版權(quán)說明:本文檔由用戶提供并上傳,收益歸屬內(nèi)容提供方,若內(nèi)容存在侵權(quán),請(qǐng)進(jìn)行舉報(bào)或認(rèn)領(lǐng)
文檔簡介
Bewarethegap:
Governance
arrangementsinthefaceofAIinnovation
REPORT798|OCTOBER2024
Aboutthisreport
ASICreviewedhow23AFSlicenseesandcreditlicenseesareusingandplanningtousearti?cialintelligence,howtheyareidentifyingandmitigatingassociatedconsumerrisks,and
theirgovernancearrangements.Thisreportoutlinesthekey?ndingsfromthatreview.
AboutASICregulatorydocuments
InadministeringlegislationASICissuesthefollowingtypesofregulatorydocuments:consultationpapers,regulatoryguides,informationsheetsandreports.
Disclaimer
Thisreportdoesnotconstitutelegaladvice.Weencourageyoutoseekyourownprofessionaladviceto?ndouthowthe
CorporationsAct2001,NationalConsumerCreditProtectionAct2009andotherapplicablelawsapplytoyou,asitisyourresponsibilitytodetermineyourobligations.Examplesin
thisreportarepurelyforillustration;theyarenotexhaustiveandarenotintendedtoimposeorimplyparticularrulesorrequirements.Forprivacyreasons,thenamesofcase-studysubjectshavebeenchanged.
ASIC?REP798
FOREWORD3
EXECUTIVESUMMARY
4
AIGOVERNANCE:QUESTIONSFORLICENSEES
8
WHYLOOKATAI?
9
FINDINGS:USEOFAI
10
FINDINGS:RISKMANAGEMENTANDGOVERNANCE
18
WHERETOFROMHEREFORLICENSEES?
33
APPENDIX1:REVIEWMETHODOLOGYANDDEFINITIONS
38
APPENDIX2:ACCESSIBLEDATAPOINTS
40
APPENDIX3:KEYTERMS
41
CONTENTS
2
Foreword
Artificialintelligence(AI)istransformingmanyaspectsofourlives,
iEnlioipwictweglaabwitiaaiilmlplisaitn,snedrbicites.
ceauortentialbenefitstobusinessandindividualsareenormous–digitalinnovationsincludingAIareestimatedtocontribute
3m1,5tobiellniniatnoeaues.tiacliilalu’spaePmbeyaq0u3a0e1rerit,temquenonpelecatutreprehentiresenissitaellabo.Itatenihilluptatem
Tofullyrealisethosebenefits,wemustbalanceinnovationandprotection.Theintegrityofourfinancialsystem–andthesafetyoftheconsumerswhointeractwithit–reliesonusfindingtherightbalance.
Forsometime,ASIChasbeenremindinglicenseesthatexistingobligationsapplytotheiruseofAI.ASIChasalsobeenbuildinganunderstandingofhowAIisactuallybeingusedinthesectorsweregulate.
ThisreportisASIC’sfirstexaminationofthewaysAustralian
financialservices(AFS)andcreditlicenseesareimplementingAIwhereitimpactsconsumers.Concerningly,itfindsthatthereisthepotentialforagovernancegap.
Simplyput,somelicenseesareadoptingAImorerapidlythan
theirriskandgovernancearrangementsarebeingupdatedto
reflecttherisksandchallengesofAI.ThereisarealriskthatsuchgapswidenasAIuseacceleratesandthismagnifiesthepotentialforconsumerharm.
ASIC?REP798
WhiletheapproachtousingAIwhereitimpactsconsumershasmostlybeencautiousforlicensees,itisworryingthatcompetitivepressures
andbusinessneedsmayincentiviseindustrytoadoptmorecomplexandconsumer-facingAIfasterthantheyupdatetheirframeworkstoidentify,mitigateandmonitorthenewrisksandchallengesthisbrings.
AstheracetomaximisethebenefitsofAIintensifies,itiscriticalthat
safeguardsmatchthesophisticationofthetechnologyandhowitis
deployed.AllentitieswhouseAIhavearesponsibilitytodososafelyandethically.
OurreviewcomesatapivotaltimeinthedevelopmentofAIregulationinAustralia.WesupporttheAustralianGovernment’sVoluntaryAISafetyStandardandintentiontointroducemandatoryguardrailsensuring
testing,transparencyandaccountabilityforAIinhigh-risksettings.
However,licenseesandthosewhogovernthemshouldnottakeawait-
and-seeapproachtolegislativeandregulatoryreform.Currentlicenseeobligations,consumerprotectionlawsanddirectordutiesaretechnologyneutralandlicenseesneedtoensurethattheiruseofAIdoesnotbreachanyoftheseprovisions.
ASIC’sworktoengagewithandmonitorlicensees’AIusewillcontinue,particularlyasweconsiderhowtheyembedthe
requirementsofanyfutureAI-specificregulatoryobligations.
Icallonindustrytoconsiderthefindings
ofthisreviewandreflectonthequestionsposedtoensurethatinnovationisbalancedwiththeresponsible,safeandethicaluseofthistechnology.
JosephLongo
ASICChair
1DepartmentofIndustry,ScienceandResources,
ListofCriticalTechnologiesintheNationalInterest
:AITechnologies3
Executivesummary
ArtificialintelligencehasthepotentialtotransformtheprovisionoffinancialservicesandcreditinAustralia.Itprovides
opportunitiesformoreefficient,accessibleandtailoredproductsandservices.However,AIcanalsoamplifyexistingrisksto
consumersandintroducenewones.Potentialharmsincludebiasanddiscrimination,provisionoffalseinformation,exploitation
ofconsumervulnerabilitiesandbehaviouralbiases,andtheerosionofconsumertrust.Tohelpshapeourunderstandingofrisktoconsumersandtoinformourregulatoryresponse,wereviewedtheuseofAIby23AFSandcreditlicensees.
Ourreview
Weanalysed624AIusecasesthat23licenseesinthebanking,credit,insuranceandfinancialadvicesectorswereusing,ordeveloping,as
atDecember2023.ThesewereusecasesthatdirectlyorindirectlyimpactedconsumersandincludedgenerativeAIandadvanceddata
analytics(ADA)models.
Aspartofourreview,wealsoaskedlicensees
abouttheirriskmanagementandgovernance
arrangementsforAI,andhowtheyplanned
touseAIinthefuture.Wemetwith12ofthe
licenseesinJune2024todiscusstheirusecasesandgovernancearrangements.
Whatwefound
WeobservedarapidaccelerationinthevolumeofAIusecases.Wealsoobservedashift
towardsmorecomplexandopaquetypesofAIsuchasgenerativeAI.
Butonthewhole,thewaylicenseesusedAI
wasquitecautiousintermsofdecisionmakingandinteractionswithconsumers:AIgenerally
augmentedratherthanreplacedhumandecision
makingandtherewasonlylimiteddirectinteractionbetweenAIandconsumers.
Themajorityoflicenseestoldustheyare
planningtoincreasetheiruseofAI.Given
thefast-movingnatureofAIandcompetitivepressuresinindustry,thereispotentialfor
thewayAIisusedandtheassociatedrisktoconsumerstoshiftquickly.
Weareconcernedthatnotalllicenseesare
wellpositionedtomanagethechallengesof
theirexpandingAIuse.Somelicenseeswere
updatingtheirgovernancearrangementsatthesametimeasincreasingtheiruseofAI.And
inthecaseoftwolicensees,AIgovernancearrangementslaggedAIuse.
Governanceandriskmanagementarrangements
are,bytheirnature,slowtochange.Itis
thereforelikelythatanygapbetweentheuseofAIandgovernancearrangementswillwidenasAIadoptionincreases.Thiscouldleavelicenseesunpreparediftheywanttorespondquicklybutsafelytoinnovationsfromcompetitors.
KEYSTATISTICS
?57%ofallusecaseswerelessthantwoyearsoldorindevelopment.
?61%oflicenseesinthereviewplannedtoincreaseAIuseinthenext12months.
?92%ofgenerativeAIusecasesreportedwerelessthanayearold,orstilltobedeployed.
GenerativeAImadeup22%ofallusecasesindevelopment.
?Only12licenseeshadpoliciesinplaceforAIthatreferencedfairnessorrelatedconceptssuchasinclusivityandaccessibility.
?Only10licenseeshadpoliciesthatreferenceddisclosureofAIusetoaffectedconsumers.
ASIC?REP798
4
Executivesummary
OURFINDINGS
UseofAI
FINDING1:TheextenttowhichlicenseesusedAIvariedsignificantly.SomelicenseeshadbeenusingformsofAIforseveralyearsandothers
wereearlyintheirjourney.Overall,adoptionofAIisacceleratingrapidly(seepage11).
FINDING2:Whilemostcurrentusecasesusedlong-established,well-understoodtechniques,thereisashifttowardsmorecomplexand
opaquetechniques.TheadoptionofgenerativeAI,inparticular,isincreasingexponentially.
Thiscanpresentnewchallengesforriskmanagement(seepage13).
FINDING3:ExistingAIdeploymentstrategiesweremostlycautious,includingforgenerative
AI.AIaugmentedhumandecisionsor
increasedefficiency;generally,AIdidnotmakeautonomousdecisions.Mostusecasesdidnotdirectlyinteractwithconsumers(seepage15).
Riskmanagementandgovernance
FINDING4:NotalllicenseeshadadequatearrangementsinplaceformanagingAIrisks(seepage19).
FINDING5:Somelicenseesassessedrisksthroughthelensofthebusinessratherthantheconsumer.Wefoundsomegapsinhowlicenseesassessedrisks,particularlyriskstoconsumersthatarespecifictotheuseofAI,suchasalgorithmicbias(seepage20).
FINDING6:AIgovernancearrangements
variedwidely.WesawweaknessesthatcreatethepotentialforgapsasAIuseaccelerates
(seepage24).
FINDING7:Thematurityofgovernance
andriskmanagementdidnotalwaysalignwiththenatureandscaleoflicensees’AI
use–insomecases,governanceandrisk
managementlaggedtheadoptionofAI,
creatingthegreatestriskofconsumerharm(seepage29).
FINDING8:ManylicenseesreliedheavilyonthirdpartiesfortheirAImodels,butnotallhadappropriategovernancearrangementsinplacetomanagetheassociatedrisks(seepage31).
ASIC?REP798
WeobservedarapidaccelerationinthevolumeofAIusecases,andashifttowardsmorecomplexandopaquetypesofAIsuchasgenerativeAI.Butonthewhole,thewaylicenseesusedAIwasquitecautious.WefoundsomegapsinhowlicenseesassessedriskstoconsumersfromAI,andforsomelicensees,governancearrangements
laggedtheirAIuse.Thiscreatesriskofconsumerharm.
5
Executivesummary
Wheretofromhereforlicensees?
ASICsupportsinnovationinthefinancialsystemthatisbalancedwithappropriateconsumer
protectionsandmarketintegritysafeguards.
Whilelicensees’deploymentstrategieswere
somewhatcautious,thereisfertilegroundforconsumerharmwhereuseofAIleapsaheadofgovernancearrangementsandcontrols.
Weexpectlicenseestocarefullyconsidertheir
readinesstodeployAIsafelyandresponsibly.
DecisionsthatlicenseesmakenowabouthowtheywillgoverntheirAIusewilldeterminewhethertheyestablishsolidfoundationsonwhichtodelivertheexpectedbenefitsandmanageriskstothemselvesandtheircustomers.
Manylicenseestoldusthattheywereupdatingtheirgovernancearrangementsinrelationto
AI.Thisiswelcome,butthereismoretodo.
AIpresentsnovelchallenges,andlicensees’
governancearrangementsshouldleadtheirAIuseasitincreasesandevolves.
Licenseesshouldconsiderthefindingsofthis
report,andthequestionsonpages35–36,to
helpthemconsidertheirreadinesstodeployAI
safely,responsiblyandincompliancewithexistingobligations.
Licensees’obligationsandresourcesforlicensees
Theregulatoryframeworkforfinancialservices
andcreditistechnologyneutral.LicenseesneedtoconsidertheirexistingregulatoryobligationsbeforedeployingAI.Inparticular,licenseesneed
toconsiderthegenerallicenseeobligations,
directors’duties,andconsumerprotection
provisions,includingprohibitionsagainst
unconscionableconductandfalseormisleadingrepresentations(seepage34).
Thereareanumberofresourcesthatlicensees
candrawonastheydeployAI,suchasthe
recentlyissued
VoluntaryAISafetyStandard
.ThisstandardgivespracticalguidancetoallAustralianorganisationsonhowtosafelyuseandinnovatewithAI.
LicenseeswhoinvestthetimenowwillalsobeinabetterpositiontocomplywithanyfutureAI-
specificregulatoryobligations.
Thefutureregulatorylandscape
ThelandscapeofAIregulationinAustraliais
evolving.TheAustralianGovernmentrecently
consultedonhowitproposestodefine‘high-riskAI’,andtheintroductionofmandatoryguardrailstopromotethesafedesign,developmentand
deploymentofhigh-riskAIuse.Theproposed
guardrailsincluderequirementsrelatedtotesting,transparencyandaccountabilityofAI.
ASICsupportstheintroductionofregulatory
measurestomandateguardrailsfortheuseofAIinhigh-risksettings.ThefindingsofthisreviewhaveinformedourcontributiontotheGovernment’s
proposals.
ASIC’sfocus
Weremainfocusedonadvancingdigitalanddataresilienceandsafety,targetingtechnology-enabledmisconductandthepooruseofAI.UnderstandingandrespondingtotheuseofAIacrosstheentitiesweregulateisakeypriorityforASIC.
Wewill:
?continuetomonitorhowourregulated
populationusesAI,andtheadequacyoftheirriskmanagementandgovernanceprocesses
?contributetotheAustralianGovernment’sdevelopmentofAI-specificregulation
?engageandcollaboratewithdomesticandinternationalregulatorcounterparts,and
?wherenecessaryandappropriate,take
enforcementactioniflicensees’useofAIresultsinbreachesoftheirobligations.
AIpresentsnovelchallenges,and
licensees’governancearrangementsshouldleadtheirAIuseasitincreasesandevolves.Licenseesshouldreviewtheirarrangementsinlinewithour
findings.
ASIC?REP798
6
Executivesummary
CASESTUDY
BewarethegapbetweenAIuseandgovernance
Onelicenseecited10AIusecasesinscope,butadoptionappearedtofront-runtheirgovernance
andriskmanagementarrangements.The
licenseehadnooverarchingAIstrategysettingouthowandwhythelicenseehaddecidedto
useAIinitsoperations.Thelicenseeproduced
nopoliciessettingoutstandardstoguidethe
design,deploymentandoversightofAI,andhadnotarticulatedthekeyrisksassociatedwithAI
andADAintheirriskmanagementframework
(e.g.alackofexplainabilityforcomplexmodels).Noneofthelicensee’susecaseswereriskrated.
ThelicenseeusedanAImodeltopredict
consumercreditdefaultriskbyproducingariskscore.Thescoreproducedbythemodelwasoneinputintocreditdecisions.Ithadthepotential
toresultinconsumersbeingrefusedcreditorofferedlesscreditthantheyotherwisewouldhavebeen.
Aninternalreporttoaseniorcommitteedatedapproximately10monthsafterdeployment
ofthemodelstatedthatitwasdeveloped
with‘limitedunderstanding’ofthethird-partyplatformused,therewas‘incompletemodel
documentationwithmissingcriticalelements’,
and‘poorgovernanceandalackofamonitoringprocess’.
Thereportfurtherdescribedthemodelasa
‘blackboxwithnoabilitytoexplainthevariablesinthescorecardortheimpacttheyarehavingonanapplicant’sscore’.
Althoughthelicensee’sreportstatedthat‘themodelhasbeenstable’,itnotedthatitsabilitytomonitorthemodelwaslimited.Thereportproposed‘torevisethe[model],toensureit
isexplainable,documented,andhasarobustgovernanceprocessinplace’.Thelicensee
continuedtousethemodelforseveralmonthsbeforereplacingitwithasimplermodel,to
ensurescoringoutcomesandthemodelwereexplainable.
DespitetheissuesidentifiedwiththeaboveAI
model,thelicenseereportedhavingplansto
expandtheiruseofAI.Theyalsonotedthat
iftheydidnotengagewiththesecapabilities,
theywouldbe‘leftbehind’bycompetitors.Thelicenseereferredtoongoingworktoupdatetheirgovernanceandriskmanagementframeworks.
However,thisexampleexemplifiestherisk
inproceedingtoadoptAIwithoutadequate
foundationsinplace,andtheriskthatgaps
betweenusecasesandgovernancewillremainorwideninthefaceofcompetitivepressures.
ASIC?REP798
7
ASIC?REP798
AIgovernance:Questionsforlicensees
OVERSIGHT
1
9
TAKINGSTOCK
WhereisAI
currentlybeingusedinyour
business?
ALIGNMENT
Areyourgovernance
arrangementsleadingorlaggingyouruseofAI?
6
WhathumanoversightdoesyourAIuserequire,andhowwillyoumonitorit?
3
FAIRNESS
Howwillyouprovideservicesefficiently,
honestlyandfairlywhenusingAI?
THIRDPARTIES
10
Howdoyoumanagethe
challengesofusingmodelsdevelopedbythirdparties?
7
2
11
POLICIES
Haveyoutranslated
REGULATORYREFORM
STRATEGY
WhatisyourstrategyforAI,nowandin
thefuture?
AreyouengagingwiththeregulatoryreformproposalsonAI?
4
ACCOUNTABILITY
Whoisaccountable
forAIuseand
yourAIstrategyintoclearexpectations
forstaff?
5
8
RISKS
Howwillyouidentifyand
RESOURCES
DoyouhavethetechnologicalandhumanresourcestomanageAI?
outcomesinyourbusiness?
manageriskstoconsumersandregulatoryrisksfromAI?
Formoredetails,seepages35–36
8
WhylookatAI?
TheuseofAIinfinancialservicesandcreditcreatesthepotentialforsignificantbenefitstoconsumers,suchasmoreefficient,
accessibleandtailoredproductsandservices.
ButAIcanamplifyexistingrisksandcreatenewriskstoconsumers.
Thepotentialriskstoconsumers
Unfairorunintendeddiscriminationdueto
biasedtrainingdataoralgorithmdesign:
BiasedAIoutputscouldhavedisproportionate,negativeimpactsonvulnerableindividuals
orgroups,includingfinancialexclusion(forexample,beingdeniedaccesstocreditorinsurance,orpayingahigherprice).
Incorrectinformationprovidedtoconsumersaboutproductsorservices:AImodelscan
provideinformationoradvicethatappears
correct,butcontainsfactualerrorsorfallacies.Thisexposesconsumerstotheriskofharm
fromrelyinguponsuchmisleadingorfalseinformation.
Manipulationofconsumersentimentor
exploitationofbehaviouralbiases:AIcanallowforfasteriterationofmarketingandadvertisingmaterial,andbespokemicro-targeting.AIcan
playoncustomers’feelingsandrestrictormanipulatetheirchoices.
Breachesofdataprivacyandsecurity:AI
modelsmaycontainorreproduceconfidentialorsensitiveinformationwithoutthepriorandinformedconsentofimpactedindividuals.AImodelscanalsobevulnerabletocyberattacksanddataleaks.
Anerosionofconsumertrustandconfidenceduetoalackof:
?explainability–AImodelsmayuse
techniquesthataretoocomplextobe
understoodandexplainedbyhumansandbetrainedondatathatistoovastandcomplexforhumanstoprocess,resultingina‘black
box’,wheredecisionsmaynotbetraceable.
?transparency–ConsumersmaynotbeinformedwhenAIhasbeenusedtomakedecisionsthatimpactthem,orwhentheyareinteractingwithAIandAIgeneratedinformation,and
?contestability–Consumersmaynotbe
providedwithaprocessandthenecessary
informationtocontesttheoutcomeofa
decisionfacilitatedbyAI.Contestabilityis
furtherunderminedifconsumersareunawareoftheuseofAI.
MANAGINGRISKSFROMAI
RisksareveryspecifictoeachAIusecase.Forexample,theycanarisefromthedatainput,
fromthetechniqueormodelused,aswell
asfromthepurpose,context,andlevelof
automationofthemodels.RiskscanalsoarisethroughouttheAIlifecycleandcanchangeovertime.
BecauseAIoperatesatscale,usingvast
amountsofdata,riskscanbeamplified,andhavethepotentialtocauseharmatscale.
ThismeansthatAIcreatesnewchallengesfor
licenseesinmanagingriskstoconsumersfromAI.WhilethisreviewdidnottesttheoutcomesfromindividualAIusecases,wehavemade
observationsonwhetherlicenseesarepreparedtomanagetheriskofharmsfromtheuseofAI.
ASIC?REP798
9
ASIC?REP798
FINDINGS:
UseofAI
10
ASIC?REP798
FINDING1
TheextentofAIusevariedsignificantlybutoveralladoptionisaccelerating
Whatwedid
Wecollecteddatafrom23licenseesonthe
numberofAIusecasesinuseorindevelopment(asatDecember2023)whereAIinteractedwithorimpactedconsumers.
Whatwefound
?AllbuttwolicenseesreportedatleastoneAIusecasethatdirectlyorindirectlyimpactedconsumers.
?Thenumberofusecaseseachlicensee
reportedvariedsignificantly–seeFigure1.
Figure1:Usecasesreportedbylicensees
Numberoflicensees
12
10
8
6
4
2
0
Fewerthan66–2526–100Morethan100
Numberofusecasesreported
Note:SeeTable2forthedatashowninthisfigure(accessibleversion).
11
ASIC?REP798
Numberofusecases
FINDING1(continued)
TheextentofAIusevariedsignificantlybutoveralladoptionisaccelerating
Whatwedid
?AIadoptionisincreasingrapidly:57%ofallusecasesreported
werelessthantwoyearsoldorindevelopment.Ofthe624usecasesreportedtous,20%werestillindevelopmentandhadnotyetbeendeployed.
?TheadoptionofgenerativeAIis,unsurprisingly,averyrecent
development:92%ofgenerative
AIusecasesweredeployedin
2022or2023,orindevelopmentasatDecember2023.
?Wecanexpectthepaceof
changetocontinue:61%of
licenseesinthereviewtoldustheyplannedtoincreasetheiruseofAIinthenext12months.TheremainderwereplanningtomaintaintheircurrentAIuse.
Wereviewedatotalof624usecases(seeAppendix1)andmappedthemtotheyeartheyweredeployed.
Whatwefound
Figure2:NumberofAIusecasesbydeploymentyear
150
120
90
60
30
0
GenerativeAI
Non-generativeAI
ChatGPTreleased(30Nov2022)
2000
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
Dev
Note1:SeeTable3forthedatashowninthisfigure(accessibleversion)
Note2:Dev=Advisedtobeindevelopment
bythelicenseeasatDec2023–seeAppendix1formoreinformation.Thedevelopmentdatesof12usecaseswerenotprovidedordidnot
haveacleardateandarenotreflectedinthisgraph.Thisgraphincludesusecasesreportedas‘inproduction’or‘indevelopment’asatDec2023.Itdoesnotincludeusecasesbuiltand
decommissionedbeforethedatacollection,orusecaseswherethemodeltechniquewasnotspecified.
12
ASIC?REP798
FINDING2
Mostcurrentusecasesappliedlong-established,well-understoodtechniques.Buttherewasashifttowardsmorecomplexandopaquetechniques,includinggenerativeAI
Whatwedid
Weassessedthecomplexityofmodeltechniquesusedineachofthe624usecases.Complex
andopaquetechniquescanposeadditional
challengesforoversight.Challengesinclude
understandingandexplaininghowAIobtainsits
results,determiningwhetherresultsarereliableandaccurate,andknowingwhetheroutputsareunfairlybiasedordiscriminatory.
Whatwefound
?Themajorityofcurrentusecasesreliedon
well-knownandestablishedmachine-learningtechniquesthatproducedexplainableand
interpretableresults.
?Weobservedanincreaseintheuseofmore
complexandopaquetechniques(suchasneuralnetworksusedindeeplearningandgenerativeAI),whichareusedfortheprocessingand
analysisoflargevolumesofimages,audio
andtextdata–seeFigure3.Togethertheserepresent32%oftheusecaseswesawunderdevelopment.
?TheuseofgenerativeAIissettoincrease
exponentially.WhilegenerativeAImadeuponly5%ofusecasesthatwereinuse,itmadeup22%ofthoseindevelopment.
Figure3:Modeltechniquesbystatus
Supervisedlearning:Classification
Supervisedlearning:Regression
Deeplearning
Unsupervisedlearning
GenerativeAI
Miscellaneous
Notspecified
Current(n=488)
Indevelopment(n=124)
WhilegenerativeAImadeuponly5%ofusecasesthatwereinuse,itmadeup22%ofthoseindevelopment
0%10%20%30%40%Note:SeeTable4forthedatashowninthisfigure(accessibleversion)
13
FINDING2(continued)
Mostcurrentusecasesappliedlong-established,well-understoodtechniques.Buttherewasashifttowardsmorecomplexandopaquetechniques,includinggenerativeAI
Howdifferentmodeltechniqueswereused
Supervisedlearning:Classificationmodelsweremostlyusedtopredictifaconsumerwaslikely
totakeoutafinancialproductusingexplainablemodelssuchaslogisticregression.
Supervisedlearning:Regressionmodelswereprimarilyusedtoderiveprices,ratesorforecastfuturetimeseries.
Deeplearningmodelsweremostlyusedfor
naturallanguageprocessingandopticalcharacterrecognition,primarilywhenscanninganalogue
formdatatospeeduploan,insurance,orotherform-heavybusinessprocesses.
Unsupervisedlearningmodelsweremostlyusedfordetectingstrangeoranomalouspatternsin
areassuchasinternalauditandfrauddetection.
GenerativeAImodelswereusedtogeneratefirstdraftsofmaterials,orresponsestocustomersincarefullyconstrainedcircumstances–seepage15formoreinformation.
Miscellaneousmodelsweremostlynon-predictivemodels,suchassearchengineoptimisationorpatternmatching.
‘Notspecified’modelsweremodelswhere
licenseesdidnotdisclosethemodeltechnique.Insomecases,theseweremodelsbuiltby
thirdparties,andlicenseesdidnothavethisinformation.
WHATISGENERATIVEAI?
GenerativeAIisatypeofAIthatfocusesoncreatingorgeneratingnovelcontentsuchasimages,text,music,video,designsor
recommendations.
UnliketraditionalAItechniquesthatproduce
outputthatisprogrammedorcopiedfrom
existingdata,generativeAItechniquesare
designedtogenerateoutputbasedonpatterns,structuresandexampleslearnedfromlargedatasetsduringthetrainingprocess.
GenerativeAImodelshavecertain
characteristicsthatmakethemparticularlypronetorisksofharm.Forexample,they:
?tendtouselargeamountsofdataforthe
trainingofthemodel.Thepresenceof
incompletedataintrainingsetsmeanthatmodelshavethepotentialtoprovidebiasedorinappropriateresults
?cangenerateoutputsthatarefalseorinaccurate
?canusecomplextechniquesthatarenot
easilyinterpretableorexplainable,and?canbesubjecttonovelcyberattacks.
ASIC?REP798
14
ASIC?REP798
FINDING3
ThewayAIwasusedwasmostlycautious
Whatwedid
WelookedatwhatlicenseeswereusingAI
for,theroleAIplayedindecision
溫馨提示
- 1. 本站所有資源如無特殊說明,都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請(qǐng)下載最新的WinRAR軟件解壓。
- 2. 本站的文檔不包含任何第三方提供的附件圖紙等,如果需要附件,請(qǐng)聯(lián)系上傳者。文件的所有權(quán)益歸上傳用戶所有。
- 3. 本站RAR壓縮包中若帶圖紙,網(wǎng)頁內(nèi)容里面會(huì)有圖紙預(yù)覽,若沒有圖紙預(yù)覽就沒有圖紙。
- 4. 未經(jīng)權(quán)益所有人同意不得將文件中的內(nèi)容挪作商業(yè)或盈利用途。
- 5. 人人文庫網(wǎng)僅提供信息存儲(chǔ)空間,僅對(duì)用戶上傳內(nèi)容的表現(xiàn)方式做保護(hù)處理,對(duì)用戶上傳分享的文檔內(nèi)容本身不做任何修改或編輯,并不能對(duì)任何下載內(nèi)容負(fù)責(zé)。
- 6. 下載文件中如有侵權(quán)或不適當(dāng)內(nèi)容,請(qǐng)與我們聯(lián)系,我們立即糾正。
- 7. 本站不保證下載資源的準(zhǔn)確性、安全性和完整性, 同時(shí)也不承擔(dān)用戶因使用這些下載資源對(duì)自己和他人造成任何形式的傷害或損失。
最新文檔
- 2024民政局官方離婚協(xié)議范本制作與婚姻法律援助服務(wù)協(xié)議3篇
- 2024年學(xué)校安全工作計(jì)劃范文(35篇)
- 產(chǎn)教融合人才培養(yǎng)共同體的構(gòu)建目標(biāo)
- 三角形-三角形的特性說課稿-2023-2024學(xué)年四年級(jí)下冊(cè)數(shù)學(xué)人教版
- 2025年幼兒園中班班務(wù)計(jì)劃開頭語模板
- 2025年皮膚科醫(yī)師工作計(jì)劃
- Unit 4 Plants around us Start to read and Reading time(說課稿)-2024-2025學(xué)年人教PEP版(2024)英語三年級(jí)上冊(cè)
- 2025年學(xué)校教師教學(xué)個(gè)人工作計(jì)劃
- 畢業(yè)活動(dòng)主題策劃方案合集
- Unit 2 My week Part C(說課稿)-2024-2025學(xué)年人教PEP版英語五年級(jí)上冊(cè)
- 2024年06月上海廣發(fā)銀行上海分行社會(huì)招考(622)筆試歷年參考題庫附帶答案詳解
- TSG 51-2023 起重機(jī)械安全技術(shù)規(guī)程 含2024年第1號(hào)修改單
- 計(jì)算機(jī)科學(xué)導(dǎo)論
- 浙江省杭州市錢塘區(qū)2023-2024學(xué)年四年級(jí)上學(xué)期英語期末試卷
- 《工程勘察設(shè)計(jì)收費(fèi)標(biāo)準(zhǔn)》(2002年修訂本)
- 2024年一級(jí)消防工程師《消防安全技術(shù)綜合能力》考試真題及答案解析
- 2024-2025學(xué)年六上科學(xué)期末綜合檢測卷(含答案)
- 安徽省森林撫育技術(shù)導(dǎo)則
- 泌尿外科護(hù)理疑難病例討論
- 英格索蘭空壓機(jī)操作規(guī)程
- 簡單通用利潤表
評(píng)論
0/150
提交評(píng)論