世代人工智能的批评者正在担心错误的知识产权问题（英）-2023.3.pdf

世代人工智能的批评者正在担心错误的知识产权问题（英）-2023.3.pdf

《世代人工智能的批评者正在担心错误的知识产权问题（英）-2023.3.pdf》由会员分享，可在线阅读，更多相关《世代人工智能的批评者正在担心错误的知识产权问题（英）-2023.3.pdf（15页珍藏版）》请在冰点文库上搜索。

CENTERFORDATAINNOVATION1CriticsofGenerativeAIAreWorryingAbouttheWrongIPIssuesByDanielCastro|March20,2023CriticsarguedevelopersofgenerativeAIsystemssuchasChatGPTandDALL-Ehaveunfairlytrainedtheirmodelsoncopyrightedworks.Thoseconcernsaremisguided.Moreover,restrictingAIsystemsfromtrainingonlegallyaccesseddatawouldsignificantlycurtailthedevelopmentandadoptionofgenerativeAIacrossmanysectors.PolicymakersshouldfocusonstrengtheningotherIPrightstoprotectcreators.Oneofthemostvisibleadvancementsinartificialintelligence（AI）isthedevelopmentofgenerativeAIAIsystemsthatcanproducenovelimages,music,ortextinresponsetouserprompts.Usersarestillexploringpotentialapplicationsofthistechnologyinmanyfields,butearlyresultsarepromising.AlreadypeoplehaveusedgenerativeAItoolstodraftnewsarticles,pressreleases,andsocialmediaposts,createhigh-qualityimages,video,andmusic,andevenwritecode.Andmanymoreapplicationsinfieldssuchasmedicine,entertainment,andeducationareonthehorizon.However,somecriticsarguethatgenerativeAIposesaseriousthreattocontentcreators.Forexample,somevisualartistshavelaunchedonlineprotestsdenouncingAIandcallingforonlineplatformstoblockAI-generatedart.1OneoftheirchiefcomplaintsisthatwhendeveloperstraingenerativeAIsystemsonpubliclyaccessiblecopyrightedcontent,theyareunfairlyexploitingtheworksofcreators.2Butthesecriticsarewrong.GenerativeAIsystemsshouldnotbeexemptfromcomplyingwithintellectualproperty（IP）laws,butneithershouldtheybeheldtoahigherstandardthanhumancreators.3ThisreportrefutesfiveofthemostcommonargumentsmadeabouthowgenerativeAIisunfairtocreators:

1.TraininggenerativeAIsystemsoncopyrightedcontentistheft.CENTERFORDATAINNOVATION22.GenerativeAIsystemsshouldnottrainoncontentwithoutthecopyrightownersexplicitpermission.3.GenerativeAIsystemsshouldcompensatecopyrightownersfortrainingontheircontent.4.GenerativeAIsystemsshouldnotbeallowedtoproducecontentbasedonthestyleofanartistwithouttheirpermission.5.GenerativeAIsystemsusefragmentsofcopyrightedcontentintheiroutputs.ThereportalsoacknowledgesthattherearelegitimateIPrightsatstake.Specifically,itidentifiesfiveharmfulactivities:

1.InfringingoncopyrightsofAI-generatedworks2.Distributingcopyrightedcontent3.Creatingforgeries4.Creatinginfringingcontent5.ImpersonatingindividualsFinally,thereportdiscussestheimpactofgenerativeAIonthoseharmfulactivitiesandrecommendspolicymakersaddressconcernsthroughrobustenforcementofexistingrights,offeringguidanceandclaritytothoseusingthesetools,newlegislationtocombatonlinepiracy,andexpandingprohibitionsonthedistributionofnonconsensualintimateimages（sometimesreferredtoas“revengeporn”）toinclude“deepfakes”（i.e.,imagesandvideocreatedusinggenerativeAI）.CRITICSAREWRONGTHATGENERATIVEAIISUNFAIRTOCREATORSThelistofgenerativeAIapplicationscontinuestogrow（seetable1）.Asitdoes,AI-generatedcontenthascreatedalotofpraiseandcontroversy.4ManywelcometheadventofgenerativeAI,seeingitasanotherpowerfultechnologylikesoftware-basedwordprocessorsandvideoeditorsthatwillempowercreatorstobetterexpressthemselves.OthersfearthatAIwilldevalueartisticworks,andartiststhemselves,byreplacingvibranthumancreativitywithacold,emotionlessalgorithm.5Therealityismorelikelytobesomewhereinthemiddle.WhileAI-generatedcontentwilllikelyserveasausefulsubstituteforcertainpurposessimplemarketingcopy,stockimages,androyalty-freemusicitmayholdlessappealforcollectorsoffineart,musicconnoisseurs,andliteraryaficionados.Indeed,artpriceshavehistoricallyoperateddifferentlythanothergoods.Thepriceoffineartisnotlinkedtoproductioncosts,butinsteadtoabstractandsubjectivequalitiessuchastheperceivedqualityofCENTERFORDATAINNOVATION3thework,thereputationoftheartist,andtheopinionsofgalleryowners,collectors,andotherauthorities.6Likewise,somepuristswhethertheybewritersormusiciansasabadgeofhonor,willlikelyrefusetousethetechnologytocreatecontent,whilemanyotherswilladoptthetechnologybecauseitmakestheirjobseasier.Indeed,ifAImakescreatorsmoreproductive,itwillgrowtheeconomy.Table1:

ListofpopulargenerativeAIapplicationsTypeExamplesImageDALL-E2,Midjourney,StableDiffusionTextChatGPT,Bing,BardMusicJukebox,MusicLMCodeCopilotButthebiggestdebateisaboutwhethergenerativeAIsystemsshouldbeallowedtotraintheirmodelsontext,audio,images,andvideosthatarelegallyaccessibletoInternetusersbutarealsoprotectedbycopyright.SomecreatorsarguethatitisunfairfordeveloperstotraintheirAIsystemsoncontenttheyhavepostedontheInternetwithouttheirconsent,credit,orcompensation.Theiroppositionisgenerallywithoutmerit,buttheirargumentsareworthconsidering.Beforegettingintothesearguments,itisimportanttonotethatpeopledonothavetherighttousecopyrightedcontentanywaytheywantjustbecausetheycanlegallyaccessitontheInternet.However,theirnothavingtherighttouseitanywaytheywantdoesnotmeantheycannotdoanythingwiththiscontent.Copyrightlawprovidescopyrightownerscertainexclusiverights,buttheserightsaresubjecttoexceptionsandlimitations,includingthoseunderthefairusedoctrine.Forexample,searchenginescanlegallycrawlwebsiteswithoutviolatingcopyrightlaws.WhileitwillultimatelybeuptothecourtstodecidewhetheraparticularuseofgenerativeAIinfringesoncopyright,thereisprecedentforthemtofindmostusestobelawfulandnotinviolationofrightsholdersexclusiverights.71.IsTrainingGenerativeAISystemsonCopyrightedContentTheft?

SomearguethattrainingAIsystemsoncopyrightedcontentistheftplainandsimple.8Indeed,stealingdigitalcontentisaseriousproblemonlinepiracyofmovies,TV,music,games,andmorecostscreatorsbillionsinlostrevenueannually.9Butonlinepiracyisclearlytheft.ThereislittledifferencebetweensomeonewatchingaDVDtheyhaveshopliftedfromaWalmartorBestBuycomparedwithsomeonewatchingavideotheyhaveCENTERFORDATAINNOVATION4streamedillegallyonlineinbothcases,theyarewatchingavideotheyhavenotpaidfor.Butseekinginspirationandlearningfromothersisnottheft.Itisnottheftifsomeonewatchesavideolegally,andthatvideoinspiresthemtofilmtheirownuniquecreation.Indeed,TikTokandothersocialmediaplatformsarefilledwithsuchvideosinspiredbyrelatedcontent.Similarly,writers,musicians,andotherartistslearntheircraftbyobservingpastcreations.Infact,allcreativeworksareshapedbypastworks,ascreatorsdonotexistinavacuum.Theinspections,impressions,andinspirationsoftheworldaroundthemarewhatgiverisetonewideas.Callingthisprocesstheftisclearlyinaccuratewhenappliedtothewayhumansobserveandlearn,anditisequallyinaccuratetodescribetrainingagenerativeAIsystem（discussedinmoredetailahead）.2.ShouldGenerativeAISystemsBeAllowedtoTrainonContentWithouttheCopyrightOwnersExplicitPermission?

SomearguethatitiswrongtotrainAIsystemsoncopyrightedcontentwithoutfirstobtainingaffirmativeconsentfromthecopyrightholder.ThesecriticssaythateveniftrainingAIsystemsdoesnotamounttotheft,copyrightownersshouldstillhavetherighttodeterminehowothersusetheirworks,sincetheycreatedit.Andindeed,thelawdoesconfercertainrightstocopyrightowners,suchastherighttoreproduceawork,therighttopreparederivativeworks,therighttoperformaworkpublicly,andtherighttodisplayaworkpublicly.10However,thisargumentfallsapartuponcloserexamination.Asnoted,copyrightownershavetherighttodecidewhethertodisplayorperformtheirworkspublicly.Butiftheychoosetodisplaytheirworkinpublic,otherscanusetheirworksincertainwayswithouttheirpermission.Forexample,photographerscantakepicturesofsculpturesorgraffitiinpublicplacesevenwhenthoseworksareprotectedbycopyright.Copyrightpreventsphotographersfromsellingthoseimages,butitdoesnotrequirethemtogetpermissionfromthecopyrightownertotakephotos.Likewise,individualsdonotneedtogetpermissionfromacopyrightownertostudyapaintingtheyseeinagalleryorasongtheyhearontheradio.Peoplearefreetoobservetheseworksandusewhattheylearnfromthemtocreatefuturecontentwithouttheexplicitpermissionofthecopyrightowners.ThereisnointrinsicrationaleforwhyusersofgenerativeAIsystemswouldneedtoobtainpermissiontotrainoncopyrightedcontenttheyhavelegalaccessto.MusiciansmightpracticeacopyrightedsongtheyheardonSpotifyhundredsoftimestolearntoplayaninstrumentorusetheirwell-honedauditorymemorytorecallelementsofpiecestheyhaveheardbefore.Learningfromlegallyaccessedworksdoesnotviolateacopyrightownersexclusivereproductionanddistributionrights.Unlesshumancreatorswillberequiredtoobtainpermissionbeforetheycanstudyanotherpersonswork,thisrequirementshouldnotbeappliedtoAI.CENTERFORDATAINNOVATION53.ShouldGenerativeAISystemsCompensateCopyrightOwnersforTrainingonTheirContent?

SomearguethatpeopleshouldpaycopyrightownerstotraingenerativeAIsystemsontheircontentbecausetheyareobtainingvaluefromthisprocess.Butcopyrightownersdonothavethissameexpectationwhenotherhumancreatorslearnfromtheirworks.Buddingmusicianslistentohoursofmusic,youngwriterscloselystudytheirfavoritenovels,andamateurpaintersspendhourslookingat