AIriskandapproachestoglobalregulatory

compliance

Introduction

Foreword 04

KeyTakeaways 05

ExecutiveSummary 06

Chapter1

:

Fromfictiontoreality–definingAI,itsevolution,opportunity,andimpact 08

1.1.WhatisAI? 08

1.2WhataretherisksofAI? 08

1.3WhatareregulatorsdoingaboutAIrisksandwhatarethechallenges? 11

Chapter2

:

WhatarethekeyinternationalAIgovernanceprocessesandwhatrisks

andissuesaretheyfocusingon? 12

2.1WhicharethekeyinternationalAIgovernanceprocesses? 12

2.2Trendsandoutlookforinternationalcooperation 14

2.3.WhatAIrisksandissuesarethesebodiesfocussedon? 14

Chapter3:Commonrisksandissuesacrossnationalandregionalregulatoryapproaches 16

3.1WhatAIrisksandissuesarenationalandregionalapproachesfocussingon? 16

Chapter4

:

Keythemesidentifiedacrossallinternationalandnationalregulatoryapproaches 17

4.1Keythemesacrossinternational,national,andregionalapproaches 17

4.2Considerationofhowkeythemesaddressedinregulatoryapproachescompare

tothebroadAIrisklandscape 17

4.3.Deloitteview 17

Chapter5:Nationalandregionalregulatoryapproaches 19

5.1.Overviewofnationalandregionalapproaches 19

5.2.Detailofspecificnationalandregionalapproaches 22

5.3.DefinitionsofAIinnationalandregionalapproaches 26

5.4.Deloitteview 27

Chapter6

:

Howcancompaniesprepareforregulatorycompliance? 28

6

.1.UnderstandtheimpactofregulationsonyourAIbusinessstrategy 30

6.

2.CreateorganisationalclarityaroundthekeyoperationalchallengesfromAIregulation 30

6.

3.Engagekeystakeholdersacrosstheorganisationtoavoidsilos 32

6.

4.DesignandimplementAIgovernanceandriskmanagementframework 32

6.

5.Implementsomenoregretsactionsnow 33

Contacts 36

04

1

2

4

5

3

6

AIriskandapproachestoglobalregulatorycompliance|Foreword

Foreword

JurisdictionsaroundtheworldarerapidlyintroducingArtificial

lnte||igence(Al)regu|ationastheybetterunderstandtherisks,bothactua|andperceived,thatAlposes.Understandingtheserisksis

amajorcha||engeforregu|ators,aswe||asfororganisationswhoaredeve|opingtheirAlstrategiesandAlsystems,anddep|oying

thesesystemsandembeddingmode|sintotheiroperationsand

products.ItcanbedifficultforcompaniesinvestinginAItomanagetheuncertaintiescreatedbyadynamicregu|atory|andscape

whi|ststi||enab|ingAlinnovation.lnparticu|ar,understandingtheregulatoryspaceandfindingthepatterns,trends,andstrategicimplicationsacrossdifferentsetsofregulationstosupport

comp|iance.

Thisregu|atoryandrisk|andscapeana|ysisfromDe|oitte,slnternetRegu|ationteamisdesignedtosupportcompanieswiththese

cha||enges.

Thereportisdividedintosixchapters:

?Chapter1definesAI,outlinesthebroadsetofrisksassociatedwithitsuse,andout|inessomeofthechoicesandcha||engesthatfaceregu|atorsindesigningregu|ation.

?Chapters2,3and4focusonidentifyingacommonsubsetofrisksandthemesthatinternationa|bodies,governmentsandregu|atorsarecurrent|yfocussedon-the“what”

?Chapter5examines“how”individua|regu|atorsareaddressingtheserisks

?Chapter6sharesperspectivesonwhatthismeansfor

companies,howtheycandeve|opappropriateguardrai|sto

addressAlrisks,andengagethemu|ti-stakeho|der|andscapetosupportregu|atorycomp|iance

Wehopethatthisreportwi||serveasausefu|resourcefor

companies|ookingtonavigatethecomp|ex|andscapeofAlriskandregu|ation.

05

1

2

4

5

3

6

AIriskandapproachestoglobalregulatorycompliance|KeyTakeaways

KeyTakeaways

?TheAIregulatorylandscapeisevolvingatbreakneckspeed.

AsofApril2024,morethan300AI-relatedlaws,guidance,or

regulationshavepassedorareindevelopmentacrosstheglobe

?Thispaceofchangeislikelytocontinueforaconsiderableperiod,creatinguncertaintyforinternationalorganisationswhoneedtomanageregulatorycompliancewhilststillinnovatinginAI

?Thereisclosealignmentacrossmajorinternationalprocessesandnationalregulatoryapproachesinthekeyareasthat

areaddressedthroughAIregulation–the“what”.These

areasarefundamental/humanrights;fairness;privacyanddatagovernance;safety;transparency;competition;andaccountabilityandhumanoversight

?TheseareasonlyrepresentasubsetofthewholeAIrisk

landscapefacingfirms.Thissuggeststhatfirmslookingacrossmultiplejurisdictions,andpursuingacompliance-firstapproachtomanagingAIrisk,maybeabletoprioritise

?Thereislessalignmentinhowdifferentjurisdictionsare

regulating,suggestingthatmanyfirmswillhavetomanagedivergence,whichcouldincreaseovertime.

?However,manyregulatoryapproachesaremultilayered(AI

specificrulessittingalongsideothertechneutralorsector

specificrulesimpactingAI).Thissuggeststhat,evenwithin

jurisdictions,firmswon’tjustbeabletosolveforAIregulationinisolation

?Elementsofariskandprinciples-basedapproacharealso

common.Alongsidedeliveringspecifictechnicalrequirements,thissuggeststhatgovernanceriskandcontrol,monitoringandtesting,documentation,auditandassurancearelikelytobekeyelementsof“how”firmscandemonstratecomplianceacross

jurisdictions

?AnumberofelementscansupporttheresponsetoglobalAIregulation,includingunderstandingthestrategicimpacts;creatingclarityaroundtheoperationalchallenges;engagingwidelyacrosstheorganisationasyoubuildyourcompliance

roadmap;anddesigningandimplementinganAIgovernanceandriskmanagementframework

?Asastartingpoint,firmsshouldformanAIgovernancecommittee,establishanAIsysteminventory,andbeginconductingAIsystemriskassessments

06

AIriskandapproachestoglobalregulatorycompliance|ExecutiveSummary

ExecutiveSummary

Fromfictiontoreality–definingAI,itsevolution,opportunity,andimpact

AIisapioneeringtechnologybutcanbehardtodefine.The

explosionoflargelanguagemodelshascatapulted“Generative

AI”(ormorecommonly“GenAI”)intothezeitgeist,butitisonly

oneofseveraldifferenttechnologiesthatmakeupAI.WhilstmanybusinessesrecognisethehugepotentialofAIandareincreasinglyadoptingitatscale,thereisalsowidespreadconcernaboutits

risks.1

AI’suniquenatureasatechnology–itsautonomyandability

tolearnandevolve-aswellasuncertaintyaboutitsultimate

potential,makeitdifficulttopreciselydefineAIrisk.2Aninterim

reportendorsedby29countries-theInternationalScientific

ReportontheSafetyofAdvancedAIwaspublishedinMay2024andprovidesacomprehensiveassessmentofthelatestscientificunderstanding,butalsoexposesthelevelofcontinueddebate

anduncertaintyaroundAIrisk.Fororganisationsbuildingtheir

approachtoAI,thisuncertaintypresentsmajorchallenges.We

haveidentifiedninebroadareasthatpresentasnapshotofthe

complexAIrisklandscape.Theseincludetechnicalandsecurity

challenges;impactsonsocialandculturaldynamics;andpotentiallong-termexistentialrisks.

Inresponsetotheserisks-inparticularthepotentialforharm

toindividuals-andtherapidupswinginboththecapabilityand

adoptionofAItechnologies,governmentsandregulatorsare

embarkingonawaveofnewregulatoryactivity.Governmentsfacekeydesignchoices,includingwhichriskstoprioritiseaddressing

intheirregulatoryapproach,howtobalanceriskmanagement

withpromotinginnovation,andhowtofuture-proofanynewAI

rules.Acommontheme,nomatterthefavouredapproach,isthatregulatoryapproachesaredevelopinginamultilayeredway,withdirectAIregulationsittingalongsideothercross-cuttingorsectorspecificregulation.Thiswillrequireorganisationstohaveastrongregulatoryhorizonscanningcapability.Commonly,legislatorsalsoappeartobepursuingelementsofaprinciples-based,risk-centricapproachakintootherlandmarkdigitalregulationsuchastheEUsDigitalServicesAct.

Whilsttheregulatorylandscapeisstillevolving,itisalreadyclearthatnewregulationwillhavemajorimpactsfororganisationsusingAI,andthatcomplyingwithregulatoryrequirementsmayrequiresignificantinvestment.GiventhebroadlandscapeofAIrisk,key

questionsfororganisationsseekingtoprioritisetheirapproachisunderstandingtheparticularareasofriskthatregulatorsarefocussingonandwhetherthereisconvergenceamongstthem.

WhatarethekeyinternationalAIgovernanceprocessesandwhatrisksandissuesaretheyfocusingon?

AIisaglobalissue,whichmeansthatthereisalreadysignificant

internationalcoordinationonunderstandingAIriskand

developingcorrespondingregulatoryapproaches.Anumberof

keyinternationalbodiessuchastheUnitedNations(UN),Groupof7(G7),Groupof20(G20)andtheOrganisationforEconomic

CooperationandDevelopment(OECD)havesoughttoestablishareasofconsensusandcommongroundamongstmember

countriesandhavebeendevelopingsharedAIPrinciples,non-

bindingCodesofConduct,orJointDeclarations.Thisprocessis

likelytocontinuefortheforeseeablefuture,includingaparticularfocusonissuesaroundAIsafety.Itispossiblethatinfuture

thesebodieswillseektotranslatevoluntarymeasuresintomore

actionablerequirements,andthatwemayseeasetofbinding

internationalrulesinsomeareas,suchasAIsafety,underpinnedbynationalframeworks.

Theoutputsoftheseinternationalprocesses,includingdescribingthecharacteristicsofsafeandtrustworthyAIandidentifyingissuestobeaddressedthroughregulatoryaction,areaninvaluable

indicationforfirmslookingtounderstandemergingareasofconsensusacrossAIregulationandtoprioritiseaccordingly.

Analysiscarriedoutforthisreportshowsthatthereishigh

convergenceacrosstheseinternationalprocessesin“what”thekeyissuestobetackledthroughregulatoryapproachesare,

withafocusaroundtacklingindividualharmsandensuringtrust

andsafety.Themostcommonlyidentifiedareasinclude:the

protectionofhuman/fundamentalrights;fairness;privacyanddatagovernance;safety;andtransparency.

Commonrisksandissuesacrossnationalandregionalregulatoryapproaches

Alongsidetheseinternationalprocesses,nationalandregional

regulatoryapproachesarenowtakingshape.Theseapproaches

areofteninformedbythediscussionsandconsensusreached

atinternationallevel.Onceagain,analysisforthisreportofa

globallyrepresentativegroupofnationalandregionalregulatory

approachesintheUnitedStates,EuropeanUnion,UnitedKingdom,Australia,SingaporeandJapan,showsthatthereisstrong

convergenceacrossnationalandregionalapproachesintermsof“what”AIrisksandharmstheyseektoaddress.

Furthermore,thereisverystrongcorrelationbetween

international,nationalandregionalprocesses.SuchahighdegreeofalignmentmayhelpfirmstoprioritiseareastofocusonwithinthebroadandcomplexAIrisklandscape.

1

2

4

5

3

6

1

StateofAIintheEnterprise2022|DeloitteUS

2InternationalScientificReportontheSafetyofAdvancedAI-InterimReport(.uk)

07

AIriskandapproachestoglobalregulatorycompliance|ExecutiveSummary

Keythemesidentifiedacrossallregulatoryapproaches

Inthisreport,wehaveidentifiedthesevenmostcommonareasintheregulatoryspaceacrossallinternational,regionalandnationalapproachesanalysedthatinternationalorganisationsplanningforregulatorycomplianceacrossmultiplejurisdictionsmaywanttoputattheforefrontoftheirownactivities,riskassessmentsandmitigations.Theseare:

?Fundamental/humanrights

?Fairness

?Privacyanddatagovernance

?Safety

?Transparency

?Competition

?Accountabilityandhumanoversight

However,regulatoryapproachesmaynotatpresentaddress

allofthebroadAIrisksthatorganisationswillidentifythemselves.Analysisinthisreportshowsthatregulatoryapproachesdonot

coverthefullAIrisklandscape,andthatsomeareasareonly

coveredalittleornotatall.Inotherwords,justbecauseanAI

isnotriskyfromaregulatoryperspective,doesntmeanthere

isnorisk.Thismeansthatorganisationsmayneedtomanage

regulatorycompliancealongsideaddressingotherAIrisksthat

theyhaveidentifiedtotheirorganisation.OneoptionformanagingbroaderrisksatanorganisationallevelistouseaTrustworthy

AIFramework.3Suchaframeworkaddressesabroadsetof

ethicalandresponsibleAIareasandcanbepartofthetoolkitforregulatorycompliancebutwillnotreplaceitgiventhespecific

requirementssetoutindifferentregulatoryapproaches.

Nationalandregionalregulatoryapproaches

Whiletheseisstrongcorrelationon“what”regulatoryapproachesareseekingtoaddress,ouranalysisshowsthatthereisalready

somedivergenceinhowthesesixkeynationalandregional

regulatoryregimesareapproachingAIregulation.Thismayreflectthedifferentmaturityoftheseregimes.Thereportidentifiesthreebroadapproachestoregulation,withthepotentialforthistoevolveinthecomingyearsasthescope,details,andinterdependenciesofAIregulationsdevelop:

?HorizontalregulationoftheuseofAIasawhole

?VerticalregulationofAIasitoccursindifferentpartsoftheeconomyorsociety

?Theapplicationofcodesofconduct,principles,ormodel

governancewhereregulatorshavenotyetdeterminedtheirpreferredapproachorbelieveitistoosoontodoso

AlltheregulatoryapproachesexaminedaddressAIina

multilayeredway,withtechnologyneutralandsectorspecific

regulationoperatingintandemwithAIspecificrules.Complying

withmultipleregulationswithinjurisdictionswillbeakeychallenge.Elementsofariskandprinciples-basedapproacharealso

common,requiringfirmstoconsidertherisksoftheirAIfromfirstprinciplesandtoapplyappropriatemitigations.

1

2

4

5

3

6

Thissuggeststhataswithotherkeypiecesoftechregulation-governance,riskandcontrol;monitoring;anddocumentationcouldprovidethefoundationsfor“how”tobuildanapproachtocomplianceacrossglobalregulations.

Howcancompaniesprepareforregulatorycompliance?

UsingDeloittesexperiencesupportingorganisationswhohave

facedotherdigitalregulatorywavesinthepast,wehaveidentifiedfiveelementstosupportanorganisationalresponsetoglobalAI

regulationandtohelpnavigatetheuncertaintyofanevolvingAI

regulatorylandscapewhilststillenablingAIinnovation.Thesestepsare:

?UnderstandingtheimpactofregulationsonyourAIbusinessstrategy

?Creatingorganisationalclarityabouttheoperationalchallengestounderstandthegaps

?Engagingkeystakeholdersacrosstheorganisationtoensurenooperationalsilosemerge

?DesigningandimplementinganAIgovernanceandriskmanagementframeworkincluding:

DevelopingandAIsystempolicy

Developingquality,privacy,safetyandsecurityguardrailsBuildingaredteamingcapability

?Implementingsomenoregretsactionsnow,whilstyoudevelopandembedabroaderorganisationalapproach.Thesenoregretsactionsare:

FormanAIgovernancecommittee

CreateanAIsysteminventoryandclassifyyourAIsystemsGatherdocumentationonexistingAIsystemsincluding

developingexplainabilityandtransparencyAIsystemnotices

orcards

Identifyandperformagapassessment

Establishdynamicregulatoryintelligence

ConductAIsystemriskassessments

Startcommunicationsacrosstheorganisationandensurecrisispreparedness

3

DeloitteTrustworthyAIinPractice

08

AIriskandapproachestoglobalregulatorycompliance|Chapter1

1.1WhatisAI?

AIisapioneeringtechnologythathasthepotentialtochange

ourviewofwhatitmeanstobehuman.Itcouldhaveaprofoundimpactonhowwedefineourselvesnotjustasindividuals,butalsooursocietyandhowitfunctions.Itsinfluencehasbeensignificantinrecentdecades,withimpactsonhowbusinessesfunction

includingbyimprovingefficiency,accuracy,anddecision-making.However,AIalsomeansverydifferentthingstodifferentpeopleandmanystruggletodefineit.

OnewayofthinkingaboutAIisasanumbrellaterm,usedto

describemultipletechnologiesandmethodsthatseektoreplicateelementsofappliedhumanintelligence.Theexplosionoflarge

languagemodelsinthelasttwoyearshascatapulted“GenerativeAI”(ormorecommonly“GenAI”)intothezeitgeistandmadeitpartofthedailyconversationinboardroomsandlivingroomsalike.

WhilsttraditionalAIsystemsrelyonexplicitprogrammingand

predefinedrulestoanalysedataandmakepredictions,GenAI

toolscancreatenewcontentbasedonlearnedpatternsanddataacrossvariousmedia(e.g.text,images,audio,code,voice,video)oftenwithseeminglymagicalresults.Increasingly,GenAIandAIareusedinterchangeably.ButGenAIisonlyoneofanumberofAItechnologies,allofwhicharebeingdevelopedatsuchspeedthatenvisagingtheirultimateimpactisvirtuallyimpossible.

Manyregulatoryapproaches,suchastheEUsAIAct,alsoapply

abroaddefinitionofAIbasedonitscorecapabilitiesratherthandistinguishingbetweendifferentAItechnologiesandsoforthe

sakeofsimplicity,withinthispaper,wewilluseAIasablankettermcoveringallofAI,includingGenAIasasubset.

AlthoughAIisalreadyarguablythemosturgentstrategicpriorityforbusinesses,inmanywaysweareonlyjustscratchingthe

surfaceofitsultimatepotential.ItiswidelypredictedthattherewillcontinuetobemajoradvancementsinAIcapabilitiesandimpact,alongsiderapidadoption,intheperiodahead.

FindingsfromDeloittesStateofAIintheEnterprise4,5theditionreportinclude:

?96%ofbusinessleadersbelieveAIiscriticaltosuccessoverthenextfiveyears,andAIdeploymentsareupsignificantlythisyear

?79%ofrespondentssaythattheyhavefullydeployedthreeormoretypesofAIcomparedtojust62%in2021

However,alongsideallthepotentialopportunities,AIalsogives

risetosignificantrisks.Inthesamesurvey,over50%ofleaderscitemanagingAIriskasoneofthecriticalchallengesinadoptingAI.

1.2WhataretherisksofAI?

ThepaceofAIdevelopmentandtheuncertaintyofitsultimate

potential,aswellasitsnatureasatechnologyitsabilitytolearnandevolve;itsautonomy;thebreadthofitspotentialapplications;thecomplexityofthecontextwhichoftenformsthebasisofits

deployment;andtheimpactofhumanbehaviouronitcreatesanalmostuniquesetofrisksthatareasyetpoorlyunderstood,givingrisetoconcernsabouthowtoensuresafetyandpromotetrustinAI.

StateoftheScienceReport

“TheintentionoftheStateoftheScienceReportis

tofacilitateasharedscience-basedunderstandingoftherisksassociatedwithfrontierAIandtosustainthatunderstandingascapabilitiescontinuetoincrease.”

Theseuncertaintiesposechallengeforcompaniesandgovernments

alikewhoareseekingtodefineAIrisks.AttheinauguralUKAISafetySummitatBletchleyParkinNovember2023,the29governmentsrepresented,highlightedthattheircollectiveunderstandingof

AIriskisstilldeveloping,andthateffectivepolicyandregulation

requiresabetterunderstandingandconsensusaroundrisk.For

thisreason,theyestablishedanexpert-ledpaneltocreatethe

StateoftheSciencereporttosynthesiseevidenceonAIriskon

anongoingbasisandtoprovideaconsistentbasisfromwhich

theycandeveloparegulatoryresponse.5Aninterimreport-

InternationalScientificReportontheSafetyofAdvancedAI–waspublishedattheendofMayaheadofthemostrecentAISafety

SummithostedbytheRepublicofKoreaandUKgovernments.6

Thereport’sauthorsreasserttheimportanceofasharedscientific,evidence-basedunderstandingofAIriskandsafetyasthe

foundationfordiscussionsanddecisions.

2

3

4

5

6

1

Chapter1

Fromfictiontoreality–definingAI,itsevolution,opportunity,andimpact

4

StateofAIintheEnterprise2022|DeloitteUS

5

StateoftheScienceReporttoUnderstandCapabilitiesandRisksofFrontierAI:StatementbytheChair,2November2023-GOV.UK(.uk)

6

InternationalScientificReportontheSafetyofAdvancedAI-InterimReport(.uk)

09

AIriskandapproachestoglobalregulatorycompliance|Chapter1

Figure1:IllustrativeexampleofdifferentAItechnologies,includingGenerativeAI

Arti?cialGeneralIntelligence(AGI

AI

Simulation

GenerativeAI

Quantum

ConversationalAI

AutonomousSystems

Deep

Virtual

Learning

VisionAI

SpeedAssistants

Recognition

MachineLearning

PredictiveAnalytics

IntelligentAutomation

Arti?cialIntelligence(AI)isabroadmarketofwhichGenerativeAIisoneofthemanytechnologiesthatcandisrupthowsociety

interactsandbusinessisconducted...

However,whilerestatingthattherearesignificantpotentialrisksfromAI,theywereunabletoreachaclearconsensusondefiningthemostseverefrontierrisks,citing“differingexpectationsaboutthestepssocietywilltaketolimitthem,theeffectivenessofthosesteps,andhowrapidlygeneral-purposeAIcapabilitieswillbe

advanced”.TheexpertpanelwillpublishafinalversionofitsfirstreportattheFranceAISummitearlyin2025.

FororganisationsthataredevelopingordeployingAI,

understandingtherisksisamajorpreoccupation,andthedifficultyofpreciselyunderstandingAIriskcreatesmajorchallenges.Businessesneedtoidentifytheimpactatthreelevels:

?Atthestrategiclevel,tounderstandthepotentialnegative

effectsofAIontheiroverallbusiness,widermarketandoverallecosystem

?Atorganisationallevel,toassesstheimpactofAIonoverallriskappetiteandtolerance

?Atthesystemlevel,togaugeifspecificAItechnologiesarefitforpurposeandplanneduse

Ofcourse,organisationsareatdifferentstagesofimplementingtheirAIstrategyandwillhavedifferentlevelsofmaturityin

understandingtheirrisks,butthebroadAIrisklandscapeencompassesthefollowingbroadareas:

?Ethicalandmoralissues:ConcernsaroundAIsystemstakingconsequentialdecisionswithouthumanoversight,andabouthowthosedecisionscouldleadtobiasedoutcomes,increasediscrimination,orthenormalisationofstructuralinequalities

?Technicalandsecuritychallenges:Thismayincludethe

riskthatmodelsarevulnerabletomanipulation;orcanbe

usedbybadactorstocommitcrimeortargetcriticalnationalinfrastructure.ThetechnicalrisksfromAIcanbeparticularlycomplexfororganisationstounderstand.Withthisinmind,wehavesummarisedtheminmoredetailbelow:

Hallucination:AIcanproduceinaccurateormisleading

contentbydrawingonincomplete,inaccurate,orbiaseddata,orsimplygeneratingfabricatedfacts

Uncertainty:Unlikehumanswhowilloftenqualifytheir

answersdependingonthelevelofcertaintytheyhaveabouttheiranswer,AImodelstendtoprovideananswerwithoutequivocation.Thisisaparticularchallengewhencombinedwiththehallucinationrisk

Explainability:Itishardtoidentifya“truth”forAImodelsif

theydonothaveaclearinformationsource.Largelanguage

modelsaretrainedtoconstructsentencesbymakingaseriesofguessesonthestatisticallylikely“token”thatcomesnext,butthereisasyetlimitedunderstandingoftheexactprocessby

whichtheyarriveattheanswerprovided.Thiscanmakeithardtoaccuratelypredictreliability

Bias:AIcanlearnbiasesbasedonpatternsinthedataitistrainedon,andleadtocontentthatisdiscriminatoryormisleading

Lackofrobustness:Despitetheappearanceofhumanlevelknowledge,AIsystemsarebrittleandlackrobustness,meaningthattheyfrequentlyfailinsituationsthataresufficientlyunliketheirtrainingdata.Thereisalsoanelementofrandomness

inAI-ifyouaskthesamequestionseveraltimesslightly

differently,youcouldgetdifferentanswers-whichmeansit’smoredifficulttoauditortrack

Jailbreaking:Itcanberelativelysimpletopromptmodelstobypasstheirsafeguardsinordertogetthemtodosomethingthattheyarentmeanttodo.Forexample,promptingthe

modeltorespondaffirmativelytoarequestorto“imaginethatitisacompulsiveliar”

Specificationproblem:TheriskthatAIsystemspursue

unintendedgoalsgiventhechallengeofpreciselydefiningtheproblemtobesolvedandteachingAIwhichbehavioursaredesirableorundesirable7

2

3

4

5

6

1

7MoreinformationontechnicalrisksofAIsystemscanbefoundhere:

CapabilitiesandrisksfromfrontierAI(.uk)

andhere:

Risksandethical

considerationsofgenerativeAI|DeloitteUK

10

AIriskandapproachestoglobalregulatorycompliance|Chapter1

?Economicandemploymentimpact:Concernsthatsystems

carryingoutrolespreviouslyundertakenbyhumanswillhave

significantimpactsonthelabourmarket,orthatAIsystemscouldexacerbateexistingeconomicinstabilities

?Impactonsocialandculturaldynamics:TheincreasinguseofAIinplaceofhumaninteractionraisesconcernsaboutits

potentialnegativesocietalimpacts,includingalossofpersonalconnectionandempathy.Additionally,theuseofAIasatoolforsocialcontrolposesarisktoindividualfreedomsandprivacy,potentiallyleadingtodiscriminationandabuseofpower

?Environment,socialandgovernancechallenges:TheuseofAImaycreatesinglepointsoffailureinkeydomains,posingsignificantrisks.WidespreadadoptionofAIispredictedto

increaseenergyusage,leadingtoenvironmentalimpact

2

4

5

3

1

6

?Legalandjudicialissues:LegalframeworksmaystruggletokeeppacewithnewAI-enabledcrimes,anditiscreatinganewsetofchallengesaroundcopyrightandprotectingintellectualproperty

?Impactonknowledgeandinformation:ThepotentialforAItoincreasethevolumeandsophisticationofmisanddisinformationposesasignificantrisk

?Human-AIinteractionandpsychology:ConcernsaroundtheimpactthatwidespreaduseofAImayhaveontrust;ortheethicalandpracticalimplicationsofAIgaining“selfawareness”

?Longtermexistentialrisks:Issuesaroundthepotentiallossofhumancontroloverhighlycapablemodelsoramisalign