2023生成式人工智能的潜在有害影响与未来之路（英文版）

epic.orgELECTRONICPRIVACYINFORMATIONCENTERGENERATNGHARMSGenerativeAl's Impact &PathsForwardMAY2023 CONTRIBUTIONSBYGrantFergussonTomMcBrienCaitrionaFitzgeraldCalliSchroederChris FrascellaBen WintersMegan lorioEnid ZhouEDITED BYGrantFergusson,Calli Schroeder,BenWinters,andEnidZhouThank youto Sarah Myers West and Katharina Koppfor your generouscomments onanearlier draftof the paper.Notes on this Paper:This is version1of thispaperand isreflectiveof documented andanticipated harms of Generative Al as of May 15,2023.Due to the fast-changing pace of development, use, and harms of Generative Al, we wanttoacknowledgethatthisisaninherentlydynamicpaper,subjecttochanges inthe future.Throughout this paper, we use a standard format to explain the typology ofharmsthatgenerativeAlcanproduce.Eachsectionfirstexplainsrelevantbackground informationand potential risks imposed bygenerativeAl,thenhighlights specifics harms and interventions that scholarsand regulatorshavepursuedtoremedyeachharm.Thispaperdrawsontwotaxonomies ofA.l.harmstoguideouranalysis:1. Danielle Citron's and Daniel Solove's Typology of Privacy Harms,comprisingphysical,economic,reputational,psychological,autonomy,discrimination,and relationshipharms;land2. Joy Buolamwini's Taxonomy of Algorithmic Harms, comprising loss ofopportunity,economicloss,andsocial stigmatization,includinglossofliberty,increased surveillance,stereotypereinforcement,andotherdignitaryharms.2These taxonomies do not necessarily cover all potential Al harms, and ouruseofthesetaxonomiesismeanttohelpreadersvisualizeandcontextualizeAlharmswithout limitingthetypesandvarietyofAlharmsthatreaders consider. TableofContentsIntroduction.Turbocharging Information Manipulation.Harassment,Impersonation,and Extortion...9Spotlight:Section230..19ProfitsOverPrivacy:IncreasedOpagueDataCollection..24IncreasingDataSecurityRisk..30ConfrontingCreativity:ImpactonIntellectualPropertyRights..33Exacerbating Effects of Climate Change....40LaborManipulation,Theft,andDisplacement...44Spotlight:Discrimination...53ThePotentialApplicationofProductsLiabilityLaw....54ExacerbatingMarketPowerandConcentration...57Recommendations..60AppendixofHarms.64References68 IntroductionOpenAl's decision to release ChatGPT, a chatbot built on the LargeLanguageModelGPT-3,lastNovemberthrustAltoolstotheforefrontofpublic consciousness.In the last six months,newAltools used to generatetext,images,video,andaudiobasedonuserpromptsexplodedinpopularity.Suddenly,phraseslikeStableDiffusion,Hallucinations,andValuecapabilities of generative Al-andtheir potentialfor harm-emerged withoutany clear indication of what would come next or what impacts thesetoolswould have.Whilegenerative Al maybe new, its harms are not.Al scholars have beenwarning us of the problems that large Al models can cause for years.3Theseoldproblemsareexacerbatedbytheindustry'sshiftingoalsfromresearchandtransparencytoprofit,opacity,andconcentrationofpower.Thewidespread availabilityand hype of these tools has led to increased harmbothindividuallyandonamassivescale.Alreplicatesracial,gender,anddisability discrimination,and these harms are weaved inextricablythrougheveryissuehighlightedinthisreport.OpenAlandothercompanies'decisionstorapidlyintegrategenerativeAltechnologyintoconsumer-facingproductsandserviceshaveunderminedlongstanding effortstomakeAldevelopmenttransparentandaccountable,leaving many regulators scrambling to prepareforthe repercussions.And itisclearthatgenerativeAlsystemscansignificantlyamplifyriskstobothindividual privacyandtodemocracyand cybersecuritygenerally.InthewordsoftheOpenAlCEO,whoindeedhadthepowernottoacceleratethereleaseofthistechnology,"I'm especiallyconcerned thatthesemodelscouldbe usedforwidespreadmisinformation...[and] offensive cyberattacks."EPICIGenerating Harm:GenerativeAl's Impactand PathsForward IntroductionThis rapid deployment of generativeAlsystems without adequatesafeguards is clear evidence that self-regulation has failed. Hundreds ofentities,fromcorporationsto mediaandgovernmententities,aredevelopingandlookingtorapidlyintegratetheseuntestedAltoolsintoawide range of systems.And this rapid rollout will have disastrous resultsbuilt infromthebeginning.We are at a critical juncture as policymakers and industry around the globearefocusingonthesubstantial risksandopportunitiesposedbyAl.Thereisanopportunitytomakethistechnologyworkforpeople.Companies shouldberequiredtoshowtheirwork,makeitclearwhenAlisinuse,andofferinformedconsentthroughoutthetraining,development,anduseprocess.Onethread ofpublic concernfocusesonAl's"existential"risks-speculativelong-termrisksinwhichrobotsreplacehumansatwork,socially,andultimately taking over,a la“I, Robot."Somelegislators on the stateandfederal level have begun to take the issueof addressing Almore seriously-however, it remains to be seen if their focus will be only on supportingcompanieswiththeirdevelopmentofAltoolsandreguiringmarginaldisclosureand transparencyrequirements.Enacting clearprohibitions onhigh-risk uses,addressin