertyRights..33 ExacerbatingEffectsofClimateChange.40 LaborManipulation,Theft,andDisplacement.44 Spotlight:Discrimination.53 ThePotentialApplicationofProductsLiabilityLaw.54 ExacerbatingMarketPowerandConcentration.57 Recommendations.60 AppendixofHarms.64 References68 Introduction OpenAl'sdecisiontoreleaseChatGPT,achatbotbuiltontheLargeLanguageModelGPT-3,lastNovemberthrustAltoolstotheforefrontofpublicconsciousness.Inthelastsixmonths,newAltoolsusedtogenerate text,images,video,andaudiobasedonuserpromptsexplodedin popularity.Suddenly,phraseslikeStableDiffusion,Hallucinations,andValue capabilitiesofgenerativeAl-andtheirpotentialforharm-emergedwithoutanyclearindicationofwhatwouldcomenextorwhatimpactsthesetools wouldhave. WhilegenerativeAlmaybenew,itsharmsarenot.AlscholarshavebeenwarningusoftheproblemsthatlargeAlmodelscancauseforyears.3Theseoldproblemsareexacerbatedbytheindustry'sshiftingoalsfromresearch andtransparencytoprofit,opacity,andconcentrationofpower.The widespreadavailabilityandhypeofthesetoolshasledtoincreasedharm bothindividuallyandonamassivescale.Alreplicatesracial,gender,anddisabilitydiscrimination,andtheseharmsareweavedinextricablythrougheveryissuehighlightedinthisreport. OpenAlandothercompanies'decisionstorapidlyintegrategenerativeAltechnologyintoconsumer-facingproductsandserviceshaveundermined longstandingeffortstomakeAldevelopmenttransparentandaccountable, leavingmanyregulatorsscramblingtopreparefortherepercussions.Andit isclearthatgenerativeAlsystemscansignificantlyamplifyriskstoboth individualprivacyandtodemocracyandcybersecuritygenerally.Inthe wordsoftheOpenAlCEO,whoindeedhadthepowernottoacceleratethereleaseofthistechnology,"I'mespeciallyconcernedthatthesemodelscouldbeusedforwidespreadmisinformation...[and]offensivecyberattacks." EPICIGeneratingHarm:GenerativeAl'sImpactandPathsForward Introduction ThisrapiddeploymentofgenerativeAlsystemswithoutadequatesafeguardsisclearevidencethatself-regulationhasfailed.Hundredsofentities,fromcorporationstomediaandgovernmententities,are developingandlookingtorapidlyintegratetheseuntestedAltoolsintoa widerangeofsystems.Andthisrapidrolloutwillhavedisastrousresults builtinfromthebeginning. Weareatacriticaljunctureaspolicymakersandindustryaroundtheglobe arefocusingonthesubstantialrisksandopportunitiesposedbyAl.Thereis anopportunitytomakethistechnologyworkforpeople.Companiesshouldberequiredtoshowtheirwork,makeitclearwhenAlisinuse,andoffer informedconsentthroughoutthetraining,development,anduseprocess. OnethreadofpublicconcernfocusesonAl's"existential"risks-speculativelong-termrisksinwhichrobotsreplacehumansatwork,socially,and ultimatelytakingover,ala“I,Robot."SomelegislatorsonthestateandfederallevelhavebeguntotaketheissueofaddressingAlmoreseriously- however,itremainstobeseeniftheirfocuswillbeonlyonsupportingcompanieswiththeirdevelopmentofAltoolsandreguiringmarginaldisclosureandtransparencyrequirements.Enactingclearprohibitionsonhigh-riskuses,addressingtheeasyspreadofdisinformation,requiring meaningfulandproactivedisclosuresthatfacilitateinformedconsent,and broadoutlineofdifferentissuesthattheuseofgenerativeAlbringsup, educatelawmakersandthepublic,andoffersomepathsforwardtomi