東盟人工智能治理與倫理指南-東南亞國家聯(lián)盟（英）

ASEANGuideon

AIGovernanceandEthics

Contents

ExecutiveSummary

03

D.National-levelRecommendations

A.Introduction

06

E.Regional-levelRecommendations

1.

Objectives

08

F.Conclusion

2.

Assumptions

08

3.

TargetAudience

09

AnnexA:AIRiskImpactAssessmentTemplate

4.

Definitions

09

AnnexB:UseCases

B.GuidingPrinciplesfortheFramework

10

1.

AboitizGroup

1.

TransparencyandExplainability

11

2.

EY

2.

FairnessandEquity

12

3.

Gojek

3.

SecurityandSafety

13

4.

UCARE.AI

4.

Human-centricity

14

5.

SmartNationGroup(SNG),Singapore

5.

PrivacyandDataGovernance

14

6.

MinistryofEducation,Singapore

6.

AccountabilityandIntegrity

15

7.

RobustnessandReliability

16

C.AIGovernanceFramework

17

1.

Internalgovernancestructuresandmeasures

18

2.

Determiningthelevelofhumaninvolvement

inAI-augmenteddecision-making

23

3.

Operationsmanagement

27

4.

Stakeholderinteractionandcommunication

41

46

53

58

60

73

74

76

78

80

83

85

ASEANGuideonAIGovernanceandEthics

ExecutiveSummary

WhattheGuideisabout

Thisdocumentservesasapracticalguidefororganisationsintheregionthatwishtodesign,develop,anddeploytraditionalAItechnologiesincommercialandnon-militaryordual-useapplications.ThisGuidefocusesonencouragingalignmentwithinASEANandfosteringtheinteroperabilityofAIframeworksacrossjurisdictions.Italsoincludesrecommendationsonnational-levelandregional-levelinitiativesthatgovernmentsintheregioncanconsiderimplementingtodesign,develop,anddeployAIsystemsresponsibly.

GuidingPrinciplesfortheFramework

Transparency

Fairnessand

Securityand

Robustnessand

and

Equity

Safety

Reliability

Explainability

Human-centricity

PrivacyandData

Accountability

Governance

andIntegrity

4KeyComponents

Internalgovernancestructuresandmeasures

Multi-disciplinary,centralgoverningbody,suchasanAIEthicsAdvisoryBoard,tooverseeAIgovernanceefforts

Developstandards,guidelines,tools,andtemplatestohelporganisationsdesign,develop,anddeployAIresponsibly

Clearlylayouttherolesandresponsibilitiesofpersonnelinvolvedintheresponsibledesign,developmentand/ordeploymentofAI

DeterminingthelevelofhumaninvolvementinAI-augmenteddecision-making

Conductrelevantriskimpactassessmentstodeterminelevelofrisk

Threebroadcategoriesofhumaninvolvementbasedonlevelofrisk–human-in-the-loop,human-over-the-loop,human-out-of-the-loop

MitigatingriskshelpsbuildtrusttowardstheacceptanceandgreateruseofAItechnologiesintheregion

3

ASEANGuideonAIGovernanceandEthics

Operationsmanagement

TheAISystemLifecycleconsistsofvariousstagesandisoftenaniterativeprocess

Conductrisk-basedassessmentsbeforestartinganydatacollectionandprocessingormodelling

Mitigaterisksofunjustbiasduetoinsufficientlyrepresentativetraining,testingandvalidationdatasets

Stakeholderinteractionandcommunication

Developtrustwithstakeholdersthroughoutthedesign,development,anddeploymentof

AI

ProvidegeneraldisclosureofwhenAIisusedinproductsand/orserviceofferings

PutinplacemeasurestohelpemployeesadapttoanAI-augmentedworkenvironment

National-levelRecommendations

NurturingAItalentandupskillingworkforce

Workcloselywithpublicandprivatesectorstoensurethatacountry’sworkforcecanadapttothenewwaysofworkingandpossessesenoughdigitalskillstointeracteffectivelywithAIsystems.

SupportingAIinnovationecosystemandpromotinginvestmentinAIstart-ups

WorkcloselywithpublicandprivatesectorstocreateasupportiveenvironmentforAIdevelopment,wherecompaniesareabletoaccessandleveragedata,digitaltechnologies,andinfrastructure.

InvestinginAIresearchanddevelopment

KeepabreastofthelatestdevelopmentsinAIandencourageresearchrelatedtothecybersecurityofAI,AIgovernance,andAIethicstoensurethatthesafetyandresiliencyofAIsystemsandtoolsalsoadvanceinparallelwithnewusecases.

PromotingadoptionofusefultoolsbybusinessestoimplementtheASEANGuideonAIGovernanceandEthics

DeploytoolstoenabletheimplementationofAIgovernanceinoperationsandensurethatdocumentationandvalidationprocessesaremoreefficient.

RaisingawarenessamongcitizensontheeffectsofAIinsociety

RaiseawarenessofthepotentialrisksandbenefitsofAIsocitizenscanmakeinformeddecisionsabouttheappropriateuseofAIandtakeappropriateactionstoprotectthemselvesfromharmfulusesofAIsystems.

4

ASEANGuideonAIGovernanceandEthics

Regional-levelrecommendations

SettingupanASEANWorkingGrouponAIGovernancetodriveandoverseeAIgovernanceinitiativesintheregion

TheWorkingGroupcanconsistofrepresentativesfromeachoftheASEANmemberstateswhocanworktogethertorollouttherecommendationslaidoutinthisGuide,aswellasprovideguidanceforASEANcountrieswhowishtoadoptcomponentsofthisGuide,andwhereappropriate,includeconsultationwithotherindustrypartnersfortheirviewsandinput.

AdaptationofthisGuidetoaddressgovernanceofgenerativeAI

Risksinclude:

Mistakesandanthropomorphism

Factuallyinaccurateresponsesanddisinformation

Deepfakes,impersonation,fraudulentandmaliciousactivities

Infringementofintellectualpropertyrights

Privacyandconfidentiality

Propagationofembeddedbiases

Governanceshouldinclude:

Adaptationofexistingframeworksandtools

Guidanceondevelopingasharedresponsibilityframework

GuidanceonincreasingthecapacitytomanagerisksofgenerativeAI

GuidanceonhowtodistinguishAI-generatedcontentversusauthenticallygeneratedones

CompilingacompendiumofusecasesdemonstratingpracticalimplementationoftheGuidebyorganisationsoperatinginASEAN

AcompendiumofusecasesshowcasesthecommitmentoftheseorganisationstoAIgovernanceandhelpsthempromotethemselvesasresponsibleAIpractitioners.

UseCases

IllustrationofcomponentsoftheASEANGuideonAIGovernanceandEthicsthroughusecasesoforganisationsoperatinginASEANthathaveimplementedAIgovernancemeasuresinAIdesign,development,anddeployment.

Gojek AboitizGroupSmartNationGroup(SNG),Singapore

UCARE.AI EYMinistryofEducation,Singapore

5

Introduction

ASEANGuideonAIGovernanceandEthics

A.Introduction

ArtificialIntelligence(AI)isthedisciplineofmakinganalyticalmachinesintelligent,enablinganorganisationtofunctionappropriatelyandwithforesight.Unlikeothertechnologies,someformsofAIadaptonitsown,learningthroughuse,sothedecisionsitmakestodaymaybedifferentfromthoseitmakestomorrow.AIandautomationhavebeenhottopics,bothfortheirtransformativepotentialandfortheircapacitytointroducenewopportunitiesbydisruptingoldmodels.SoutheastAsiaisnoexception.AIsystemsshouldbetreateddifferentlyfromothersoftwaresystemsbecauseofitsuniquecharacteristicsandrisks.CapabilitiesofAIsystemsfuelledbytechniquesevolutionandbreakthroughsarequicklyoutpacingthemonitoringandvalidationtools.ThedevelopmentofAIisalsodecentralisedduetolowbarrierstoentryandtheproliferationofopen-sourcetechnologies.GiventheprofoundimpactthatAIpotentiallybringstoorganisationsandindividualsinASEAN,itisimportantthatthedecisionsmadebyAIarealignedwithnationalandcorporatevalues,aswellasbroaderethicalandsocialnorms.

Also,theASEANDigitalMasterplan2025(ADM2025)developedbyASEANMemberStates(“AMS”)envisionsASEANasaleadingdigitalcommunityandeconomicbloc,poweredbysecureandtransformativedigitalservices,technologies,andecosystem.Inthatcontext,theADM2025hasidentifiedEnablingAction(EA)2.7thatsuggeststhedevelopmentandadoptionofaregionalpolicytodeliverbestpracticeguidanceonAIgovernanceandethics.Inrecentyears,governmentsandinternationalorganisationshavebegunissuingprinciples,frameworksandrecommendationsonAIethicsandgovernance.ExamplesincludeSingapore’sModelAIGovernanceFramework1andOECD’sRecommendationoftheCouncilonAI2.However,therehasnotyetbeenanintergovernmentalcommonstandardforAIthatdefinestheprinciplesofAIgovernanceandprovidesguidanceforpolicymakersintheregiontoutiliseAIsystemsinaresponsibleandethicalmanner.IntheprocessofdraftingthisGuide,existingAIgovernanceframeworksandguidelinessuchasUNESCO’sRecommendationontheEthicsofArtificialIntelligence3andEU’sEthicsGuidelinesforTrustworthyAIwerereferenced4.

TheASEANGuideonAIGovernanceandEthicsaimstoempowerorganisationsandgovernmentsintheregiontodesign,develop,anddeploytraditionalAIsystemsresponsiblyandincreaseusers’trustinAI.

InfocommMediaDevelopmentAuthority,“ModelArtificialIntelligenceGovernanceFrameworkSecondEdition”(21January2020)

<

.sg/-/media/files/pdpc/pdf-files/resource-for-organisation/ai/sgmodelaigovframework2.pdf

>

OrganisationforEconomicCo-operationandDevelopment,“RecommendationoftheCouncilonArtificialIntelligence”(22May2019)<

/en/instruments/OECD-LEGAL-0449

>

UnitedNationsEducational,ScientificandCulturalOrganisation,“RecommendationontheEthicsofArtificialIntelligence”(23November2021)

/ark:/48223/pf0000381137

>

EuropeanCommission“EthicsGuidelinesforTrustworthyAI”<

https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai

>(8April2019)

7

ASEANGuideonAIGovernanceandEthics

1.Objectives

Thisdocumentservesasapracticalguidefororganisationsintheregionthatwishtodesign,develop,anddeploytraditionalAItechnologiesincommercialandnon-militaryordual-useapplications.ThisGuidefocusesonencouragingalignmentwithinASEANandfosteringtheinteroperabilityofAIframeworksacrossjurisdictions.Italsoincludesrecommendationsonnational-levelandregional-levelinitiativesthatgovernmentsintheregioncanconsiderimplementingtodesign,develop,anddeployAIsystemsresponsibly.

TheASEANGuideonAIGovernanceandEthicsencompassesabroadrangeofconsiderationsthatneedstobetailoredtotheorganisationsdesigning,developing,anddeployingAIdependingonthenatureoftheindustry,complexityofthetechnology,andassociatedrisksoftheAIsystems.LocalgovernmentauthoritiesareencouragedtorefertothisGuidewhenformulatingtheirpoliciesandapproaches.

Whileadoptionoftheframeworklaidouthereinisvoluntary,thisGuidecanhelporganisationsbuildtrustamongstakeholdersandthepublicaswellasaligntheirAIpracticeswithinternationalstandardsandbestpractices,amongothers.OrganisationsareencouragedtorefertotheguidelinesinthisdocumenttounderstandhowtoassesstherisksassociatedwithAIandtakemeasurestodesign,develop,anddeployAIresponsiblyinthecontextoftheirorganisations.

ThisGuideismeanttobealivingdocumentthatshouldbeperiodicallyreviewedandassessedbyrelevantASEANsectorialbodies,inconsultationwithindustrypartners,toensurethatitisuptodatewiththelatestregulationsandadvancementsintheAIspace.UpdatestothisGuidemaybepublishedsubsequentlytokeepupwithevolutionsandgrowthingovernanceandstandards.

2.Assumptions

AIsystemsneedtobemanagedholistically,includingitsecosystemandallcomponents–humanoperator,InternetofThings(IoT),robotics,traditionaltechnology,vendors,etc.InadditiontotherecommendationssetoutinthisGuidepertainingtoAIgovernance,organisationsarealsoencouragedtofollowandrefertointernationalstandardsandbestpracticesinrelatedfieldslikeinformationsecuritymanagementsystems(ISMS),datamanagementandgovernance,softwaredevelopmentandtesting,cybersecurity,IoT,etc.

Developersanddeployersneedtoadheretoapplicablenationallawsandregulations,includingsector-specificlawsandconstitutionswhendesigning,developing,anddeployingAItechnologies.TheASEANGuideonAIGovernanceandEthicsdoesnotreplaceorsupersedeanyexistingorupcominglawsandonlyservesasaguideforresponsibledesign,development,anddeploymentofAIintheregion.BeforedeployingAI,itisalsoimportantfordevelopersanddeployerstoconsidertherelevantlegalandregulatoryrequirementsoftherespectivecountrieswheretheAIsystemswillbedeployed,aswellastheusecontextforlegal,policy,andregulatoryconcerns.

Giventhefast-evolvingspaceofAI,developersanddeployersofAIsystemsshouldbemindfulofthelateststateoftheartvis-à-visgovernancetoolsandtechnologiesandconducttherelevantassessmentstoevaluatethefeasibilityandusefulnessofthesetoolsintheimplementationofAIgovernancepracticesinthedesign,development,anddeploymentofAIsystems.

8

ASEANGuideonAIGovernanceandEthics

3.TargetAudience

TheASEANGuideonAIGovernanceandEthicsisanASEAN-endorsedframeworkfororganisationsintheregiontorefertowhendesigning,developing,anddeployingtraditionalAItechnologies.Itprovidesguidelinesandrecommendationsforadiverserangeofindividualsandorganisationsalongtheentirevaluechain.TheseincludeAIdevelopersanddeployers,academicprofessionals,andeveryonethatisinterestedinutilisingorscalingupAIsystems.TheGuidealsoincludessectionsonnational-levelandregional-levelrecommendationsthataremoretargetedtowardspolicymakersinASEAN.

4.Definitions

ThedefinitionsofseveralkeytermsusedinthisGuidearesetoutbelow.

ASEAN:referstotheAssociationofSoutheastAsianNations.Itisapoliticalandeconomicunionof10memberstatesinSoutheastAsia,whichpromotesintergovernmentalcooperationandfacilitateseconomic,political,security,military,educational,andsocioculturalintegrationbetweenitsmembersandcountriesintheAsia-Pacific.

ArtificialIntelligence(AI):isanengineeredormachine-basedsystemthatcan,foragivensetofobjectives,generateoutputssuchaspredictions,recommendations,ordecisionsinfluencingrealorvirtualenvironments.

AIsystem:isamachine-basedsystemthatiscapableofinfluencingtheenvironmentbyproducinganoutput(predictions,recommendationsordecisions)foragivensetofobjectives.Itusesmachineand/orhuman-baseddataandinputsto(i)perceiverealand/orvirtualenvironments;(ii)abstracttheseperceptionsintomodelsthroughanalysisinanautomatedmanner(e.g.,withmachinelearning),ormanually;and(iii)usemodelinferencetoformulateoptionsforoutcomes.AIsystemsaredesignedtooperatewithvaryinglevelsofautonomy.

DeepLearning:isasubsetofmachinelearning.Itisverylooselybasedontheinformation-processingarchitectureofthebrain–albeitfarfrommatchingitsability–enablingsystemstofindwaysofre-representinginputdatathatfacilitatemakinghighlyaccuratepredictions.

Deployer:isanentitythatusesorimplementsanAIsystem,whichcouldeitherbedevelopedbytheirin-houseteamorviaathird-partydeveloper.

Developer:isanentitythatdesigns,codes,orproducesanAIsystem.

MachineLearning:isasubfieldinAIwherealgorithmslearnbyidentifyingpatternsandcorrelationswithindatausingstatisticaltechniquestoenhanceperformance,allwithoutbeingexplicitlyprogrammed.

User:isanentityorperson(internalorexternal)thatinteractswithanAIsystemoranAI-enabledserviceandcanbeaffectedbyitsdecisions.

9

GuidingPrinciplesfortheFramework

ASEANGuideonAIGovernanceandEthics

B.GuidingPrinciplesfortheFramework

ThesevenguidingprinciplesbelowhelptoensuretrustinAIandthedesign,development,anddeploymentofethicalAIsystems.TheyalsoprovideguidanceonhowAIsystemsshouldbedesigned,developed,anddeployedinwayswhichconsiderthebroadersocietalimpact.

1.TransparencyandExplainability

TransparencyreferstoprovidingdisclosureonwhenanAIsystemisbeingusedandtheinvolvementofanAIsystemindecision-making,whatkindofdataituses,anditspurpose.BydisclosingtoindividualsthatAIisusedinthesystem,individualswillbecomeawareandcanmakeaninformedchoiceofwhethertousetheAI-enabledsystem.

ExplainabilityistheabilitytocommunicatethereasoningbehindanAIsystem’sdecisioninawaythatisunderstandabletoarangeofpeople,asitisnotalwaysclearhowanAIsystemhasarrivedataconclusion.ThisallowsindividualstoknowthefactorscontributingtotheAIsystem’srecommendation.

InordertobuildpublictrustinAI,itisimportanttoensurethatusersareawareoftheuseofAItechnologyandunderstandhowinformationfromtheirinteractionisusedandhowtheAIsystemmakesitsdecisionsusingtheinformationprovided.

Inlinewiththeprincipleoftransparency,deployershavearesponsibilitytoclearlydisclosetheimplementationofanAIsystemtostakeholdersandfostergeneralawarenessoftheAIsystembeingused.WiththeincreasinguseofAIinmanybusinessesandindustries,thepublicisbecomingmoreawareandinterestedinknowingwhentheyareinteractingwithAIsystems.KnowingwhenandhowAIsystemsinteractwithusersisalsoimportantinhelpingusersdiscernthepotentialharmofinteractingwithanAIsystemthatisnotbehavingasintended.Inthepast,AIalgorithmshavebeenfoundtodiscriminateagainstfemalejobapplicantsandhavefailedtoaccuratelyrecognisethefacesofdark-skinnedwomen.ItisimportantforuserstobeawareoftheexpectedbehaviouroftheAIsystemssotheycanmakemoreinformeddecisionsaboutthepotentialharmofinteractingwithAIsystems.AnexampleoftransparencyinanAI-enabledecommerceplatformisinformingusersthattheirpurchasehistoryisusedbytheplatform’srecommendationalgorithmtoidentifysimilarproductsanddisplaythemontheusers’feeds.

Inlinewiththeprincipleofexplainability,developersanddeployersdesigning,developing,anddeployingAIsystemsshouldalsostrivetofostergeneralunderstandingamongusersofhowsuchsystemsworkwithsimpleandeasytounderstandexplanationsonhowtheAIsystemmakesdecisions.UnderstandinghowAIsystemsworkwillhelphumansknowwhentotrustitsdecisions.Explanationscanhavevaryingdegreesofcomplexity,rangingfromasimpletextexplanationofwhichfactorsmoresignificantlyaffectedthedecision-makingprocesstodisplayingaheatmapovertherelevanttextorontheareaofanimagethatledtothesystem’sdecision.Forexample,whenanAIsystemisusedtopredictthelikelihoodofcardiacarrestinpatients,explainabilitycanbeimplementedbyinformingmedicalprofessionalsofthemostsignificantfactors(e.g.,age,bloodpressure,etc.)thatinfluencedtheAIsystem’sdecisionsothattheycansubsequentlymakeinformeddecisionsontheirown.

Where“blackbox”modelsaredeployed,renderingitdifficult,ifnotimpossibletoprovideexplanationsastotheworkingsoftheAIsystem,outcome-basedexplanations,withafocusonexplainingtheimpactofdecision-makingorresultsflowingfromtheAIsystemmaybereliedon.

11

ASEANGuideonAIGovernanceandEthics

Alternatively,deployersmayalsoconsiderfocusingonaspectsrelatingtothequalityoftheAIsystemorpreparinginformationthatcouldbuilduserconfidenceintheoutcomesofanAIsystem’sprocessingbehaviour.Someofthesemeasuresare:

DocumentingtherepeatabilityofresultsproducedbytheAIsystem.SomepracticestodemonstraterepeatabilityincludeconductingrepeatabilityassessmentstoensuredeploymentsinliveenvironmentsarerepeatableandperformingcounterfactualfairnesstestingtoensurethattheAIsystem’sdecisionsarethesameinboththerealworldandinthecounterfactualworld.Repeatabilityreferstotheabilityofthesystemtoconsistentlyobtainthesameresults,giventhesamescenario.Repeatabilityoftenapplieswithinthesameenvironment,withthesamedataandthesamecomputationalconditions.

EnsuringtraceabilitybybuildinganaudittrailtodocumenttheAIsystemdevelopmentanddecision-makingprocess,implementingablackboxrecorderthatcapturesallinputdatastreams,orstoringdataappropriatelytoavoiddegradationandalteration.

Facilitatingauditabilitybykeepingacomprehensiverecordofdataprovenance,procurement,pre-processing,lineage,storage,andsecurity.Suchinformationcanalsobecentraliseddigitallyinaprocesslogtoincreasecapacitytocaterthepresentationofresultstodifferenttiersofstakeholderswithdifferentinterestsandlevelsofexpertise.Deployersshould,however,notethatauditabilitydoesnotnecessarilyentailmakingcertainconfidentialinformationaboutbusinessmodelsorintellectualpropertyrelatedtotheAIsystempubliclyavailable.Arisk-basedapproachcanbetakentowardsidentifyingthesubsetofAI-enabledfeaturesintheAIsystemforwhichimplementedauditabilityisnecessarytoalignwithregulatoryrequirementsorindustrypractices.

UsingAIModelCards,whichareshortdocumentsaccompanyingtrainedmachinelearningmodelsthatdisclosethecontextinwhichmodelsareintendedtobeused,detailsoftheperformanceevaluationprocedures,andotherrelevantinformation.

IncaseswhereAIsystemsareprocureddirectlyfromdevelopers,deployerswillhavetoworktogetherwiththesedeveloperstoachievetransparency.MoreonthiswillbecoveredinlatersectionsoftheGuide.

2.FairnessandEquity

Deployersshouldhavesafeguardsinplacetoensurethatalgorithmicdecisionsdonotfurtherexacerbateoramplifyexistingdiscriminatoryorunjustimpactsacrossdifferentdemographicsandthedesign,development,anddeploymentofAIsystemsshouldnotresultinunfairbiasnessordiscrimination.Anexampleofsuchsafeguardswouldincludehumaninterventionsandchecksonthealgorithmsanditsoutputs.DeployersofAIsystemsshouldconductregulartestingofsuchsystemstoconfirmifthereisbiasandwherebiasisconfirmed,makethenecessaryadjustmentstorectifyimbalancestoensureequity.

WiththerapiddevelopmentsintheAIspace,AIsystemsareincreasinglyusedtoaiddecision-making.Forexample,AIsystemsarecurrentlyusedtoscreenresumesinjobapplicationprocesses,predictthecreditworthinessofconsumersandprovideagronomicadvicetofarmers.Ifnotproperlymanaged,anAIsystem’soutputsusedtomakedecisionswithsignificantimpactonindividualscouldperpetuateexistingdiscriminatoryorunjustimpactstospecificdemographics.Tomitigatediscrimination,itisimportantthatthedesign,development,anddeploymentofAIsystemsalignwithfairnessandequityprinciples.Inaddition,thedatasetsusedtotraintheAIsystemsshouldbediverseandrepresentative.Appropriatemeasuresshouldbetakentomitigatepotentialbiasesduringdatacollectionandpre-processing,training,andinference.Forexample,the

12

ASEANGuideonAIGovernanceandEthics

trainingandtestdatasetforanAIsystemusedintheeducationsectorshouldbeadequatelyrepresentativeofthestudentpopulationbyincludingstudentsofdifferentgendersandethnicities.

3.SecurityandSafety

AIsystemsshouldbesafeandsufficientlysecureagainstmaliciousattacks.

Safetyreferstoensuringthesafetyofdevelopers,deployers,andusersofAIsystemsbyconductingimpactorriskassessmentsandensuringthatknownriskshavebeenidentifiedandmitigated.Ariskpreventionapproachshouldbeadopted,andprecautionsshouldbeputinplacesothathumanscanintervenetopreventharm,orthesystemcansafelydisengageitselfintheeventanAIsystemmakesunsafedecisions-autonomousvehiclesthatcauseinjurytopedestriansareanillustrationofthis.EnsuringthatAIsystemsaresafeisessentialtofosteringpublictrustinAI.SafetyofthepublicandtheusersofAIsystemsshouldbeofutmostpriorityinthedecision-makingprocessofAIsystemsandrisksshouldbeassessedandmitigatedtothebestextentpossible.BeforedeployingAIsystems,deployersshouldconductriskassessmentsandrelevanttestingorcertificationandimplementtheappropriatelevelofhumaninterventiontopreventharmwhenunsafedecisionstakeplace.Therisks,limitations,andsafeguardsoftheuseofAIshouldbemadeknowntotheuser.Forexample,inAI-enabledautonomousvehicles,developersanddeployersshouldputinplacemechanismsforthehumandrivertoeasilyresumemanualdrivingwhenevertheywish.

SecurityreferstoensuringthecybersecurityofAIsystems,whichincludesmechanismsagainstmaliciousattacksspecifictoAIsuchasdatapoisoning,modelinversion,thetamperingofdatasets,byzantineattacksinfederatedlearning5,aswellasotherattacksdesignedtoreverseengineerpersonaldatausedtotraintheAI.DeployersofAIsystemsshouldworkwithdeveloperstoputinplacetechnicalsecuritymeasureslikerobustauthenticationmechanismsandencryption.Justlikeanyothersoftware,deployersshouldalsoimplementsafeguardstoprotectAIsystemsagainstcyberattacks,datasecurityattacks,andotherdigitalsecurityrisks.ThesemayincludeensuringregularsoftwareupdatestoAIsystemsandproperaccessmanagementforcriticalorsensitivesystems.DeployersshouldalsodevelopincidentresponseplanstosafeguardAIsystemsfromtheaboveattacks.

Itisalsoimportantfordeployerstomakeaminimumlistofsecuritytesting(e.g.vulnerabilityassessmentandpenetrationtesting)andotherapplicablesecuritytestingtools.Someotherimportantconsiderationsalsoinclude:

Businesscontinuityplan

Disasterrecoveryplan

Zero-dayattacks

IoTdevices

Abyzantineattackinfederatedlearningisatypeofmaliciousactwhereoneormoreofthedevicesorserversinvolvedinthefederatedlearningprocessbehaveserraticallyorprovidesmisleadingupdatestothecentralmodel,withtheintenttocorruptormanipulatethelearningprocessoroutcomes.

13

ASEANGuideonAIGovernanceandEthics

4.Human-centricity

AIsystemsshouldrespecthuman-centredvaluesandpursuebenefitsforhumansociety,includinghumanbeings’well-being,nutrition,happiness,etc.

ItiskeytoensurethatpeoplebenefitfromAIdesign,development,anddeploym