2023人工智能信任全球洞察（英文）-KPMG

Trustin

Artificial

Intelligence

Aglobalstudy

2023

.au

KPMG.com.au

Citation

Gillespie,N.,Lockey,S.,Curtis,C.,Pool,J.,&

Akbari,A.(2023).TrustinArti?cialIntelligence:

AGlobalStudy.TheUniversityofQueensland

andKPMGAustralia.doi:

10.14264/00d3c94

UniversityofQueenslandResearchers

ProfessorNicoleGillespie,DrSteveLockey,

DrCaitlinCurtisandDrJavadPool.

TheUniversityofQueenslandteamledthe

design,conduct,analysisandreportingof

thisresearch.

KPMGAdvisors

JamesMabbott,RitaFentenervanVlissingen,

JessicaWyndham,andRichardBoele.

Acknowledgements

Wearegratefulfortheinsightfulinput,expertise

andfeedbackonthisresearchprovidedby

DrAliAkbari,DrIanOpperman,RossanaBianchi,

ProfessorShaziaSadiq,MikeRichmond,and

DrMortezaNamvar,andmembersofthe

Trust,EthicsandGovernanceAllianceatThe

UniversityofQueensland,particularlyDrNatalie

Smith,AssociateProfessorMartinEdwards,

DrShannonColvilleandAlexMacdade.

Funding

ThisresearchwassupportedbyanAustralian

GovernmentResearchSupportPackagegrant

providedtoTheUniversityofQueenslandAI

Collaboratory,andbytheKPMGChairinTrust

grant(ID2018001776).

AcknowledgementofCountry

TheUniversityofQueensland(UQ)

acknowledgestheTraditionalOwnersandtheir

custodianshipofthelands.Wepayourrespects

totheirAncestorsandtheirdescendants,who

continueculturalandspiritualconnections

toCountry.Werecognisetheirvaluable

contributionstoAustralianandglobalsociety.

?2023TheUniversityofQueenslandABN:63942912684CRICOSProviderNo:00025B.

?2023KPMG,anAustralianpartnershipandamember?rmoftheKPMGglobalorganisationofindependentmember?rmsaf?liatedwithKPMGInternationalLimited,aprivateEnglishcompanylimitedbyguarantee.Allrightsreserved.TheKPMGnameandlogoaretrademarksusedunderlicensebytheindependentmember?rmsoftheKPMGglobalorganisation.

LiabilitylimitedbyaschemeapprovedunderProfessionalStandardsLegislation.

Contents

Executivesummary02

Introduction07

Howweconductedtheresearch08

1.TowhatextentdopeopletrustAIsystems?11

2.Howdopeopleperceivethebene?tsandrisksofAI?22

3.Whoistrustedtodevelop,useandgovernAI?29

4.Whatdopeopleexpectofthemanagement,governance

andregulationofAI?34

5.HowdopeoplefeelaboutAIatwork?43

6.HowwelldopeopleunderstandAI?53

7.WhatarethekeydriversoftrustinandacceptanceofAI?60

8.HowhavetrustandattitudestowardsAIchangedovertime?66

Conclusionandimplications70

Appendix1:Methodandstatisticalnotes73

Appendix2:Countrysamples75

Appendix3:Keyindicatorsforeachcountry7

7

?2023TheUniversityofQueenslandABN:63942912684CRICOSProviderNo:00025B.

?2023KPMG,anAustralianpartnershipandamember?rmoftheKPMGglobalorganisationofindependentmember?rmsaf?liatedwithKPMGInternationalLimited,aprivateEnglishcompanylimitedbyguarantee.Allrightsreserved.TheKPMGnameandlogoaretrademarksusedunderlicensebytheindependentmember?rmsoftheKPMGglobalorganisation.

LiabilitylimitedbyaschemeapprovedunderProfessionalStandardsLegislation.

TRUSTINARTIFICIALINTELLIGENCE2

Executive

summary

Arti?cialIntelligence(AI)hasbecomeaubiquitouspartofeverydaylifeandwork.AIisenablingrapidinnovationthatistransformingthewayworkisdoneandhowservicesaredelivered.Forexample,generativeAItoolssuchasChatGPTarehavingaprofoundimpact.Giventhemanypotentialandrealisedbene?tsforpeople,organisationsandsociety,investmentinAIcontinuestogrowacrossallsectors

1

,withorganisationsleveragingAIcapabilitiestoimprovepredictions,optimiseproductsandservices,augmentinnovation,enhanceproductivityandef?ciency,andlowercosts,amongstotherbene?cialapplications.

However,theuseofAIalsoposesrisksandchallenges,raisingconcernsaboutwhetherAIsystems(inclusiveofdata,algorithmsandapplications)areworthyoftrust.Theseconcernshavebeenfuelledbyhighpro?lecasesofAIusethatwerebiased,discriminatory,manipulative,unlawful,orviolatedhumanrights.Realisingthebene?tsAIoffersandthereturnoninvestmentinthesetechnologiesrequiresmaintainingthepublic’strust:peopleneedtobecon?dentAIisbeingdevelopedandusedinaresponsibleandtrustworthymanner.SustainedacceptanceandadoptionofAIinsocietyarefoundedonthistrust.

Thisresearchisthe?rsttotakeadeepdiveexaminationintothepublic’strustandattitudestowardstheuseofAI,andexpectationsofthemanagementandgovernanceofAIacrosstheglobe.

Wesurveyedover17,000peoplefrom17countriescoveringallglobalregions:Australia,Brazil,Canada,China,Estonia,Finland,France,Germany,India,Israel,Japan,theNetherlands,Singapore,SouthAfrica,SouthKorea,theUnitedKingdom(UK),andtheUnitedStatesofAmerica(USA).ThesecountriesareleadersinAIactivityandreadinesswithintheirregion.Eachcountrysampleisnationallyrepresentativeofthepopulationbasedonage,gender,andregionaldistribution.

WeaskedsurveyrespondentsabouttrustandattitudestowardsAIsystemsingeneral,aswellasAIuseinthecontextoffourapplicationdomainswhereAIisrapidlybeingdeployedandlikelytoimpactmanypeople:inhealthcare,publicsafetyandsecurity,humanresourcesandconsumerrecommenderapplications.

Theresearchprovidescomprehensive,timely,globalinsightsintothepublic’strustandacceptanceofAIsystems,includingwhoistrustedtodevelop,useandgovernAI,theperceivedbene?tsandrisksofAIuse,communityexpectationsofthedevelopment,regulationandgovernanceofAI,andhoworganisationscansupporttrustintheirAIuse.ItalsoshedslightonhowpeoplefeelabouttheuseofAIatwork,currentunderstandingandawarenessofAI,andthekeydriversoftrustinAIsystems.WealsoexplorechangesintrustandattitudestoAIovertime.

Next,wesummarisethekey?ndings.

?2023TheUniversityofQueenslandABN:63942912684CRICOSProviderNo:00025B.

?2023KPMG,anAustralianpartnershipandamember?rmoftheKPMGglobalorganisationofindependentmember?rmsaf?liatedwithKPMGInternationalLimited,aprivateEnglishcompanylimitedbyguarantee.Allrightsreserved.TheKPMGnameandlogoaretrademarksusedunderlicensebytheindependentmember?rmsoftheKPMGglobalorganisation.

LiabilitylimitedbyaschemeapprovedunderProfessionalStandardsLegislation.

TRUSTINARTIFICIALINTELLIGENCE3

MostpeoplearewaryabouttrustingAIsystemsandhavelowormoderateacceptanceofAI:however,trustandacceptancedependontheAIapplication

Acrosscountries,threeoutof?vepeople(61%)arewaryabouttrustingAIsystems,reportingeitherambivalenceoranunwillingnesstotrust.TrustisparticularlylowinFinlandandJapan,wherelessthanaquarterofpeoplereporttrustingAI.Incontrast,peopleintheemergingeconomiesofBrazil,India,ChinaandSouthAfrica(BICS

2

)havethehighestlevelsoftrust,withthemajorityofpeopletrustingAIsystems.

PeoplehavemorefaithinAIsystemstoproduceaccurateandreliableoutputandprovidehelpfulservices,andaremorescepticalaboutthesafety,securityandfairnessofAIsystemsandtheextenttowhichtheyupholdprivacyrights.

TrustinAIsystemsiscontextualanddependsonthespeci?capplicationorusecase.Oftheapplicationsweexamined,peoplearegenerallylesstrustingandacceptingofAIuseinhumanresources(i.e.foraidinghiringandpromotiondecisions),andmoretrustingofAIuseinhealthcare(i.e.foraidingmedicaldiagnosisandtreatment)wherethereisadirectbene?ttothem.Peoplearegenerallymorewillingtorelyon,thanshareinformationwithAIsystems,particularlyrecommendersystems(i.e.forpersonalisingnews,socialmedia,andproductrecommendations)andsecurityapplications(i.e.foraidingpublicsafetyandsecuritydecisions).

ManypeoplefeelambivalentabouttheuseofAI,reportingoptimismorexcitementontheonehand,whilesimultaneouslyreportingworryorfear.Overall,two-thirdsofpeoplefeeloptimisticabouttheuseofAI,whileabouthalffeelworried.Whileoptimismandexcitementaredominantemotionsinmanycountries,particularlytheBICScountries,fearandworryaredominantemotionsforpeopleinAustralia,Canada,France,andJapan,withpeopleinFrancethemostfearful,worried,andoutragedaboutAI.

Peoplerecognisethemanybene?tsofAI,butonlyhalfbelievethebene?tsoutweightherisks

People’swarinessandambivalencetowardsAIcanbepartlyexplainedbytheirmixedviewsofthebene?tsandrisks.Mostpeople(85%)believeAIresultsinarangeofbene?ts,andthinkthat‘process’bene?tssuchasimprovedef?ciency,innovation,effectiveness,resourceutilisationandreducedcosts,aregreaterthanthe‘people’bene?tsofenhancingdecision-makingandimprovingoutcomesforpeople.

However,onaverage,onlyoneintwopeoplebelievethebene?tsofAIoutweightherisks.PeopleinthewesterncountriesandJapanareparticularlyunconvincedthatthebene?tsoutweightherisks.Incontrast,themajorityofpeopleintheBICScountriesandSingaporebelievethebene?tsoutweightherisks.

PeopleperceivetherisksofAIinasimilarway

acrosscountries,withcybersecurityratedasthe

topriskglobally

WhiletherearedifferencesinhowtheAIbene?t-risk

ratioisviewed,thereisconsiderableconsistencyacross

countriesinthewaytherisksofAIareperceived.

Justunderthree-quarters(73%)ofpeopleacrosstheglobereportfeelingconcernedaboutthepotentialrisksofAI.Theserisksincludecybersecurityandprivacybreaches,manipulationandharmfuluse,lossofjobsanddeskilling,systemfailure,theerosionofhumanrights,andinaccurateorbiasedoutcomes.

Inallcountries,peopleratedcybersecurityrisksastheirtoponeortwoconcerns,andbiasasthelowestconcern.JoblossduetoautomationisalsoatopconcerninIndiaandSouthAfrica,andsystemfailureranksasatopconcerninJapan,potentiallyre?ectingtheirrelativeheavydependenceonsmarttechnology.

These?ndingsreinforcethecriticalimportanceofprotectingpeople’sdataandprivacytosecureandpreservetrust,andsupportingglobalapproachesandinternationalstandardsformanagingandmitigatingAIrisksacrosscountries.

ThereisstrongglobalendorsementfortheprinciplesoftrustworthyAI:trustiscontingentonupholdingandassuringtheseprinciplesareinplace

Our?ndingsrevealstrongglobalpublicsupportfortheprinciplesandrelatedpracticesorganisationsdeployingAIsystemsareexpectedtoupholdinordertobetrusted.EachoftheTrustworthyAIprinciplesoriginallyproposedbytheEuropeanCommission

3

areviewedashighlyimportantfortrustacrossall17countries,withdataprivacy,securityandgovernanceviewedasmostimportantinallcountries.ThisdemonstratesthatpeopleexpectorganisationsdeployingAIsystemstoupholdhighstandardsof:

–dataprivacy,securityandgovernance

–technicalperformance,accuracyandrobustness

–fairness,non-discriminationanddiversity

–humanagencyandoversight

–transparencyandexplainability

–accountabilityandcontestability

–riskandimpactmitigation

–AIliteracysupport

?2023TheUniversityofQueenslandABN:63942912684CRICOSProviderNo:00025B..

?2023KPMG,anAustralianpartnershipandamember?rmoftheKPMGglobalorganisationofindependentmember?rmsaf?liatedwithKPMGInternationalLimited,aprivateEnglishcompanylimitedbyguarantee.Allrightsreserved.TheKPMGnameandlogoaretrademarksusedunderlicensebytheindependentmember?rmsoftheKPMGglobalorganisation.

LiabilitylimitedbyaschemeapprovedunderProfessionalStandardsLegislation.

TRUSTINARTIFICIALINTELLIGENCE4

PeopleexpecttheseprinciplestobeinplaceforeachoftheAIuseapplicationsweexamined(e.g.,HumanResources,Healthcare,Security,Recommender,andAIsystemsingeneral),suggestingtheiruniversalapplication.ThisstrongpublicendorsementprovidesablueprintfordevelopingandusingAIinawaythatsupportstrustacrosstheglobe.

OrganisationscandirectlybuildtrustandconsumerwillingnesstouseAIsystemsbysupportingandimplementingassurancemechanismsthathelppeoplefeelcon?denttheseprinciplesarebeingupheld.ThreeoutoffourpeoplewouldbemorewillingtotrustanAIsystemwhenassurancemechanismsareinplacethatsignalethicalandresponsibleuse,suchasmonitoringsystemaccuracyandreliability,independentAIethicsreviews,AIethicscerti?cations,adheringtostandards,andAIcodesofconduct.Thesemechanismsareparticularlyimportantgiventhecurrentrelianceonindustryregulationandgovernanceinmanyjurisdictions.

Peoplearemostcon?dentinuniversitiesanddefenceorganisationstodevelop,useandgovernAIandleastcon?dentingovernmentandcommercialorganisations

Peoplehavethemostcon?denceintheirnationaluniversitiesandresearchinstitutions,aswellastheirdefenceorganisations,todevelop,useandgovernAIinthebestinterestofthepublic(76–82%con?dent).Incontrast,theyhavetheleastcon?denceingovernmentsandcommercialorganisationstodothis.Athirdofpeoplelackcon?denceingovernmentandcommercialorganisationstodevelop,useandregulateAI.ThisisproblematicgiventheincreasingscopewithwhichgovernmentsandcommercialorganisationsareusingAI,andthepublic’sexpectationthattheseentitieswillresponsiblygovernandregulateitsuse.AnimplicationisthatgovernmentandbusinesscanpartnerwithmoretrustedentitiesintheuseandgovernanceofAI.

Therearesigni?cantdifferencesacrosscountriesinpeople’strustoftheirgovernmenttouseandgovern

AI,withabouthalfofpeoplelackingcon?denceintheirgovernmentinSouthAfrica,Japan,theUKandtheUSA,whereasthemajorityinChina,IndiaandSingaporehavehighcon?denceintheirgovernment.Thispatternmirrorspeople’sgeneraltrustintheirgovernments:wefoundastrongassociationbetweenpeople’sgeneraltrustingovernment,commercialorganisationsandotherinstitutionsandtheircon?denceintheseentitiestouseandgovernAI.These?ndingssuggestthattakingactiontostrengthentrustininstitutionsgenerallyisanimportantfoundationfortrustinspeci?cAIactivities.

PeopleexpectAItoberegulatedwithsomeformofexternal,independentoversight,butviewcurrentregulationsandsafeguardsasinadequate

Thelargemajorityofpeople(71%)expectAItoberegulated.WiththeexceptionofIndia,themajorityinallothercountriesseeregulationasnecessary.This?ndingcorroboratespriorsurveys

4

indicatingstrongdesireforregulationofAIandisnotsurprisinggivenmostpeople(61%)believethelong-termimpactofAIonsocietyisuncertainandunpredictable.

Peoplearebroadlysupportiveofmultipleformsofregulation,includingregulationbygovernmentandexistingregulators,adedicatedindependentAIregulator,andco-regulationandindustryregulation,withgeneralagreementoftheneedforsomeformofexternal,independentoversight.

DespitethestrongexpectationsofAIregulation,onlytwoin?vepeoplebelievecurrentregulations,lawsandsafeguardsaresuf?cienttomakeAIusesafe.Thisalignswithprevioussurveys

5

showingpublicdissatisfactionwiththeregulationofAI,andisproblematicgiventhestrongrelationshipbetweencurrentsafeguardsandtrustinAIdemonstratedbyourmodelling.ThishighlightstheimportanceofstrengtheningandcommunicatingtheregulatoryandlegalframeworkgoverningAIanddataprivacy.

Thereare,however,substantialcountrydifferences,withpeopleinIndiaandChinamostlikelytobelieveappropriatesafeguardsareinplace(74–80%agree)followedbyBrazilandSingapore(52–53%).Incontrast,peopleinJapanandSouthKoreaaretheleastconvinced(13–17%agree)asarethemajorityofpeopleinwesterncountries.ThesedifferencesintheperceivedadequacyofregulationsmaypartlyexplainthehighertrustandacceptanceofAIamongpeopleintheBICScountries.

MostpeoplearecomfortablewiththeuseofAIto

augmentworkandinformmanagerialdecision-

making,butwanthumanstoretaincontrol

MostpeoplearecomfortablewiththeuseofAIatworktoaugmentandautomatetasks,butarelesscomfortablewhenAIisfocusedonthemasemployees,forexampleforHRandpeoplemanagement(e.g.tomonitorandevaluateemployees,andsupportrecruitment).Onaverage,halfthepeoplearewillingtotrustAIatwork,forexamplebyrelyingontheoutputitprovides.PeopleinAustralia,Canada,FranceandGermanyaretheleastcomfortablewiththeuseofAIatwork,whilethoseintheBICScountriesandSingaporearethemostcomfortable.

?2023TheUniversityofQueenslandABN:63942912684CRICOSProviderNo:00025B.

?2023KPMG,anAustralianpartnershipandamember?rmoftheKPMGglobalorganisationofindependentmember?rmsaf?liatedwithKPMGInternationalLimited,aprivateEnglishcompanylimitedbyguarantee.Allrightsreserved.TheKPMGnameandlogoaretrademarksusedunderlicensebytheindependentmember?rmsoftheKPMGglobalorganisation.

LiabilitylimitedbyaschemeapprovedunderProfessionalStandardsLegislation.

TRUSTINARTIFICIALINTELLIGENCE5

MostpeopleviewAIuseinmanagerialdecision-makingasacceptable,andactuallypreferAIinvolvementtosolehumandecision-making.However,thepreferredoptioniseithera25%-75%or50%-50%AI-humancollaboration,withhumansretainingmoreorequalcontrol.ThisindicatesaclearpreferenceforAItobeusedasadecisionaid,andalackofsupportforfullyautomatedAIdecision-makingatwork.

WhileabouthalfbelieveAIwillenhancetheircompetenceandautonomyatwork,lessthanoneinthreepeoplebelieveAIwillcreatemorejobsthanitwilleliminate.However,mostmanagersbelievetheopposite–thatAIwillcreatejobs.Thisre?ectsabroadertrendofmanagersbeingmorecomfortable,trustingandsupportiveofAIuseatworkthanotheremployees,withmanualworkerstheleastcomfortableandtrustingofAIatwork.GivenmanagersaretypicallythedriversofAIadoptioninorganisations,thesedifferingviewsmaycausetensionsintheimplementationofAIatwork.

Aminorityofpeopleinwesterncountries,JapanandSouthKoreareportthattheiremployingorganisationinvestsinAIadoption,recogniseseffortstointegrateAI,orsupportstheresponsibleuseofAI.ThisstandsincontrasttoamajorityofpeopleintheBICScountriesandSingapore.

PeoplewanttolearnmoreaboutAIbutcurrentlyhavelowunderstanding

While82%ofpeopleareawareofAI,oneintwopeoplereportfeelingtheydonotunderstandAIorwhenandhowitisused.UnderstandingofAIishighestinChina,India,SouthKorea,andSingapore.Twooutof?vepeopleareunawarethatAIenablescommonapplicationstheyuse.Forexample,eventhough87%ofpeopleusesocialmedia,45%donotknowAIisusedinsocialmedia.

PeoplewhobetterunderstandAIaremorelikelytotrustandacceptit,andperceivegreaterbene?tsofAIuse.ThissuggestsunderstandingAIsetsafoundationfortrust.Mostpeopleacrossallcountries(82%)wanttoknowmoreaboutAI.Consideredtogether,these?ndingssuggestastrongneedandappetiteforpubliceducationonAI.

Youngergenerations,theuniversityeducatedandmanagersaremoretrusting,acceptingandgenerallyholdmorepositiveattitudestowardsAI

Youngergenerations,theuniversityeducated,andmanagersshowaconsistentanddistinctlymorepositiveorientationtowardsAIacrossthe?ndings,comparedtooldergenerations,thosewithoutauniversityeducationandnon-managers.TheyaremoretrustingandacceptingofAIsystems,includingtheiruseatwork,andaremorelikelytofeelpositiveaboutAIandreportusingit.

TheyhavegreaterknowledgeofAIandarebetterabletoidentifywhenAIisused,andhavegreaterinterestinlearningaboutAI.Theyperceivemorebene?tsofAI,butremainthesameasothergroupsintheirperceptionsoftherisksofAI.TheyaremorelikelytobelieveAIwillcreatejobs,butalsomoreawarethatAIcanperformkeyaspectsoftheirwork.Theyaremorecon?dentinentitiestodevelop,useandgovernAI,andmorelikelytobelievethatcurrentsafeguardsaresuf?cienttomakeAIusesafe.ItisnoteworthythatweseeveryfewmeaningfuldifferencesacrossgenderintrustandattitudestowardsAI.

Therearestarkdifferencesintrustandattitudesacrosscountries:peopleintheemergingeconomiesofBrazil,India,China,andSouthAfricaaremoretrustingandacceptingofAIandhavemorepositiveattitudestowardsAI

Akeyinsightfromthesurveyisthestarkdifferencesintrust,attitudesanduseofAIbetweenpeopleintheemergingeconomiesofBrazil,India,ChinaandSouthAfricaandthoseinothercountries.

PeopleintheemergingeconomiesaremoretrustingandacceptingofAIandholdmorepositivefeelingsandattitudestowardsAIthanpeopleinothercountries.Thesedifferencesheldevenwhencontrollingfortheeffectsofageandeducation.Singaporefollowedthispositiveorientationonseveralindicators,particularlycomfort,trustandfamiliaritywiththeuseofAIatwork,adequacyofcurrentAIregulationandgovernance,con?denceincompaniestouseandgovernAI,andthebeliefthatAIwillcreatejobs.

Ourdatasuggeststhatthishightrustisnotblindtotherisks.PeopleinBICScountriesandSingaporedidnotperceivetherisksofAI,ortheuncertainimpactofAIonsociety,anylowerthanpeopleinothercountries.NordidtheydifferfromothercountriesontheimportanceoftheprinciplesandpracticesrequiredtoensureAIistrustworthy.Rather,akeydifferentiatoristhatmostpeopleintheBICScountriesandSingaporebelievethebene?tsofAIoutweightherisks,whereasaminorityofpeopleinwesterncountries,suchasAustralia,Canada,France,Netherlands,theUKandUSA,holdthisview.

ThehighertrustandmorepositiveattitudesintheBICScountriesislikelyduetothegreaterbene?tsaffordedbytechnologicaladvancesanddeploymentinemergingeconomiesandtheincreasinglyimportanteconomicroleofAItechnologiesinthesecountries.

?2023TheUniversityofQueenslandABN:63942912684CRICOSProviderNo:00025B..

?2023KPMG,anAustralianpartnershipandamember?rmoftheKPMGglobalorganisationofindependentmember?rmsaf?liatedwithKPMGInternationalLimited,aprivateEnglishcompanylimitedbyguarantee.Allrightsreserved.TheKPMGnameandlogoaretrademarksusedunderlicensebytheindependentmember?rmsoftheKPMGglobalorganisation.

LiabilitylimitedbyaschemeapprovedunderProfessionalStandardsLegislation.

TRUSTINARTIFICIALINTELLIGENCE6

Thismayencourageagrowthmindsetthatmotivatesacceptanceanduseoftechnologyasameanstoaccelerateeconomicprogress,prosperity,andqualityoflife.AnimplicationisthatthesecountriesmaybeuniquelypositionedtorapidlyaccelerateinnovationandtechnologicaladvantagethroughAI.Itisnotable,however,thatoninternationalrankingsthesecountriesranklowongovernanceandregulationframeworkstoensuretheethicalandresponsibleuseofAIcomparedtowesterncountries.

6

AIawareness,understandingandtrustinAIhasincreasedovertime,butinstitutionalsafeguardscontinuetolag

Wehadtheopportunitytoexaminehowtrustands