CONTENTS

ExecutiveSummary

3

Accountability

6

Data

9

TrustedDevelopment

12

andDeployment

IncidentReporting

16

TestingandAssurance

19

Security

21

ContentProvenance

23

SafetyandAlignmentR&D26

AIforPublicGood

28

Conclusion

31

Acknowledgements

32

FurtherDevelopment

34

3

MODELAIGOVERNANCEFRAMEWORKFORGENERATIVEAI

EXECUTIVESUMMARY

GenerativeAIhascapturedtheworld’simagination.Whileitholdssignificanttransformativepotential,italsocomeswithrisks.Buildingatrustedecosystemisthereforecritical—ithelpspeopleembraceAIwithconfidence,givesmaximalspaceforinnovation,andservesasacorefoundationtoharnessingAIforthePublicGood.

AI,asawhole,isatechnologythathasbeendevelopingovertheyears.PriordevelopmentanddeploymentissometimestermedtraditionalAI.1TolaythegroundworktopromotetheresponsibleuseoftraditionalAI,Singaporereleasedthefirstversionofthe

ModelAIGovernanceFramework

in2019,andupdateditsubsequentlyin2020.2TherecentadventofgenerativeAI3hasreinforcedsomeofthesameAIrisks(e.g.,bias,misuse,lackofexplainability),andintroducednewones(e.g.,hallucination,copyrightinfringement,valuealignment).Theseconcernswerehighlightedinourearlier

DiscussionPaperonGenerativeAI:Implicationsfor

TrustandGovernance,

4issuedinJune2023.Thediscussionsandfeedbackhavebeeninstructive.

Existinggovernanceframeworksneedtobereviewedtofosterabroadertrustedecosystem.Acarefulbalanceneedstobestruckbetweenprotectingusersanddrivinginnovation.Therehavealsobeenvariousinternationaldiscussionspullingintherelatedandpertinenttopicsofaccountability,copyrightandmisinformation,amongothers.Theseissuesareinterconnectedandneedtobeviewedinapracticalandholisticmanner.Nosingleinterventionwillbeasilverbullet.

ThisModelAIGovernanceFrameworkforGenerativeAIthereforeseekstosetforthasystematicandbalancedapproachtoaddressgenerativeAIconcernswhilecontinuingtofacilitateinnovation.Itrequiresallkeystakeholders,includingpolicymakers,industry,theresearchcommunityandthebroaderpublic,tocollectivelydotheirpart.ThereareninedimensionswhichtheFrameworkproposestobelookedatintotality,tofosteratrustedecosystem.

a)Accountability—AccountabilityisakeyconsiderationtoincentiviseplayersalongtheAIdevelopmentchaintoberesponsibletoend-users.Indoingso,werecognisethatgenerativeAI,likemostsoftwaredevelopment,involvesmultiplelayersinthetechstack,andhencetheallocationofresponsibilitymaynotbeimmediatelyclear.WhilegenerativeAIdevelopmenthasuniquecharacteristics,usefulparallelscanstillbedrawnwithtoday’scloudandsoftwaredevelopmentstacks,andinitialpracticalstepscanbetaken.

1TraditionalAIreferstoAImodelsthatmakepredictionsbyleveraginginsightsderivedfromhistoricaldata.TypicaltraditionalAImodelsincludelogisticregression,decisiontreesandconditionalrandomfields.Othertermsusedtodescribethisinclude“discriminativeAI”.

2ThefocusoftheModelAIGovernanceFrameworkistosetoutbestpracticesforthedevelopmentanddeploymentoftraditionalAIsolutions.ThishasbeenincorporatedintoandexpandedundertheTrustedDevelopmentandDeploymentdimensionoftheModelAIGovernanceFrameworkforGenerativeAI.

3GenerativeAIareAImodelscapableofgeneratingtext,imagesorothermediatypes.Theylearnthepatternsandstructureoftheirinputtrainingdataandgeneratenewdatawithsimilarcharacteristics.Advancesintransformer-baseddeepneuralnetworksenablegenerativeAItoacceptnaturallanguagepromptsasinput,includinglargelanguagemodels(LLM)suchasGPT-4,Gemini,ClaudeandLLaMA.

4TheDiscussionPaperwasjointlypublishedbytheInfocommMediaDevelopmentAuthorityofSingapore(IMDA),AicadiumandAIVerifyFoundation.See

https://aiverifyfoundation.sg/downloads/Discussion

_Paper.pdf

4

MODELAIGOVERNANCEFRAMEWORKFORGENERATIVEAI

b)Data—Dataisacoreelementofmodeldevelopment.Itsignificantlyimpactsthequalityofthemodeloutput.Hence,whatisfedtothemodelisimportantandthereisaneedtoensuredataquality,suchasthroughtheuseoftrusteddatasources.Incaseswheretheuseofdataformodeltrainingispotentiallycontentious,suchaspersonaldataandcopyrightmaterial,itisalsoimportanttogivebusinessclarity,ensurefairtreatment,andtodosoinapragmaticway.

c)TrustedDevelopmentandDeployment—Modeldevelopment,andtheapplicationdeploymentontopofit,areatthecoreofAI-driveninnovation.Notwithstandingthelimitedvisibilitythatend-usersmayhave,meaningfultransparencyaroundthebaselinesafetyandhygienemeasuresundertakeniskey.Thisinvolvesindustryadoptingbestpracticesindevelopment,evaluation,andthereafter“foodlabel”-typetransparencyanddisclosure.Thiscanenhancebroaderawarenessandsafetyovertime.

d)IncidentReporting—Evenwiththemostrobustdevelopmentprocessesandsafeguards,nosoftwareweusetodayiscompletelyfoolproof.ThesameappliestoAI.Incidentreportingisanestablishedpractice,andallowsfortimelynotificationandremediation.Establishingstructuresandprocessestoenableincidentmonitoringandreportingisthereforekey.ThisalsosupportscontinuousimprovementofAIsystems.

e)TestingandAssurance—Foratrustedecosystem,third-partytestingandassuranceplaysacomplementaryrole.Wedothistodayinmanydomains,suchasfinanceandhealthcare,toenableindependentverification.AlthoughAItestingisanemergingfield,itisvaluableforcompaniestoadoptthird-partytestingandassurancetodemonstratetrustwiththeirend-users.ItisalsoimportanttodevelopcommonstandardsaroundAItestingtoensurequalityandconsistency.

f)Security—GenerativeAIintroducesthepotentialfornewthreatvectorsagainstthemodelsthemselves.Thisgoesbeyondsecurityrisksinherentinanysoftwarestack.Whilethisisanascentarea,existingframeworksforinformationsecurityneedtobeadaptedandnewtestingtoolsdevelopedtoaddresstheserisks.

g)ContentProvenance—AI-generatedcontent,becauseoftheeasewithwhichitcanbecreated,canexacerbatemisinformation.Transparencyaboutwhereandhowcontentisgeneratedenablesend-userstodeterminehowtoconsumeonlinecontentinaninformedmanner.Governmentsarelookingtotechnicalsolutionslikedigitalwatermarkingandcryptographicprovenance.Thesetechnologiesneedtobeusedintherightcontext.

h)SafetyandAlignmentResearch&Development(R&D)—Thestate-of-the-sciencetodayformodelsafetydoesnotfullycoverallrisks.AcceleratedinvestmentinR&Disrequiredtoimprovemodelalignmentwithhumanintentionandvalues.GlobalcooperationamongAIsafetyR&Dinstituteswillbecriticaltooptimiselimitedresourcesformaximumimpact,andkeeppacewithcommerciallydrivengrowthinmodelcapabilities.

i)AIforPublicGood—ResponsibleAIgoesbeyondriskmitigation.ItisalsoaboutupliftingandempoweringourpeopleandbusinessestothriveinanAI-enabledfuture.DemocratisingAIaccess,improvingpublicsectorAIadoption,upskillingworkersanddevelopingAIsystemssustainablywillsupporteffortstosteerAItowardsthePublicGood.

5

3.Trusted

Developmentand

Deployment

Enhancing

transparencyaround

baselinesafetyand

hygienemeasures

basedonindustry

bestpractices

indevelopment,

evaluationand

disclosure

5.TestingandAssurance

Providingexternal

validationand

addedtrustthrough

third-partytesting,

anddeveloping

commonAItesting

standardsfor

consistency

4.Incident

Reporting

Implementingan

incidentmanagement

systemfortimely

notification,

remediation

andcontinuous

improvements,asno

AIsystemisfoolproof

7.ContentProvenance

Transparencyaboutwherecontentcomes

fromusefulforend-

assignalsusers

MODELAIGOVERNANCEFRAMEWORKFORGENERATIVEAI

FosteringaTrustedAIEcosystem

1.Accountability

PuttinginplacetherightincentivestructurefordifferentplayersintheAIsystemdevelopmentlifecycletoberesponsibletoend-users

2.Data

Ensuringdataquality

andaddressing

potentiallycontentious

trainingdataina

pragmaticway,as

dataiscoretomodel

development

6.Security

AddressingnewthreatvectorsthatarisethroughgenerativeAImodels

8.SafetyandAlignmentR&D

AcceleratingR&DthroughglobalcooperationamongAISafetyInstitutestoimprovemodelalignmentwithhumanintentionandvalues

9.AIforPublicGood

ResponsibleAIincludesharnessingAItobenefitthepublicbydemocratisingaccess,

improvingpublicsectoradoption,upskillingworkersanddevelopingAIsystemssustainably

ThisFrameworkbuildsonthepolicyideashighlightedinourDiscussionPaperonGenerativeAIanddrawsfrominsightsanddiscussionswithkeyjurisdictions,internationalorganisations,researchcommunitiesandleadingAIorganisations.TheFrameworkwillevolveastechnologyandpolicydiscussionsdevelop.

ACCOUNTABILITY

7

MODELAIGOVERNANCEFRAMEWORKFORGENERATIVEAI

FOSTERINGATRUSTEDAIECOSYSTEM

ACCOUNTABILITY

Accountabilityisakeyconsiderationinfosteringatrustedecosystem.PlayersalongtheAIdevelopmentchainneedtoberesponsibletowardsend-users,5andthestructuralincentivesshouldalignwiththisneed.Theseplayersincludemodeldevelopers,applicationdeployers6andcloudserviceproviders(whooftenprovideplatformsonwhichAIapplicationsarehosted).GenerativeAI,likemostsoftwaredevelopment,involvesmultiplelayersinthetechstack.Whiletheallocationofresponsibilitymaynotbeimmediatelyclear,usefulparallelscanbedrawnwithtoday’scloudandsoftwaredevelopment,andpracticalstepscanbetaken.

Design

Todothiscomprehensively,thereshouldbeconsiderationforhowresponsibilityisallocatedbothupfrontinthedevelopmentprocess(ex-ante)asbestpractice,andguidanceonhowredresscanbeobtainedifissuesarediscoveredthereafter(ex-post).

ExAnte—AllocationUpfront

ResponsibilitycanbeallocatedbasedonthelevelofcontrolthateachstakeholderhasinthegenerativeAIdevelopmentchain,sothattheablepartytakesnecessaryactiontoprotectend-users.Asareference,whiletheremaybevariousstakeholdersinthedevelopmentchain,thecloudindustry7hasbuiltandcodifiedcomprehensivesharedresponsibilitymodelsovertime.Theobjectiveistoensureoverallsecurityofthecloudenvironment.Thesemodelsallocateresponsibilitybyexplainingthecontrolsandmeasuresthatcloudserviceproviders(whoprovidethebaseinfrastructurelayer)andtheircustomers(whohostapplicationsonthelayerabove)respectivelyundertake.

ThereisvalueinextendingthisapproachtoAIdevelopment.CloudserviceprovidershaverecentlyextendedsomeelementsoftheircloudsharedresponsibilitymodelstocoverAI,placinginitialfocusonsecuritycontrols.8Thisisagoodstart,andasimilarapproachcanbetakentoaddressothersafetyconcerns.TheAIsharedresponsibilityapproachmayalsoneedtoconsiderdifferentmodeltypes(e.g.,closed-source,open-source9oropen-weights10),giventhedifferentlevelsofcontrolthatapplicationdeployershaveforeachmodeltype.Responsibilityin

5WhiletheFrameworkplacesemphasisonallocatingresponsibilitiesforAIdevelopment,end-usershaveseparateresponsibilitiesforAIuse(e.g.,abidingbytermsofuse).

6WerecognisethatthegenerativeAIdevelopmentchainiscomplex.Applicationdevelopers(whodevelopsolutionsorapplicationsthatmakeuseofAItechnology)andapplicationdeployers(whoprovideAIsolutionsorapplicationstoend-users)cansometimesbetwodifferentparties.Forsimplicity,thispaperusestheterm“applicationdeployers”torefertobothapplicationdevelopersanddeployers.

7ThisincludesGoogleCloud,MicrosoftAzureandAmazonWebServices.

8Microsoft,whichisbothacloudandmodelserviceprovider,hasinitiatedsomeelementsofthis.See

https://learn.microsoft

.com/en-us/azure/security/fundamentals/shared-responsibility-ai

9Open-sourcingmakesavailablethefullsourcecodeandinformationrequiredforre-trainingthemodelfromscratch,includingmodelarchitecturecode,trainingmethodologyandhyperparameters,originaltrainingdatasetanddocumentation.Modelsthatareclosertothisendofthespectrum(butnotfullyopen)includeDollyandBLOOMZ.

10Open-weightsmakesavailablepre-trainedparametersorweightsofthemodelitself,butnotthetrainingcode,dataset,methodology,etc.Existingopen-weightsmodelsincludeLlaMa2,Falcon-40B-InstructandMistral7B-Instruct.

8

MODELAIGOVERNANCEFRAMEWORKFORGENERATIVEAI

thiscase,forexamplewhenusingopen-sourceoropen-weightsmodels,shouldrequireapplicationdeployerstodownloadmodelsfromreputableplatformstominimisetheriskoftamperedmodels.Beingthemostknowledgeableabouttheirownmodelsandhowtheyaredeployed,modeldevelopersarewell-placedtoleadthisdevelopmentinaconcertedmanner.11Thiswillprovidestakeholderswithgreatercertaintyupfront,andfosterasaferecosystem.

ExPost—SafetyNets

Sharedresponsibilitymodelsserveasanimportantfoundationforaccountability—theyprovideclarityonredresswhenissuesoccur.However,theymaynotbeabletocoverallpossiblescenarios.Allocatingresponsibilitywhenthereareneworunanticipatedissuesmayalsobepracticallychallenging.Itwillbeworthconsideringadditionalmeasures—includingconceptsaroundindemnityandinsurance—tobettercoverend-users.

Thisexistsinalimitedformtoday.Inclearerareaswhereredressisneeded,theindustryhasmovedaccordingly.Somemodeldevelopers12havebeguntounderwritecertainrisks,suchasthird-partycopyrightclaimsarisingfromtheuseoftheirAIproductsandservices.Indoingso,developersimplicitlyacknowledgetheirresponsibilityformodeltrainingdataandhowtheirmodelsareused.

Therewillinevitablybeotherareasthatarenotasclearandnotwell-covered.Thismayincluderisksthathavedisproportionateimpactonsocietyasawhole,andwhichmayonlyemergeasAIisused.Itisthereforeusefultoconsiderupdatinglegalframeworkstomakethemmoreflexible,andtoallowemergingriskstobeeasilyandfairlyaddressed.Thisisakintohowend-usersofphysicalproductstodayenjoysafetyprotections.OneexampleofsucheffortsistheEU’sproposedAILiabilityDirectiveandsoon-to-beapprovedRevisedProductLiabilityDirective.TheseDirectivesaimtomakeitsimplerforend-userstoprovedamagecausedbyAI-enabledproductsandservices.Thisensuresthatnopartyisunfairlydisadvantagedbythecompensationprocess.

Finally,thereareboundtoberesidualissuesthatfallthroughthecracks.Thisisaverynascentdiscussion,andalternativesolutionssuchasno-faultinsurance13couldbeconsideredasasafetynet.

11Thedetailsofhowresponsibilitieswillbeallocatedarekeyandwillneedtobeworkedoutgradually.

12Forexample,Adobe,Anthropic,Google,MicrosoftandOpenAI.

13Underano-faultinsuranceapproach,stakeholders’expensesarecoveredregardlessofwhoisatfault.ItiscurrentlyadoptedintheUSforsometypesofmotoraccidentclaims.ThisinsuranceapproachintheAIcontextwarrantsfurtherstudy.

DATA

10

MODELAIGOVERNANCEFRAMEWORKFORGENERATIVEAI

FOSTERINGATRUSTEDAIECOSYSTEM

DATA

Dataisacoreelementofmodelandapplicationdevelopment.AlargecorpusofdataisneededtotrainrobustandreliableAImodels.Givenitsimportance,businessesrequireclarityandcertaintyonhowtheycanusedatainmodeldevelopment.Thisincludespotentiallycontentiousareassuchaspubliclyavailablepersonaldataandcopyrightmaterial,whicharetypicallyincludedinweb-scrapeddatasets.Insuchcases,itisimportanttorecognisecompetingconcerns,ensurefairtreatment,andtodosoinapragmaticway.Inaddition,developingamodelwellrequiresgoodqualitydata,andinsomecircumstances,representativedata.Itisalsoimportanttoensuretheintegrityofavailabledatasets.14

Design

TrustedUseofPersonalData

Aspersonaldataoperateswithinexistinglegalregimes,ausefulstartingpointisforpolicymakerstoarticulatehowexistingpersonaldatalawsapplytogenerativeAI.Thiswillfacilitatetheuseofpersonaldatainamannerthatstillprotectstherightsofindividuals.15Forexample,policymakersandregulatorscanclarifyconsentrequirementsorapplicableexceptions,andprovideguidanceongoodbusinesspracticesfordatauseinAI.16

Anemerginggroupoftechnologies,knowncollectivelyasPrivacyEnhancingTechnologies(PETs),hasthepotentialtoallowdatatobeusedinthedevelopmentofAImodelswhileprotectingdataconfidentialityandprivacy.SomePETssuchasanonymisationtechniquesarenotnew,whileothertechnologiesarestillnascentandevolving.17TheunderstandingofhowPETscanbeappliedtoAIwillbeanimportantareatoadvance.

BalancingCopyrightwithDataAccessibility

Fromamodeldevelopmentperspective,theuseofcopyrightmaterialintrainingdatasetsandtheissueofconsentfromcopyrightownersisstartingtoraiseconcerns,particularlyastoremunerationandlicensingtofacilitatesuchuses.Modelsarealsoincreasinglybeingusedforgeneratingcreativeoutput—someofwhichmimicthestylesofexistingcreatorsandgiverisetoconsiderationsofwhetherthiswouldconstitutefairuse.18

14Datapoisoningattackstrainingdatasetsbyintroducing,modifyingordeletingspecificdatapoints.Forexample,withknowledgeoftheexacttimemodeldeveloperscollectcontent(e.g.,viasnapshots)fromsourceslikeWikipedia,badactorscan“poison”theWikipediawebpageswithfalsecontent,whichwillbescrapedandusedtotrainthegenerativeAImodel.Evenifthesourcemoderatorsundothechangesmadetothewebpages,thecontentwouldhavebeenscrapedandused.

15Thecollectionanduseofpersonaldataisalreadyprotectedundermanyexistingdataregimes.

16OneexampleofthisistheSingaporePersonalDataProtectionCommission’sAdvisoryGuidelinesonUseofPersonalDatainAIRecommendationandDecisionSystems.See

.sg/guidelines-and-consultation/2024/02/advisory-guidelines-on-use-of-personal-data-in

-ai-recommendation-and-decision-systems

17IMDA’sPETSandboxhelpstofacilitateexperimentationbasedonreal-worldusecases,includingusingPETsforAI.ThisenablesindustrytoexploreinnovativeusesofthisemergingtechnologywhileensuringPETsaredeployedinasafeandcompliantmanner.See

.sg/

how-we-can-help/data-innovation/privacy-enhancing-technology-sandboxes

18Thecopyrightissuehasgivenrisetovariedinterestsandconcernsamongstdifferentstakeholders,withpolicymakersstudyingtofindthebestwayforward.Copyrightownershaverequestedforrenumerationforuseoftheirworkstotrainmodels,concernedthatsuchsystemsmaycompetewiththemandimpacttheirlivelihood.Theyhaveadvocatedforlicensing-basedsolutionstofacilitatetextanddataminingactivitiesformachinelearning(ML),aswellasanopt-outsystemforcopyrightownersfromstatutoryexceptionsfortextanddatamining,andMLactivitiestoavoidundulyimpingingontheircommercialinterests.Othershavearguedthattextanddatamining,andMLdonotinfringecopyrightbecausetrainingdoesnotinvolvethecopyinganduseofthecreativeexpressioninworks.Therearealsopracticalconsiderationssurroundingobtainingconsentfromeverycopyrightowner,aswellastrade-offsinmodelperformance.

11

MODELAIGOVERNANCEFRAMEWORKFORGENERATIVEAI

GiventhelargevolumeofdatainvolvedinAItraining,thereisvalueindevelopingapproachestoresolvethesedifficultissuesinaclearandefficientmanner.Today,legalframeworkshavenotyetcoalescedaroundsuchanapproach.SomecopyrightownershaveinstitutedlawsuitsagainstgenerativeAIcompaniesintheUSandUKcourts.Variouscountriesarealsoexploringnon-legislativesolutionssuchascopyrightguidelines19andcodesofpracticefordevelopersandend-users.20

Giventhevariousinterestsatstake,policymakersshouldfosteropendialogueamongstallrelevantstakeholderstounderstandtheimpactofthefast-evolvinggenerativeAItechnology,andensurethatpotentialsolutionsarebalancedandinlinewithmarketrealities.

FacilitatingAccesstoQualityData

Asanoverallhygienemeasureatanorganisationallevel,itwouldbegooddisciplineforAIdeveloperstoundertakedataqualitycontrolmeasuresandadoptgeneralbestpracticesindatagovernance,includingannotatingtrainingdatasetsconsistentlyandaccurately,andusingdataanalysistoolstofacilitatedatacleaning(e.g.,debiasingandremovinginappropriatecontent).

Globally,itisworthconsideringaconcertedefforttoexpandtheavailablepooloftrusteddatasets.ReferencedatasetsareimportanttoolsinbothAImodeldevelopment(e.g.,forfine-tuning)aswellasbenchmarkingandevaluation.21Governmentscanalsoconsiderworkingwiththeirlocalcommunitiestocuratearepositoryofrepresentativetrainingdatasetsfortheirspecificcontext(e.g.,inlowresourcelanguages).Thishelpstoimprovetheavailabilityofqualitydatasetsthatreflecttheculturalandsocialdiversityofacountry,whichinturnsupportsthedevelopmentofsaferandmoreculturallyrepresentativemodels.

19JapanandtheRepublicofKoreahaveannouncedthedevelopmentofcopyrightguidelinestoaddressgenerativeAIissues,thoughtheyhavenotyetbeenissued.

20UKhasannouncedthatitisdevelopingavoluntarycodeofpracticebetweenend-usersandrightsholdersthroughaworkinggroupwithdiverseparticipationfromtechnology,creativeandresearchsectors.Thestatedaimsoftheworkinggrouparetomakelicensesfordataminingmoreavailable,tohelptoovercomebarriersthatAIfirmsandend-userscurrentlyface,andtoensurethereareprotectionsforrightsholders.

21Thisisakintoreferencestandardsin,forexample,thepharmaceuticalindustry,whichareusedasabasisforevaluationfordrugs.

TRUSTED

DEVELOPMENTAND

DEPLOYMENT

13

MODELAIGOVERNANCEFRAMEWORKFORGENERATIVEAI

FOSTERINGATRUSTEDAIECOSYSTEM

TRUSTEDDEVELOPMENT

ANDDEPLOYMENT

Modeldevelopment,andtheapplicationdeploymentontopofit,areatthecoreofAI-driveninnovation.Today,however,thereisalackofinformationontheapproachesbeingtakentoensuretrustworthymodels.Evenincasesof“open-source”models,someimportantinformationlikethemethodologyanddatasetsmaynotbemadeavailable.

Goingforward,itisimportantthattheindustrycoalescesaroundbestpracticesindevelopmentandsafetyevaluation.Thereafter,meaningfultransparencyaroundbaselinesafetyandhygienemeasuresundertakenwillalsobekey.ThiswillenablesafermodelusebyallstakeholdersintheAIecosystem.Suchtransparencywillneedtobebalancedwithlegitimateconsiderationssuchassafeguardingbusinessandproprietaryinformation,andnotallowingbadactorstogamethesystem.

Design

SafetybestpracticesneedtobeimplementedbymodeldevelopersandapplicationdeployersacrosstheAIdevelopmentlifecycle,arounddevelopment,disclosureandevaluation.Groundworkforthishasbeenlaidinthe2020versionoftheModelAIGovernanceFramework,whichsetsoutbestpracticesforthedevelopmentanddeploymentoftraditionalAIsolutions.22TheprinciplesarticulatedtherecontinuetoberelevantandareextendedhereforgenerativeAI.

Development—BaselineSafetyPractices

Safetymeasuresaredevelopingrapidly,andmodeldevelopersandapplicationdeployersarebestplacedtodeterminewhattouse.Evenso,industrypracticesarestartingtocoalescearoundsomecommonsafetypractices.

Forexample,afterpre-training,fine-tuningtechniquessuchasReinforcementLearningfromHumanFeedback(RLHF)23canguidethemodeltogeneratesaferoutputthatismorealignedwithhumanpreferencesandvalues.Acrucialstepforsafetyisalsotoconsiderthecontextoftheusecaseandconductariskassessment.Forexample,furtherfine-tuningorusinguserinteractiontechniques(suchasinputandoutputfilters)canhelptoreduceharmfuloutput.TechniqueslikeRetrieval-AugmentedGeneration(RAG)24andfew-shotlearningarealsocommonlyusedtoreducehallucinationsandimproveaccuracy.

Disclosure—“FoodLabels”

Transparencyaroundthesesafetymeasuresundertaken,thatformthecoreoftheAImodel’smake-upisthenkey.Thisisakinto“foodoringredientlabels”.Byprovidingrelevantinformationtodownstreamusers,theycanmakemoreinformeddecisions.Whileleadingmodeldevelopersalreadydisclosesomeinformation,

22See

.sg/-/media/Files/PDPC/PDF-Files/Resource-for-Organisation/AI/SGModelAIGovFramework2.pdf

23RLHFisatechniqueusedtoimproveLLMsbyusinghumanfeedbacktotrainapreferencemodel,thatinturnstrainstheLLMusingreinforcementlearning.RLHFcanbecomplementedwithmechanismstoassessconfidenceduringcontentgenerationtoalertmodeldevelopersorapplicationdeployerstoriskswherehumanverificationandvalidationisrequired.

24RAGisatechniquethathelpsmodelsprovidemorecontextuallyappropriateandcurrentresponsesthatarespecifictoanorganisationorindustry.ThisisdonebylinkinggenerativeAIservicestoexternalresources,therebygivingmodelssourcestociteandenhancingtheaccuracyandreliabilityofgenerativeAImodelswithfactsfetchedfromtrustedsources.

14

MODELAIGOVERNANCEFRAMEWORKFORGENERATIVEAI

standardisingdisclosurewillfacilitatecomparabilityacrossmodelsandpromotesafermodeluse.Relevanta