世界經(jīng)濟(jì)論壇：2024年生成人工智能時(shí)代的治理報(bào)告 -彈性的360度方法政策和法規(guī)

WJECCN

FO

IncollaborationwithAccenture

GovernanceintheAge

A360oApproachPolicyand

WHITE

OCTOBER

9

11

10

12

12

18

21

21

22

25

25

26

27

28

33

Contents

Images:GettyImages,Midjourney

3

4

5

6

6

Foreword

ExecutivesummaryIntroduction

1Harnesspast

1.1ExamineexistingregulationscomplicatedbygenerativeAIattributes

1.2Resolvetensionsbetweenpolicyobjectivesofmultipleregulatoryregimes

1.3Clarifyexpectationsaroundresponsibilityallocation

1.4Evaluateexistingregulatoryauthoritycapacityforeffectiveenforcement

2Buildpresent

2.1Addresschallengesofstakeholdergroups

2.2Facilitatemultistakeholderknowledge-sharingandinterdisciplinaryefforts

3Planfuture

3.1Targetedinvestmentsandupskilling

3.2Horizonscanning

3.3Strategicforesight

3.4Impactassessmentsandagileregulations

3.5Internationalcooperation

ConclusionContributorsEndnotes

Disclaimer

Thisdocumentispublishedbythe

WorldEconomicForumasacontributiontoaproject,insightareaorinteraction.

Thefindings,interpretationsand

conclusionsexpressedhereinarearesultofacollaborativeprocessfacilitatedand

endorsedbytheWorldEconomicForumbutwhoseresultsdonotnecessarily

representtheviewsoftheWorldEconomicForum,northeentiretyofitsMembers,

Partnersorotherstakeholders.

?2024WorldEconomicForum.Allrightsreserved.Nopartofthispublicationmaybereproducedortransmittedinanyformorbyanymeans,includingphotocopyingandrecording,orbyanyinformation

storageandretrievalsystem.

GovernanceintheAgeofGenerativeAI2

GovernanceintheAgeofGenerativeAI3

October2024

GovernanceintheAgeofGenerativeAI:

A360oApproachforResilientPolicyandRegulation

Foreword

ArnabChakraborty

ChiefResponsible

AIOfficer,Accenture

Wearelivinginatimeofrapidinnovationandglobaluncertainty,inwhichgenerativeartificialintelligence(AI)standsoutasatransformativeforce.This

technologyimpactsvariousindustries,economiesandsocietiesworldwide.WiththeEuropeanUnion’s(EU’s)AIActnowineffect,wehaveaprecedent

forcomprehensiveAIregulation.TheUS,Canada,Brazil,theAfricanUnion,JapanandChinaarealsodevelopingtheirownregulatoryapproaches.Thispivotalmomentcallsforvisionaryleadershipandacollaborativeapproachtoanticipatorygovernance.

Overthepastyear,theAIGovernanceAlliance

hasunitedindustryandgovernmentwithcivil

societyandacademia,establishingaglobal

multistakeholderefforttoensureAIservesthe

greatergoodwhilemaintainingresponsibility,

inclusivityandaccountability.Wehavebeenabletopositionourselvesasasoundingboardfor

policy-makerswhoaregrapplingwiththedifficultiesofdevelopingAIregulatoryframeworks,andto

conveneallplayersfromtheAIvaluechaintocreateameaningfuldialogueonemergingAIdevelopmentissues.

WithAccentureasitsknowledgepartner,the

Alliance’sResilientGovernanceandRegulation

workinggroup(composedofover110members),hascontributedtoshapingasharedunderstandingoftheglobalregulatorylandscape.Thegrouphas

workedtoestablishacomprehensivegovernanceframeworkthatcouldbeusedtoregulate

generativeAIusewellintothefuture.

CathyLi

Head,AI,DataandMetaverse;DeputyHead,Centreforthe

FourthIndustrialRevolution;

Member,ExecutiveCommittee,WorldEconomicForum

Thispaperisaculminationofthoseeffortsand

equipspolicy-makersandregulatorswithaclear

roadmapforaddressingthecomplexitiesof

generativeAIbyexaminingexistingregulatory

gaps,theuniquegovernancechallengesof

variousstakeholdersandtheevolvingformsofthistechnology.Theoutputsofthispaperaredesignedtobepracticalandimplementable,providingglobalpolicy-makerswiththetoolstheyneedtoenhancegenerativeAIgovernancewithintheirjurisdictions.Throughthispaper,ourAIGovernanceAlliance:

BriefingPaperSeries,launchedinJanuary2024,

andoureventsandcommunitymeetings,weseektocreateatangibleimpactinAIliteracyandknowledgedissemination.

Giventheinternationalcontextinwhichthis

technologyoperates,weadvocateforaharmonizedapproachtogenerativeAIgovernancethat

facilitatescooperationandinteroperability.SuchanapproachisessentialforaddressingtheglobalchallengesposedbygenerativeAIandforensuringthatitsbenefitsaresharedequitably,particularlywithlow-resourceeconomiesthatstandtogainsignificantlyfromitsresponsibledeployment.

Weinvitepolicy-makers,industryleaders,

academicsandcivilsocietytojoinusinthis

endeavour.Together,wecanshapeafuturewheregenerativeAIcontributespositivelytoourworldandensuresaprosperous,inclusiveandsustainable

futureforall.

Executivesummary

Governmentsshouldaddressregulatorygaps,engagemultiplestakeholdersinAIgovernanceandprepareforfuture

generativeAIrisks.

GovernanceintheAgeofGenerativeAI4

1

2

3

TherapidevolutionandswiftadoptionofgenerativeAIhavepromptedgovernmentstokeeppaceandprepareforfuturedevelopmentsandimpacts.

Policy-makersareconsideringhowgenerative

artificialintelligence(AI)canbeusedinthe

publicinterest,balancingeconomicandsocial

opportunitieswhilemitigatingrisks.Toachievethispurpose,thispaperprovidesacomprehensive

360°governanceframework:

Harnesspast:UseexistingregulationsandaddressgapsintroducedbygenerativeAI.

Theeffectivenessofnationalstrategiesfor

promotingAIinnovationandresponsible

practicesdependsonthetimelyassessmentoftheregulatoryleversathandtotacklethe

uniquechallengesandopportunitiespresentedbythetechnology.PriortodevelopingnewAIregulationsorauthorities,governmentsshould:

–AssessexistingregulationsfortensionsandgapscausedbygenerativeAI,

coordinatingacrossthepolicyobjectivesofmultipleregulatoryinstruments

–Clarifyresponsibilityallocationthroughlegalandregulatoryprecedentsand

supplementeffortswheregapsarefound

–Evaluateexistingregulatoryauthoritiesforcapacitytotacklegenerative

AIchallengesandconsiderthe

trade-offsforcentralizingauthoritywithinadedicatedagency

Buildpresent:Cultivatewhole-of-societygenerativeAIgovernanceandcross-sectorknowledgesharing.

Governmentpolicy-makersandregulatorscannotindependentlyensuretheresilient

governanceofgenerativeAI–additional

stakeholdergroupsfromacrossindustry,civilsocietyandacademiaarealsoneeded.Governmentsmustuseabroadersetof

governancetools,beyondregulations,to:

–Addresschallengesuniquetoeach

stakeholdergroupincontributingto

whole-of-societygenerativeAIgovernance

–Cultivatemultistakeholder

knowledge-sharingandencourageinterdisciplinarythinking

–LeadbyexamplebyadoptingresponsibleAIpractices

Planfuture:IncorporatepreparednessandagilityintogenerativeAIgovernanceand

cultivateinternationalcooperation.

GenerativeAI’scapabilitiesareevolving

alongsideothertechnologies.Governmentsneedtodevelopnationalstrategiesthat

considerlimitedresourcesandglobaluncertainties,andthatfeatureforesight

mechanismstoadaptpoliciesandregulationstotechnologicaladvancementsandemerging

risks.Thisnecessitatesthefollowingkeyactions:

–TargetedinvestmentsforAIupskillingandrecruitmentingovernment

–HorizonscanningofgenerativeAIinnovationandforeseeablerisks

associatedwithemergingcapabilities,convergencewithothertechnologiesandinteractionswithhumans

–Foresightexercisestoprepareformultiplepossiblefutures

–Impactassessmentandagile

regulationstoprepareforthedownstreameffectsofexistingregulationandforfutureAIdevelopments

–Internationalcooperationtoalignstandardsandrisktaxonomiesandfacilitatethesharingofknowledgeandinfrastructure

Introduction

A360°frameworkisneededforresilient

generativeAIgovernance,balancinginnovationandriskacrossdiversejurisdictions.

FIGURE1

A360oapproachforresilientpolicyandregulation

t

s

a

p

s

s

e

n

r

a

H

:

1

r

e

r

p

d

l

i

u

B

:

2

r

a

l

l

i

P

a

l

l

i

P

t

n

se

a

u

t

u

r

e

P

i

l

l

a

r

Asorganizationsandindividualsconsiderhowbesttoadoptgenerativeartificialintelligence(AI),new

powerfulcapabilitiescontinuetoemerge.Forsome,humanity’sfuturewithgenerativeAIcanfeelfullofpromise,andforothers,concern.Indeed,across

industriesandsectors,generativeAIpresents

bothopportunitiesandrisks.Forexample–will

generativeAIenhancepersonalizedtreatment

plansimprovingpatients’healthoutcomes,orwillitinducenovelbiosecurityrisks?Willjournalismbedemocratizedthroughnewstorytellingtools,orwilldisinformationbescaled?

ThereisnosingleguaranteedfutureforgenerativeAI.Rather,howsocietyadaptstothetechnologywilldependonthedecisionshumansmakein

researching,developing,deployingandexploitingitscapabilities.Policy-makers,througheffective

governance,canhelptoensurethatgenerativeAIfacilitateseconomicopportunityandfairdistributionofbenefits,protectshumanrights,promotes

greaterequityandencouragessustainable

practices.Governancedecisionsmadenowwillshapethelivesofpresentandfuture

generations,how(andwhether)thistechnologybenefitssocietyandwhoisleftbehind.

Inresponsetothecontinuedgrowthofthe

generativeAIindustryandrapidadoptionof

itsapplicationsacrosstheworld,thispaper’s

360°frameworkoutlineshowtobuildresilientgovernancethatfacilitatesAIinnovationwhile

mitigatingrisks,fromthedevelopmentstagetoitsuse.Theframeworkisdesignedtosupport

policy-makersandregulatorsinthedevelopmentofholisticanddurablegenerativeAIgovernance.Thespecificimplementationoftheframework,

however,willdifferbetweenjurisdictions,

dependingonthenationalAIstrategy,maturityofAInetworks,economicandgeopoliticalcontexts,individuals’expectationsandsocialnorms.

MakeuseofexistingregulationsandaddressgapscausedbygenerativeAI.

360o

governance

Encouragewhole-of-societygenerative

AIgovernanceandcross-sectorknowledgesharing.

3:Plnf

IncorporatepreparednessandagilityingenerativeAIgovernanceand

facilitateinternationalcooperation.

GovernanceintheAgeofGenerativeAI5

Harnesspast

Greaterclarityandcertaintyregarding

existingregulatoryenvironmentsisnecessarytoaddressemerginggenerativeAIchallengesandopportunities.

1.1

oWithincreasingdigitalizationandagrowingtrend

ofmonetizing

personaland

professionaldata,protectionof

privacyisboth

vitalandcomplex.Policy-makersarelookingtoprioritizeprivacy-preservingconsiderations.

ExamineexistingregulationscomplicatedbygenerativeAIattributes

GovernanceintheAgeofGenerativeAI6

Pillar1

Successfulimplementationofnationalstrategiesforresponsibleandtrustworthygovernance

ofgenerativeAIrequiresatimelyassessmentofexistingregulatorycapacity–amongothergovernancetools–totackletheunique

opportunitiesandrisksposedbythetechnology.Thisincludesexaminationoftheadequacyofexisting

legalinstruments,lawsandregulations,resolutionofregulatorytensionsandgaps,clarification

ofresponsibilityallocationamonggenerativeAI

supplychainactorsandevaluationofcompetent

regulatoryauthorities’effectivenessandcapacities.Suchassessmentsmustrespectthefundamental

rightsandfreedomsalreadycodifiedininternationalhumanrightslaw,suchastheprotectionof

particulargroups(e.g.minorityrights1andchildren’srights2)aswellaslegalinstrumentsthataredomain-specific(e.g.tocybercrime3andclimatechange4).5

WhilegenerativeAI’semergingpropertiesand

capabilitiesmaywarrantnovelregulations,policy-makersandregulatorsshouldfirstexaminetheirjurisdiction’sexistingregulationsforaddressing

newchallenges.Theyshouldalsoidentifywhereexistingregulationsmaybeapplied,adaptedorforegonetofacilitatetheobjectivesofanationalAIstrategy.NavigatinggenerativeAI’sinteractionswithexistingregulationsrequiresanuanced

understandingofboththetechnicalaspects

andthelegalprinciplesunderlyingtheimpactedregulations.Table1discussesexamplesofhowregulatoryinstrumentscanbecomplicatedinthecontextofgenerativeAI.

Privacyanddataprotection

GenerativeAImodelsamplifyprivacy,safetyandsecurityrisksduetotheirrelianceonvastamountsoftrainingdata,powerfulinferencecapabilityandsusceptibilitytouniqueadversarialattacksthat

canunderminedigitaltrust.6Anumberofrisks

arisefromtheinclusionofpersonal,sensitiveandconfidentialinformationintrainingdatasetsanduser

inputs,lackoftransparencyoverthelawfulbasisforcollectingandprocessingdata,theabilityofmodelstoinferpersonaldataandthepotentialformodelstomemorizeanddiscloseportionsoftrainingdata.Withincreasingdigitalizationandagrowingtrendofmonetizingpersonalandprofessionaldata,

protectionofprivacyisbothvitalandcomplex.

Policy-makersarelookingtoprioritizeprivacy-

preservingconsiderationsapplicabletodigitaldatawhilealsocreatingaffordancesfordatapoolingthatcouldleadtoAI-facilitatedbreakthroughs.7Suchaffordancescouldbemadetopromoteinnovationforpublicgoodsinareassuchasagriculture,

healthandeducation,orwithinnarrowlyspecifiedexceptionsfordataconsortiathatfacilitatethe

trainingofAImodelstoachievepublicpolicy

objectives.8Anotheremergingissueforpolicy-

makersisthatofensuringgenerativeAIsafety

andsecurity,evenwhenitmayinvolveinteractionwithpersonaldata,asinthecaseofinvestigatingandrespondingtosevereincidents.Thiscould

beaddressedthroughthecreationofregulatoryexceptionsandguardrailstoensurebothprivacyandresponsibleAIoutcomes.

GovernanceintheAgeofGenerativeAI7

ISSUESPOTLIGHT1

TraininggenerativeAIsystemsoncopyright-protecteddata,andtensionswiththetextanddataminingexception

Copyrightand

intellectualproperty

GenerativeAIraisesseveralissuesrelatingto

copyrightinfringement,plagiarismandintellectualproperty(IP)ownership(seeIssuespotlight1),

someofwhicharecurrentlybeingconsideredbycourtsinvariousjurisdictions.Rightsrelatedtoprotectinganindividual’slikeness,voiceandotherpersonalattributesarealsoimplicatedby

thecreationof“deepfakes”usinggenerativeAI.

AblanketrulingonAItrainingisuncertainand

judgescoulddeterminethefairnessofcertaindatausesforspecificproductsbasedontheproduct’sfeaturesoroutputs’frequencyandsimilarityto

trainingdata.9Lookingahead,thereisapressingneedforcomprehensiveexaminationofregulatoryframeworksandfornecessaryguidanceon

documentinghumancreativityinthegenerationofcontentasameansofassertingIPprotection.

Textanddatamining(TDM)istheautomated

processofdigitallyreproducingandanalysing

largequantitiesofdataandinformationto

identifypatternsanddiscoverresearchinsights.Variousjurisdictionsaroundtheworld–such

asJapan,Singapore,Estonia,SwitzerlandandtheEuropeanUnion(EU)–haveintroduced

specificexemptionswithintheircopyrightlawstoenableTDMextractionfromcopyright-protectedcontenttoinnovate,advancescienceandcreatebusinessvalue.

GiventhevastamountsofdatathatgenerativeAIsystemsusetotrainonandgeneratenew

content,jurisdictionsshouldestablishregulatoryclarityregardingTDMforthepurposeof

generativeAItraining.Thiscouldbedone,forexample,byconfirmingwhetherAIdevelopmentconstitutes“fairdealing”or“fairuse”(akey

defenceagainstcopyrightinfringement)or

fallswithintheexemptionsrecognizedinsome

copyrightlaws.CountriesliketheUKareexploringsuchregulatoryexceptions,seekingtopromote

apro-innovationAIagenda.10Ultimately,thereismountingpressureongovernmentstoresolvethecopyrighttensiondefinitively.11

Licensinganddataaccessonan“opt-in”or

“opt-out”basisarealsounderexaminationto

addressTDMconcerns,inadditiontoarangeoftechnologiesandstandardsthatattempttocedecontroltocreators,allowingthemtooptoutfrommodeltrainers.12Licensingproponentsargue

thatscrapingforgenerativeAItrainingwithout

payingcreatorsconstitutesunlawfulcopyingandisaformofreducingcompetition.13AIdevelopers,however,arguethatrequirementstopay

copyrightownersforcontentusedintraining

wouldconstrainmodeldevelopment,negativelyimpactventurecapital(VC)fundingandreducecompetitionamonggenerativeAImodels.14WhiletheydonoteliminateIPlawconcernsentirely,

opt-in/outandlicensingeffortscouldcontributetosettingstandardsthatgenerativeAIfoundationmodelproviderswouldbeexpectedtouphold.

Consumerprotectionandproductliability

WhileAI-specificregulationremainsvoluntaryor

pendinginjurisdictionsoutsideoftheEU,consumerregulationandproductliabilitylawscontinueto

beapplicable,regardlessofwhethertheystrictlycontemplateAIorothertechnologies.GenerativeAIhasthepotentialtoinfluencetheconsumer

marketbyautomatingvarioustasksandservices.Thismay,however,alsochallengetraditional

approachestoriskassessmentandmitigation(duetothetechnology’sbroadapplicabilityandability

tocontinuallylearnandgeneratenewanduniquecontent),aswellasproductsafetystandards

(forexample,inhealthandphysicalsafety).Thedevelopmentofstandardsshouldbeaniterative,multidisciplinaryprocessthatkeepspacewith

technologicaladvancements.

TABLE1

SelectionofcomplexitiesintroducedbygenerativeAIforexistingregulatoryareas

accesstoavastnumberofusers,contributingtoeconomiesofscalethatchallengecompetition.15Inresponse,competitionauthoritiesaroundtheglobearestartingtoprovideguidanceoncompetition

risksandexpectationsingenerativeAImarkets.16CompetitioncomplexitiesateachlayeroftheAI

stackwillneedtobeevaluatedasthetechnologyevolvestoenableaccessandchoiceacrossAI

models,includinggeneral(e.g.ChatGPT),area-

specific(e.g.modelsdesignedforhealthcare)andpersonalusemodels.Suchevaluationswillalso

needtobeconsideredalongsideexistinglegislationrelatingtonationalsecurity,freedomofexpression,mediaandassembly.

Competition

Marketauthoritiesmustensurethatthecompetitiveconditionsdrivingtherapidpaceofinnovation

continuetobenefitconsumers.Althoughexisting

competitionlawsremainapplicable,generativeAI

raisesnewconcernsrelatedtotheconcentrationofcontrolovercriticalcomponentsofthetechnologyandcertainpartnershiparrangements.Forexample,generativeAI’scapabilitiesareenhancedwith

accesstohigh-performancecomputecapacitiesandcertaindatasetsthatmayprovecriticalformodeldevelopment.Thelattercandependon

GovernanceintheAgeofGenerativeAI8

Emergingcomplexities(non-exhaustive)

Regulatoryarea

Emergingstrategiesunderconsiderationbyregulators(non-exhaustive)

LegalbasisforuserdatabeingusedtotraingenerativeAImodels

Incidentalcollectionofpersonaldatabyweb-crawlers

Enforcementofdata-minimizationprinciples17andopt-in/outrightsbygenerativeAIprovidersanddeployers18

Clarifyingwebterms-of-serviceagreementsandencouragingprivacy-enhancingtechnologicalmeasuressuchasthe

detectionandredactionofpersonallyidentifiableinformation19

Guidanceforpurposethresholdswithindomain-specificregulations,e.g.financialservices20

Positionstatementshighlightingexpectationsforsafetymeasuresandpreferencesforemergingbestpractices21

Privacyanddataprotection

Specifyingpurposelimitationsfordatacollection

Onlinesafetyandprotectionofvulnerable

groups,especiallyminors,fromharmfuloutputs

Copyrightinfringementoftrainingdata

ClearpolicypositionsandaccumulationoflegalprecedentsontherelationsbetweencopyrightandgenerativeAI22

IPrightsandownershipofworksgenerated

Guidanceonassessingtheprotectableelements

byAI

ofAI-generatedworks23

CopyrightandIPAttributionandfaircompensationforartists

andcreators

Investmentsinsolutionsforattributionandauthorrecognitionsuchaswatermarkingandcontentprovenance,alongwith

privacyanddataprotection

ExtensionofgenerativeAImodeltraining

ConsiderationsofnewIPchallengesandclassificationsrelated

toadditionaldatamodalities(e.g.sensory,biological,motion)

toemergingdatamodalities

Liabilityobligationsresultingfromscopeofmultipleapplicableregulations

Considerationsaroundwhetherandinwhichcasesaconcerniscoveredbytheexistingregulations

Thelackofaspecificpurposeofthegenerative

CombiningtheconventionalAIfaultanddefectiveness

AImodelbeforeitsimplementationcomplicates

criteriawithnewmethodsdesignedforgenerativeAI’s

liabilityarisingfromdefectivenessandfault

technicalnuances

Consumer

protectionandproductliability

Efficacyofevidentialdisclosurerequirements

Broadeningthedisclosurerequirementtoencourage

transparencyviaexplainability,traceabilityandauditability,andincludesystemsthatarenotjustclassifiedashigh-risk

Businessconductoragreementsthatenableadominantfirmtoexcluderivals

InitiatingsectoralstudiestodevelopabaselineunderstandingofthecompetitivedynamicsoftheAItechnologystack,

reviewingagreementsbetweenindustryplayersandexaminingsinglefirmconduct24

Unfairordeceptivepractice

Competition

Issuingguidanceonunfairordeceptivepracticeprohibitionsifitdoesnotexist25

StakeholderconsultationsonhowgenerativeAIimpactscompetitioninimportantmarkets,e.g.searchengines,

onlineadvertising,cloudcomputingandsemiconductors26

Impactofdownstreamapplicationsoncompetitionacrossseveralsectors

1.2Resolvetensionsbetweenpolicyobjectivesofmultipleregulatoryregimes

oRegulatorsmustaddressemergingtensionsand

mitigatetheriskofundermininglegalcertaintyandrespect

forlegitimateexpectations.

throughtheEU’sGeneralDataProtectionRegulation(GDPR).Asimilartensionemergesbetweencopyrightlaw–whichprotectstherightsofcreatorsand

inventorsensuringthattheycancontrolandprofitfromtheircreations–andgenerativeAIinnovation,whichoftenusescopyrightedmaterialfortraining.

Addressingtensionsbetween

horizontalandverticalregulations

Horizontalregulationsmayalsoconflictwith

verticalregulationstailoredtospecificsectors.Forinstance,financialinstitutionsusinggenerativeAI

mayencounterchallengesbalancinghorizontalprivacyregulationswithfinancialsectorknow-

your-client(KYC)procedures.Wheredata

protectionregulationsrequireorganizationsto

minimizepersonaldatacollectionlinkedtoa

specificpurpose,KYCguidelinesrequirefinancialinstitutionstoconductthoroughduediligenceonclientstoensurecompliancewithanti-money-

launderinglaws.

TheintersectionalnatureofgenerativeAI

technologiesandtheapplicabilityofmultiple

regulatoryinstrumentscreatesacomplex

environmentwhereregulatoryframeworksoftenoverlapandconflictduetocompetingpolicy

objectives.Astechnologyevolvesandbecomesmorewidelyadopted,regulatorsmustaddressemergingtensionsandmitigatetheriskof

undermininglegalcertaintyandrespectforlegitimateexpectations.

Addressingtensionsbetweenhorizontalregulations

Multiplehorizontalregulations,whichaimtocreatebroad,industry-agnosticstandards,mayconflictwhentheyimposerequirementsthataredifficult

toreconcileacrossgenerativeAIcontextsor

applications.Forexample,generativeAImodeldevelopersmayhavetroubleidentifyingthe

appropriatelawfulbasisfordataprocessingand

deliveryaccordingtodataprotectionrightsarticulated

GovernanceintheAgeofGenerativeAI9

ChallengesandconsiderationsforgenerativeAIresponsibilityallocation(non-exhaustive)

AsdefinedintheWorldEconomicForum’sDigitalTrustFramework,27maintainingaccountability

andoversightfortrustworthydigitaltechnologiesrequiresclearlyassignedandwell-definedlegalresponsibilitiesalongsideremedyprovisionsforupholdingindividualandsocialexpectations.

GenerativeAIintroducescomplexitiesinto

traditionalresponsibilityallocationpractices,as

examinedinTable2.Policy-makersshouldconsiderwheresupplementaryeffortsareneededtoaddressgapsandwherelegalandregulatoryprecedents

canhelptoclarifygenerativeAIresponsibility.Theissuanceofeffectiveguidancerequires

considerationofhowliabilitywithinthegenerativeAIsupplychaincanvaryfordifferentrolesandactorsaswellasconsiderationofretroactiveliabilities

anddispute-resolutionprovisions.Unresolvedambiguityinresponsibilityallocationcanlimitinvestorconfidence,createanunevenplayingfieldforvarioussupplychainactorsandleaverisksunaddressedandharmswithoutredress.

TABLE2

GovernanceintheAgeofGenerativeAI10

Clarifyexpectationsaroundresponsibilityallocation

1.3

Examplechallenges

Considerationsforpolicy-makers

–Case-basedreview:Policy-makersshould

providegeneralallocationguidancetocultivate

predictability,butincludemechanismsthatallowcasecomplexitiestodeterminepreciseallocation.Requiringactorstoidentifyresponsibilityhand-offsisoneapproachbeingexaminedbyjurisdictions.

–Terminology:Policy-makersshouldcollaborate

toarriveatsharedtermsformodels,applicationsandroles,e.g.inlinewithISO420001fromthe

InternationalOrganizationforStandardization(ISO).

–Regulatorycarve-outs:Policy-makersshouldlimitinstanceswhenusecanleadtounfairadvantages,suchaswheresomeentitiesareabletobypass

crucialsafeguardsandaccountabilitymeasuresorengageinregulatoryarbitrage.

–Proportionality:Policy-makersshouldconsiderthecontrol,influenceandresourceseachactorhasinthegenerativeAIlifecycle,andabilitytoredressissuesresultinginharm.

–Third-partycertifications:Policy-makersshould

considerappropriatenessandnecessityofusingthirdpartiesforarobustAIcertificationsystem

(potentiallydefinedthrough