生成式人工智能与版权法 Generative Artificial Intelligence and Copyright Law_第1页
已阅读1页,还剩12页未读 继续免费阅读

下载本文档

版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领

文档简介

LegalSidebari

GenerativeArtificialIntelligenceandCopyrightLaw

UpdatedJune16,2025

Innovationsinartificialintelligence(AI)haveraisedseveralnewquestionsinthefieldof

copyrightlaw

.

GenerativeAI

programs—suchasOpenAI’s

DALL-E

and

ChatGPT

programs,StabilityAI’s

Stable

Diffusion

program,andMidjourney’s

self-titledprogram

—areabletogeneratenewimages,texts,and

othercontent(or“outputs”)inresponsetoauser’stextualorotherprompts.GenerativeAIprogramsaretrainedtocreatesuchoutputspartlybyexposingthemtolargequantitiesofexistingwritings,photos,

paintings,orotherworks.

ThisLegalSidebarexploresquestionsthatcourtsandtheU.S.CopyrightOfficehaveconfronted

regardingwhethergenerativeAIoutputsmaybecopyrighted,aswellaslegaldebatesaboutwhether

trainingandusinggenerativeAIprogramsmayinfringecopyrightsinotherworks.OtherLegalSidebarsexplorequestionsAIraisesintheintellectualpropertyfieldsof

patents

and

therightofpublicity

.

CopyrightinWorksCreatedwithGenerativeAI

DoCopyrightedWorksRequireaHumanAuthor?

ThequestionofwhethercopyrightprotectionmaybeaffordedtoAIoutputs—suchasimagescreatedbyMidjourneyortextscreatedbyChatGPT—hingeslargelyonthelegalconceptof“authorship.”ArticleI,Section8,Clause8oftheU.S.Constitution,oftenreferredtoastheIntellectualProperty(IP)Clause,

empowers

Congressto“secur[e]forlimitedTimestoAuthors...theexclusiveRighttotheir...

Writings.”Basedonthisauthority,the

CopyrightAct

affordscopyrightprotectionto“originalworksofauthorship.”WhiletheConstitutionandCopyrightActdonotexplicitlydefinewho(orwhat)maybean“author,”U.S.courtstodatehavenotrecognizedcopyrightinworksthatlackahumanauthor—includingworkscreatedautonomouslybyAIsystems.

BeforetheproliferationofgenerativeAI,U.S.courtsdidnotextendcopyrightprotectiontovarious

nonhumanauthors,holdingthat

amonkey

whotookphotosofhimselflackedstandingtosueunderthe

CopyrightAct;thathumanauthorshipwasrequiredtocopyrightabookpurportedlyinspiredby

celestial

beings;

andthata

livinggarden

couldnotbecopyrighted.TheU.S.CopyrightOfficehasalsolong

maintained

thatcopyrightedworksmustbe“createdbyahumanbeing”andtherefore

refused

toregister

CongressionalResearchService

LSB10922

CRSLegalSidebar

PreparedforMembersandCommitteesofCongress

CongressionalResearchService2

worksthatare“producedbyamachineormeremechanicalprocessthatoperatesrandomlyorautomaticallywithoutanycreativeinputorinterventionfromahumanauthor.”

Atleastonelawsuithas—unsuccessfully,thusfar—challengedthehuman-authorshiprequirementinthecontextofAI.InJune2022,StephenThaler

sued

theCopyrightOfficefordenyinghisapplicationto

registeravisualartworkthathe

claims

wasauthored“autonomously”byanAIprogram.Dr.Thaler

arguedthathumanauthorshipisnotrequiredbytheCopyrightAct.InAugust2023,aU.S.districtcourt

granted

summaryjudgmentinfavoroftheCopyrightOffice.Thecourtheldthat“humanauthorshipisanessentialpartofavalidcopyrightclaim,”reasoningthatonlyhumanauthorsneedcopyrightasan

incentivetocreateexpressiveworks.

InMarch2025,theU.S.CourtofAppealsfortheD.C.Circuit

affirmed

thedistrictcourt’sdecisionin

Thalerv.Perlmutter,

holding

thattheCopyrightAct“requiresalleligibleworktobeauthoredinthefirstinstancebyahumanbeing.”Thecourt

reasoned

thatseveralprovisionsoftheCopyrightActimplythatitusestheword“author”onlytorefertohumanbeings,includingprovisions(1)vestingcopyright

ownership

“initiallyintheauthor”;(2)limitingcopyright

duration

to70yearsafter“theauthor’sdeath”;(3)providingfor

inheritance

ofcertainrightsbytheauthor’s“widoworwidower”or“survivingchildrenorgrandchildren”;(4)requiringa

signature

totransfercopyrightownership;(5)extendingprotectionto

unpublishedworks

regardlessoftheauthor’s“nationalityordomicile”;and(6)defininga“

jointwork

basedontheauthors’“intention”tomergetheircontributionsinacertainway.Inaddition,thecourt

observed

thattheCopyrightOfficehadadoptedthehuman-authorshiprequirementyearsbeforeCongressenactedthecurrentCopyrightAct.Thecourtthus

inferred

thatCongressmeanttoadoptthehuman-

authorshiprequirementwhenitenactedthelaw.BasedonitsholdingthattheCopyrightActrequires

humanauthorship,thecourtfounditunnecessarytoevaluatetheCopyrightOffice’s

argument

thattheConstitution’sIPClauserequireshumanauthorshipforcopyrightedworks.OnMay12,2025,thecourt

denied

Dr.Thaler’spetitiontorehearthecaseenbanc(i.e.,byallofthecourt’sjudges).

MayHumansCopyrightWorksThatTheyCreateUsingAI?

Assumingthatcopyrightableworksrequireahumanauthor,workscreatedbyhumanswiththeassistanceofgenerativeAImightbeentitledtocopyrightprotectiondependingonthenatureofhumaninvolvementinthecreativeprocess.Asdiscussedbelow,theCopyrightOfficehassoughttodelineatewhatauthors

mustdotosatisfythehuman-authorshiprequirementwhenusinggenerativeAI.

InMarch2023,theCopyrightOfficereleased

CopyrightRegistrationGuidance

regarding“works

containingmaterialgeneratedby[AI]”(theAIGuidance).AllowingthathumanauthorsmayuseAIinthecreativeprocess,theAIGuidance

states

that“whatmattersistheextenttowhichthehumanhadcreativecontroloverthework’sexpression.”Thus,theAIGuidance

states

,whenAI“determinestheexpressive

elementsofitsoutput,thegeneratedmaterialisnottheproductofhumanauthorship.”Ontheotherhand,works

containing

AI-generatedmaterialmaybecopyrightedundersomecircumstances,suchas

“sufficientlycreative”humanarrangementsormodificationsofAI-generatedmaterialorworksthatcombineAI-generatedandhuman-authoredmaterial.TheAIGuidance

states

thatauthorsmayclaimcopyrightprotectiononly“fortheirowncontributions”tosuchworks,andtheymustidentifyand

disclaimAI-generatedpartsoftheworkswhenapplyingtoregistertheircopyright.

CongressionalResearchService3

Threecopyrightregistrationdenials

highlighted

bytheCopyrightOfficeillustratethat,ingeneral,the

officewillnotfindhumanauthorshipwhereanAIprogramgeneratesworksinresponsetouserprompts:

ZaryaoftheDawn:

AFebruary2023decisionthatAI-generatedillustrationsforagraphicnovelwerenotcopyrightable,althoughthehuman-authoredtextofthenovelandoverallselectionandarrangementoftheimagesandtextinthenovelcouldbecopyrighted.

ThéâtreD,opéraSpatial:

ASeptember2023decisionthatanartworkgeneratedbyAIandthenmodifiedbytheapplicantcouldnotbecopyrighted,sincetheapplicantfailedto

identifyanddisclaimtheAI-generatedportionsoftheworkasrequiredbytheAIGuidance.

SURYAST:

ADecember2023decisionthatanartworkgeneratedbyanAIsystem

combininga“baseimage”(anoriginalphototakenbytheapplicant)anda“styleimage”theapplicantselected(VincentvanGogh’sTheStarryNight)couldnotbecopyrighted,sincetheAIsystemwas“responsiblefordetermininghowtointerpolate[i.e.,combine]thebaseandstyleimages.”

WhiletheCopyrightOffice’sdecisionsindicatethatitmaynotbepossibletoobtaincopyrightprotectionformanyAI-generatedworks,theissueremainsunsettled.AnapplicantmayfilesuitinU.S.districtcourtto

challenge

theCopyrightOffice’sfinaldecisiontorefusetoregisteracopyright.TheputativeauthorofThéâtreD,opéraSpatial,forinstance,has

sued

theCopyrightOfficefordecliningtoregisterthatwork.

WhiletheCopyrightOffice

notes

thatcourtssometimesgiveweighttotheoffice’sexperienceandexpertise,courtsarenotboundtoadopttheoffice’sinterpretationsoftheCopyrightAct,suchasitsapplicationoftheauthorshiprequirementtoAI-assistedworks.

InJanuary2025,theCopyrightOfficepublishedthe

Copyrightability

partofitsCopyrightandArtificialIntelligencereport.SimilartotheAIGuidance’semphasison“creativecontrol,”thereport

concludes

that,“givencurrentgenerallyavailabletechnology,promptsalonedonotprovidesufficienthumancontrolto

makeusersofanAIsystemtheauthorsoftheoutput.”Thereport

contends

thattheCopyrightAct’s

distinctionbetweencopyrightable“works”andnoncopyrightable“

ideas

”precludescopyrightabilityforworksgeneratedbyAIinresponsetouserprompts.Asthereport

argues

,“[p]romptsessentiallyfunctionasinstructionsthatconveyunprotectibleideas”and“donotcontrolhowtheAIsystemprocessesthemingeneratingtheoutput.”

Somecommentators

assertthatcertainAI-generatedworksshouldreceivecopyrightprotection,

comparingAIprogramstoothertoolshumanauthorshaveusedtocreatecopyrightedworks.For

example,theU.S.SupremeCourtheldinthe1884case

Burrow-GilesLithographicCo.v.Sarony

that

photographscanbeentitledtocopyrightprotectionwherethephotographermakesdecisionsregarding

creativeelementssuchascomposition,arrangement,andlighting.Somecopyrightapplicants

argue

thatgenerativeAIprogramsmayfunctionastools,analogoustocameras.TheCopyrightOfficedisputesthe

photographyanalogy,arguingthatAIusersdonotexercisesufficientcontroltocharacterizegenerativeAIasatoolusedbyanauthor.Instead,theCopyrightOfficehas

compared

anAIuserto“aclientwhohiresanartist”andgivesthatartistonly“generaldirections.”Atleastoneapplicant’sattorneyhas

argued

thattheCopyrightActdoesnotrequiresuchexactingcreativecontrol,observingthatcertainphotographsandmodernartincorporateadegreeofhappenstance.

Regardingworksthatcontainacombinationofhuman-authoredandAI-generatedmaterial,theCopyrightOffice

reports

thatit“hasregisteredhundredsofworksthatincorporateAI-generatedmaterial,withtheregistrationcoveringthehumanauthor’scontributiontothework,”inthetimesinceitissuedtheAI

Guidance.Theoffice

contends

thatnewlegislationregarding“thecopyrightabilityofAI-generated

material”iscurrentlynotneeded,indicatingthatcourts“willprovidefurtherguidanceonthehuman

authorshiprequirementasitappliestospecificusesofAI”andthat,sinceeachworkmustbeanalyzedindividually,“greaterclaritywouldbedifficulttoachieve”throughlegislation.

CongressionalResearchService4

CopyrightInfringementbyGenerativeAIPrograms

DoestheAITrainingProcessInfringeCopyrightsinOtherWorks?

AIsystemsare

trained

tocreateliterary,visual,andotherartisticworksbyexposingthesesystemsto

largeamountsofdata,whichmayincludetext,images,andotherworksdownloadedfromtheinternetorotherwiseobtainedbyAIcompanies.Thistrainingprocessofteninvolvesmakingdigitalcopiesof

existingworks.AstheU.S.PatentandTrademarkOfficehas

described

,theprocess“willalmostby

definitioninvolvethereproductionofentireworksorsubstantialportionsthereof.”OpenAI,forexample,

acknowledged

thatitsprogramsaretrainedon“large,publiclyavailabledatasetsthatincludecopyrightedworks”andthatthisprocess“involvesfirstmakingcopiesofthedatatobeanalyzed”(althoughitnow

offersanoption

toremoveimagesfromtrainingfutureAImodels).

SomecopyrightownersandcommentatorshaveassertedthatcreatingdigitalcopiesofworkswithoutpermissiontotrainAIinfringestheowners’

exclusiveright

tomakereproductionsoftheirwork.

Copyrightownershavefiled

severaldozenlawsuits

againstAIcompaniesmakingsomeversionofthisclaim.

Incontrast,anumberofAIcompaniesandsomelegal

scholars

arguethatAItrainingprocessesconstitutefairuseandarethereforenoninfringing.Whetherornotcopyingconstitutesfairusedependson

four

nonexclusivefactors

thatCongresssetforthintheCopyrightAct:

1.thepurposeandcharacteroftheuse,includingwhethersuchuseisofacommercialnatureorisfornonprofiteducationalpurposes;

2.thenatureofthecopyrightedwork;

3.theamountandsubstantialityoftheportionusedinrelationtothecopyrightedworkasawhole;and

4.theeffectoftheuseuponthepotentialmarketfororvalueofthecopyrightedwork.

AstheSupremeCourthas

stated

,fairuseisa“flexible”doctrine,and“itsapplicationmaywellvary

dependinguponcontext.”Astothefirstfactor,theCourthas

held

thatsomeuseswitha“transformative”purpose(suchasparodies)maybefair,althoughin2023it

cautioned

thattransformativenessis“amatterofdegree.”TheCourthas

described

thefourthfactorasthe“mostimportant”one.

Regardingthefirstfactor,OpenAI

hasargued

itspurposeis“transformative”becausethetrainingprocesscreates“ausefulgenerativeAIsystem”thatdidnotpreviouslyexist.Forsupport,OpenAIcites

The

AuthorsGuild,Inc.v.Google,Inc.

,inwhichtheU.S.CourtofAppealsfortheSecondCircuitheldthat

Google’scopyingofentirebookstocreateasearchabledatabasethatdisplayedexcerptsofthosebooksconstitutedfairuse.Ontheotherhand,somestakeholdershaveobservedthatmanygenerativeAI

programshaveacommercialpurposeand

claim

thatusingcopyrightedworkstotrainAItocreatesimilarkindsofworksisnot“transformative.”

Regardingthefourthfair-usefactor,stakeholdersandcommentatorsdisputewhetherAIoutputsarelikelytocompetewithorharmthemarketforcopyrightedworksusedintheirtrainingdataand,ifso,what

kindsofcompetitionarerelevanttothefair-useanalysis.Some

argue

thatthefourthfactorweighsagainstfairusetotheextentthat“outputsthatmimicorareotherwisebasedontheingestedworksundermine

marketdemandforthoseworks.”Others

contend

thatcompetitionfromAIoutputsshouldweighagainstfairuseonlywherethoseoutputsreproduce“thecopyrightowner’soriginalexpression.”Insomecases,AIcompanieshavevoluntarilyadoptedmeasuresthatmayarguablymitigateconcernsaboutharmingthemarketforworksusedtotraintheAI.Forinstance,OpenAI

states

thatDALL-E3“isdesignedtodeclinerequeststhataskforanimageinthestyleofalivingartist.”

CongressionalResearchService5

InFebruary2025,aU.S.districtcourt

ruled

thatitwasnotfairuseforacompany(RossIntelligence)tocopycasesummariesfromWestlaw,alegalresearchplatform,totrainanAIprogramtoquotepassagesfromlegalopinionsinresponsetouserquestions.Thecourtconcludedthatthe

firstfactor

weighed

againstfairuse,sincethecopyinghadacommercialpurpose.Further,sinceRoss’sAIprogramand

Westlawhadthesamepurposeofassistinglegalresearch,thecourtfoundthatthecopyingwasnot

sufficiently“

transformative

”tosupportfairuse.Thecourtruledthatthe

fourthfactor

alsoweighed

againstfairuse,asRossmeanttocompetewithWestlawbycreatingasubstitutelegalresearchplatformwhilepotentiallyunderminingWestlaw’sabilitytolicenseitscasesummariestotrainAIsystems.The

courtconcludedthatthe

secondandthirdfactors

supportedfairuse—sincetheWestlawcasesummariesshowedonly“minimal”creativity(factor2)andRoss’sproductdidnotmakethosesummariesavailabletothepublic(factor3)—butthatthesefactorswereoutweighedbytheothers.Theimportanceofthis

decisiontoongoinglitigationregardinggenerativeAIprogramsisdebatable,sincefairuseisafact-specificanalysisand,asthecourt

observed

,theRossAItechnologywas“non-generativeAI.”

InMay2025,theCopyrightOfficereleasedaprepublicationversionofthe

GenerativeAITraining

partofitsCopyrightandArtificialIntelligencereport.Basedonitsanalysisofthefourfair-usefactors,thereportsection

concluded

that“itisnotpossibletoprejudgelitigationoutcomes,”anticipatingthat“someusesofcopyrightedworksforgenerativeAItrainingwillqualifyasfairuse,andsomewillnot.”

DoAIOutputsInfringeCopyrightsinOtherWorks?

SomeoutputsofAIprogramsmightinfringecopyrightsinotherworkstheyresemblethatwereusedto

traintheAI.Copyrightownersmaybeabletoestablishthatsuchoutputsinfringetheircopyrightsifthe

AIprogramboth(1)hadaccesstotheirworksand(2)created“

substantiallysimilar

”outputs.First,to

establishcopyrightinfringement,aplaintiffmustprovetheinfringer“

actuallycopied

”theunderlying

work.Thiselementissometimesprovencircumstantiallybyevidencethattheinfringer“

hadaccesstothe

work

.”ForAIoutputs,accessmightbeshownbyevidencethattheAIprogramwastrainedusingthe

underlyingwork.Suchevidencemightshow,forinstance,thatacopyoftheunderlyingworkwaslocatedonaninternetsitethatwasdownloadedor“scraped”totraintheAIprogram.

Second,aplaintiffmustprovethatthenewworkis“

substantiallysimilar

”totheunderlyingworkto

establishinfringement.Thesubstantialsimilaritytestisdifficulttodefine.Courtshavevariously

describedthetestasrequiring,forexample,thattheworkshave“asubstantiallysimilar

totalconceptand

feel

”or“

overalllookandfeel

”orthat“

theordinaryreasonablepersonwouldfailtodifferentiatebetween

thetwoworks

.”Leadingcaseshavealsostatedthatthisdeterminationconsidersboth“

thequalitativeand

quantitativesignificanceofthecopiedportioninrelationtotheplaintiff’sworkasawhole

.”ForAI-

generatedoutputs,nolessthanfortraditionalworks,the“substantialsimilarity”analysismayrequirecourtstomakethesekindsofcomparisonsbetweentheAIoutputandtheunderlyingwork.

OpenAI

hasargued

that“[w]ell-constructedAIsystemsgenerallydonotregenerate,inanynontrivialportion,unaltereddatafromanyparticularworkintheirtrainingcorpus.”Thus,

according

toOpenAI,copyright-infringingoutputswouldbe“anunlikelyaccidentaloutcome”ofsuchsystems.

Onestudy

found“asignificantamountofcopying”inlessthan2%oftheimagescreatedbyStableDiffusion,

thoughtheauthorsclaimedthattheirmethodology“likelyunderestimatesthetruerate”ofcopying.

TwokindsofAIoutputsmayraisespecialconcerns.First,someAIprogramsmaybeusedtocreateworksinvolvingexistingfictionalcharacters.Theseworksmayrunaheightenedriskofinfringement,since

characters

sometimesenjoycopyrightprotection

distinctfromthespecificworksinwhichtheyappear.

Second,someAIprogramsmaybepromptedtocreateworks“inthestyleof”aparticularartistorauthor,although—asnotedabove—someAIprogramsmaynowbedesignedto“

decline

”suchprompts.Theseoutputsarenotnecessarilyinfringing,ascopyrightlawgenerallyprotectsonlyagainstthecopyingof

specificworksratherthananartist’soverallstyle.Forexample,asonggeneratedbyAIinthestyleand

CongressionalResearchService6

simulatedvoiceofahumanperformer

mightnotinfringe

anycopyright,althoughvoicesimulationsmaypotentiallyviolatesome

stateright-of-publicitylaws

.Asaseparateissuefromwhethertheoutputitselfisinfringing,theCopyrightOffice

contends

thattheuseofAItocreateoutputsinthestyleofanauthor

couldpotentiallyweighagainstafair-usedefenseforcopyingtheauthor’sworktotraintheAI,assuchoutputsmayreducedemandfortheauthor’sworkvia“

marketdilution

.”Asnotedabove,somelegal

scholars

disagree

thatthecreationofnoninfringingoutputsshouldweighagainstfairusefortrainingAI.

IfagenerativeAIoutputinfringesacopyrightinanexistingwork,boththeAIuserandtheAIcompanycouldpotentiallybeliableundercurrentlaw.Forinstance,theusermightbedirectlyliableforpromptingtheAIprogramtogenerateaninfringingoutput.Itmaybechallengingtoanalyzetheuser’sliabilityin

somecases,sincetheusermightnothavedirectaccessto—orevenbeawareof—acopyrightedworkpurportedlyinfringedbyanAIoutput.TheAIcompanycouldalsopotentiallyfaceliabilityunderthedoctrineof“vicariousinfringement.”Vicariousinfringement

applies

todefendantswhohave“therightandabilitytosupervisetheinfringingactivity”and“adirectfinancialinterestinsuchactivities.”

ConsiderationsforCongress

CongressmayconsiderwhethertoaddressanyofthecopyrightlawquestionsraisedbygenerativeAI

programsthroughamendmentstotheCopyrightActorotherlegislation.Congressmay,forexample,

considerlegislationclarifyingwhetherAI-generatedworksarecopyrightableorunderwhatcircumstancestheprocessoftraininggenerativeAIprogramsmayconstitutefairuse.Giventhelimitedtimecourtshavehadtoaddresstheseissues,Congressmayalternativelyadoptawait-and-seeapproach.Asthecourts

decidecasesinvolvinggenerativeAI,theymaybeabletoprovidegreaterguidanceandpredictabilityinthisarea.Basedontheoutcomesofthesecases,Congressmayreassesswhetherlegislationisneeded.

CongressmayalsoconsiderthepracticalimplicationsofrequiringAIcompaniestoidentify,seek

permissionfrom,orcompensatecopyrightownersshouldcourtdecisionsorfuturelegislationdeterminethattraininggenerativeAIsystemsisnotafairuseofcopyrightedworks.Commentatorshavedebatedwhetheritisfeasibletorequirecompaniestoidentifyandpayownersofthelargenumberofworks

neededtotrainAIsystems,aswellaswhetherthevalueofsuchcompensationtoownerswouldbe

outweighedbytransactionoradministrationcosts.Onescholar,

acknowledging

that“[i]twould...beimpossibleforanAIdevelopertoidentifyandclearbillionsofrightsclaimsonanindividualbasis,”

arguesthatitmaybefeasibleinsteadtocreatemarketsforAItrainingdataviameanssuchascontent

aggregation(e.g.,TVstreamingservices),collectivemanagementorganizations(orCMOs,suchasthosethatmanagerightsto

musicalworks)

,compulsory

温馨提示

  • 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
  • 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
  • 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
  • 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
  • 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
  • 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
  • 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。

评论

0/150

提交评论