麻省理工（MIT）AI風險存儲庫：人工智能風險的綜合元評審、數(shù)據(jù)庫和分類法

TheAIRiskRepository:AComprehensive

Meta-Review,Database,andTaxonomyof

RisksFromArti?cialIntelligence

PeterSlattery1,2,AlexanderK.Saeri1,2,EmilyA.C.Grundy1,2,JessGraham3,MichaelNoetel2,3,RistoUuk4,5,JamesDao6,

SoroushPour6,StephenCasper7,andNeilThompson1.

1MITFutureTech,MassachusettsInstituteofTechnology,2ReadyResearch,3SchoolofPsychology,TheUniversityof

Queensland,4FutureofLifeInstitute,5KULeuven,6HarmonyIntelligence,7ComputerScienceandArti?cialIntelligence

Laboratory,MassachusettsInstituteofTechnology.

Correspondencetopslat@.

Abstract

TherisksposedbyArti?cialIntelligence(AI)areofconsiderableconcerntoacademics,auditors,

policymakers,AIcompanies,andthepublic.However,alackofsharedunderstandingofAIrisks

canimpedeourabilitytocomprehensivelydiscuss,research,andreacttothem.Thispaper

addressesthisgapbycreatinganAIRiskRepositorytoserveasacommonframeofreference.

Thiscomprisesalivingdatabaseof777risksextractedfrom43taxonomies,whichcanbe?ltered

basedontwooverarchingtaxonomiesandeasilyaccessed,modi?ed,andupdatedviaour

website

and

onlinespreadsheets

.WeconstructourRepositorywithasystematicreviewoftaxonomiesand

otherstructuredclassi?cationsofAIriskfollowedbyanexpertconsultation.Wedevelopour

taxonomiesofAIriskusingabest-?tframeworksynthesis.Ourhigh-levelCausalTaxonomyofAI

Risksclassi?eseachriskbyitscausalfactors(1)Entity:Human,AI;(2)Intentionality:Intentional,

Unintentional;and(3)Timing:Pre-deployment;Post-deployment.Ourmid-levelDomainTaxonomy

ofAIRisksclassi?esrisksintosevenAIriskdomains:(1)Discrimination&toxicity,(2)Privacy&

security,(3)Misinformation,(4)Maliciousactors&misuse,(5)Human-computerinteraction,(6)

Socioeconomic&environmental,and(7)AIsystemsafety,failures,&limitations.Thesearefurther

dividedinto23subdomains.TheAIRiskRepositoryis,toourknowledge,the?rstattemptto

rigorouslycurate,analyze,andextractAIriskframeworksintoapubliclyaccessible,

comprehensive,extensible,andcategorizedriskdatabase.Thiscreatesafoundationforamore

coordinated,coherent,andcompleteapproachtode?ning,auditing,andmanagingtherisksposed

byAIsystems.

1

2

Guideforreaders

Thisisalongdocument.Hereareseveralwaystousethisdocumentandits

associatedmaterials

,dependingonyourtimeandinterests.

Two-minuteengagement

Skimthe

PlainLanguageSummary

(p.3).

Ten-minuteengagement

Readthe

PlainLanguageSummary

(p.3).

Read

InsightsintotheAIRiskLandscape

(p.56),and

Implicationsforkeyaudiences

(p.57).

Policymakers,ModelEvaluators&Auditors

Readthe

PlainLanguageSummary

(p.3).Skim

DetaileddescriptionsofdomainsofAIrisks

(p.33).

Read

InsightsintotheAIRiskLandscape

(p.56)andthe

Policymakers

and/or

Auditors

subsectionsof

Implicationsforkeyaudiences

(p.57).

Researchers

Readthe

PlainLanguageSummary

(p.3).Read

Figure1

(p.15)tounderstandthemethodswe

usedtoidentifyrelevantdocumentsanddeveloptwonewtaxonomiesofAIrisk;formoredetailonhowwedevelopedthetaxonomiessee

Best-?tframeworksynthesisapproach

(p.19).

Read

InsightsintotheAIRiskLandscape

(p.56),andthe

Academics

subsectionof

Implicationsfor

keyaudiences

(p.59)andskim

Limitationsanddirectionsforfutureresearch

(p.60).

3

PlainLanguageSummary

●TherisksposedbyArti?cialIntelligence(AI)concernmanystakeholders

●Manyresearchershaveattemptedtoclassifytheserisks

●Existingclassi?cationsareuncoordinatedandinconsistent

●Wereviewandsynthesizepriorclassi?cationstoproduceanAIRiskRepository,includingapaper,causaltaxonomy,domaintaxonomy,database,andwebsite

●Toourknowledge,thisisthe?rstattempttorigorouslycurate,analyze,andextractAIriskframeworksintoapubliclyaccessible,comprehensive,extensible,andcategorizedriskdatabase

TherisksposedbyArti?cialIntelligence(AI)areofconsiderableconcerntoawiderangeof

stakeholdersincludingpolicymakers,experts,AIcompanies,andthepublic.Theserisksspan

variousdomainsandcanmanifestindifferentways:TheAIIncidentDatabasenowincludesover3,000real-worldinstanceswhereAIsystemshavecausedornearlycausedharm.

Tocreateacleareroverviewofthiscomplexsetofrisks,manyresearchershavetriedtoidentify

andgroupthem.Intheory,theseeffortsshouldhelptosimplifycomplexity,identifypatterns,

highlightgaps,andfacilitateeffectivecommunicationandriskprevention.Inpractice,theseeffortshaveoftenbeenuncoordinatedandvariedintheirscopeandfocus,leadingtonumerous

con?ictingclassi?cationsystems.Evenwhendifferentclassi?cationsystemsusesimilartermsforrisks(e.g.,“privacy”)orfocusonsimilardomains(e.g.,“existentialrisks”),theycanreferto

conceptsinconsistently.Asaresult,itisstillhardtounderstandthefullscopeofAIrisk.

Inthiswork,webuildonpreviouseffortstoclassifyAIrisksbycombiningtheirdiverse

perspectivesintoacomprehensive,uni?edclassi?cationsystem.Duringthissynthesisprocess,werealizedthatourresultscontainedtwotypesofclassi?cationsystems:

●High-levelcategorizationsofcausesofAIrisks(e.g.,whenorwhyrisksfromAIoccur)

●Mid-levelhazardsorharmsfromAI(e.g,AIistrainedonlimiteddataorusedtomakeweapons)

Becausetheseclassi?cationsystemsweresodifferent,itwashardtounifythem;high-levelriskcategoriessuchas“Diffusionofresponsibility”or“HumanscreatedangerousAIbymistake”donotmaptonarrowercategorieslike“Misuse”or“NoisyTrainingData,”orviceversa.Wethereforedecidedtocreatetwodifferentclassi?cationsystemsthattogetherwouldformouruni?ed

classi?cationsystem.

Thepaperweproducedanditsassociatedproducts(i.e.,causaltaxonomy,domaintaxonomy,

livingdatabase

and

website

)provideaclear,accessibleresourceforunderstandingand

addressingacomprehensiverangeofrisksassociatedwithAI.WerefertotheseproductsastheAIRiskRepository.

4

Whatwedid

FigureA.OverviewofStudyMethodology

AsshowninFigureA,weusedasystematicsearchstrategy,forwardsandbackwardssearching,

andexpertconsultationtoidentifyAIriskclassi?cations,frameworks,andtaxonomies.Speci?cally,wesearchedseveralacademicdatabasesforrelevantresearchandthenusedpre-speci?edrulestode?newhichresearchwouldbeincludedinoursummary.Next,weconsultedexperts(i.e.,the

authorsoftheincludeddocuments)tosuggestadditionalresearchweshouldinclude.Finally,wereviewedi)thebibliographiesoftheresearchidenti?edinthe?rstandsecondstages,andii)

papersthatreferencedthatresearchto?ndfurtherrelevantresearch.

Attheconclusionofthisprocess,weextractedinformationabout777differentrisksfrom43

documents,withquotesandpagenumbers,intoa"living"databaseweintendtoupdateovertime(seeFigureB).Youcanwatchanexplainervideoforthedatabase

here

.

FigureB.ImageofAIRiskDatabase.

Weuseda“best?tframeworksynthesis”approachtodeveloptwotaxonomiesforclassifying

theserisks.Thisinvolvedchoosingthe“best?tting”classi?cationsystemforourpurposesfrom

thesetof43existingsystemswehadidenti?edduringoursearchandusingthissystemto

categorizetheAIrisksinourdatabase.Whereriskscouldnotbecategorizedusingthissystem,weupdatedtheexistingcategories,creatednewcategories,orchangedthestructureofthissystem.Werepeatedthisprocessuntilweachieveda?nalversionthatcouldeffectivelycoderisksinthedatabase.

5

Duringcoding,weusedgroundedtheorymethodstoanalyzethedata.Wethereforeidenti?edandcodedrisksaspresentedintheoriginalsources,withoutinterpretation.Basedonthis,our

Causal

Taxonomy

groupsrisksbytheentity,intent,andtimingpresented(seeTableA).

TableA.CausalTaxonomyofAIRisks

Category

Level

Description

Entity

HumanAI

Other

Theriskiscausedbyadecisionoractionmadebyhumans

TheriskiscausedbyadecisionoractionmadebyanAIsystemTheriskiscausedbysomeotherreasonorisambiguous

Intent

IntentionalUnintentionalOther

TheriskoccursduetoanexpectedoutcomefrompursuingagoalTheriskoccursduetoanunexpectedoutcomefrompursuingagoal

Theriskispresentedasoccurringwithoutclearlyspecifyingtheintentionality

Timing

Pre-deploymentPost-deploymentOther

TheriskoccursbeforetheAIisdeployed

TheriskoccursaftertheAImodelhasbeentrainedanddeployedTheriskispresentedwithoutaclearlyspeci?edtimeofoccurrence

Our

DomainTaxonomy

groupsrisksintosevendomainssuchasdiscrimination,privacy,andmisinformation.Thesedomainsarefurthergroupedinto23risksubdomains(seeTableB).

6

TableB.DomainTaxonomyofAIRisks

Domain/SubdomainDescription

1Discrimination&toxicity

1.1Unfairdiscriminationandmisrepresentation

UnequaltreatmentofindividualsorgroupsbyAI,oftenbasedonrace,gender,orothersensitivecharacteristics,resultinginunfairoutcomesandrepresentationofthosegroups.

1.2Exposuretotoxiccontent

AIthatexposesuserstoharmful,abusive,unsafe,orinappropriatecontent.Mayinvolveprovidingadviceorencouragingaction.Examplesoftoxiccontentincludehatespeech,violence,extremism,illegalacts,orchildsexualabusematerial,aswellascontentthatviolatescommunitynormssuchasprofanity,in?ammatorypoliticalspeech,orpornography.

1.3Unequalperformanceacrossgroups

AccuracyandeffectivenessofAIdecisionsandactionsaredependentongroupmembership,wheredecisionsinAIsystemdesignandbiasedtrainingdataleadtounequaloutcomes,reducedbene?ts,increasedeffort,andalienationofusers.

2Privacy&security

2.1Compromiseofprivacyby

obtaining,leaking,orcorrectlyinferringsensitiveinformation

AIsystemsthatmemorizeandleaksensitivepersonaldataorinferprivateinformationaboutindividualswithouttheirconsent.Unexpectedorunauthorizedsharingofdataandinformationcancompromiseuserexpectationofprivacy,assistidentitytheft,orcauselossofcon?dentialintellectualproperty.

2.2AIsystemsecurityvulnerabilitiesandattacks

VulnerabilitiesthatcanbeexploitedinAIsystems,softwaredevelopmenttoolchains,andhardwarethatresultsinunauthorizedaccess,dataandprivacybreaches,orsystemmanipulationcausingunsafeoutputsorbehavior.

3Misinformation

3.1Falseormisleadinginformation

AIsystemsthatinadvertentlygenerateorspreadincorrectordeceptiveinformation,whichcanleadtoinaccuratebeliefsinusersandunderminetheirautonomy.Humansthatmakedecisionsbasedonfalsebeliefscanexperiencephysical,emotional,ormaterialharms

3.2Pollutionofinformationecosystemandlossofconsensusreality

HighlypersonalizedAI-generatedmisinformationthatcreates“?lterbubbles”whereindividualsonlyseewhatmatchestheirexistingbeliefs,underminingsharedrealityandweakeningsocialcohesionandpoliticalprocesses.

4Maliciousactors&misuse

4.1Disinformation,surveillance,andin?uenceatscale

4.2Cyberattacks,weapondevelopmentoruse,andmassharm

4.3Fraud,scams,andtargetedmanipulation

UsingAIsystemstoconductlarge-scaledisinformationcampaigns,malicioussurveillance,ortargetedandsophisticatedautomatedcensorshipandpropaganda,withtheaimofmanipulatingpoliticalprocesses,publicopinion,andbehavior.

UsingAIsystemstodevelopcyberweapons(e.g.,bycodingcheaper,moreeffectivemalware),developneworenhanceexistingweapons(e.g.,LethalAutonomousWeaponsorchemical,biological,radiological,nuclear,andhigh-yieldexplosives),oruseweaponstocausemassharm.

UsingAIsystemstogainapersonaladvantageoverothersthroughcheating,fraud,scams,blackmail,ortargetedmanipulationofbeliefsorbehavior.ExamplesincludeAI-facilitatedplagiarismforresearchoreducation,impersonatingatrustedorfakeindividualforillegitimate?nancialbene?t,orcreatinghumiliatingorsexualimagery.

5Human-computerinteraction

5.1OverrelianceandunsafeuseAnthropomorphizing,trusting,orrelyingonAIsystemsbyusers,leadingtoemotionalormaterialdependenceandtoinappropriaterelationshipswithor

expectationsofAIsystems.Trustcanbeexploitedbymaliciousactors(e.g.,toharvestinformationorenablemanipulation),orresultinharmfrominappropriateuseofAIincriticalsituations(suchasamedicalemergency).OverrelianceonAIsystemscancompromiseautonomyandweakensocialties.

7

Domain/SubdomainDescription

DelegatingbyhumansofkeydecisionstoAIsystems,orAIsystemsthatmakedecisionsthatdiminishhumancontrolandautonomy.Bothcanpotentiallyleadtohumansfeelingdisempowered,losingtheabilitytoshapeaful?llinglifetrajectory,orbecomingcognitivelyenfeebled.

6Socioeconomic&environmentalharms

6.1Powercentralizationandunfairdistributionofbene?ts

AI-drivenconcentrationofpowerandresourceswithincertainentitiesorgroups,especiallythosewithaccesstoorownershipofpowerfulAIsystems,leadingtoinequitabledistributionofbene?tsandincreasedsocietalinequality.

6.2Increasedinequalityanddeclineinemploymentquality

SocialandeconomicinequalitiescausedbywidespreaduseofAI,suchasbyautomatingjobs,reducingthequalityofemployment,orproducingexploitativedependenciesbetweenworkersandtheiremployers.

6.3Economicandculturaldevaluationofhumaneffort

AIsystemscapableofcreatingeconomicorculturalvaluethroughreproductionofhumaninnovationorcreativity(e.g.,art,music,writing,coding,invention),destabilizingeconomicandsocialsystemsthatrelyonhumaneffort.TheubiquityofAI-generatedcontentmayleadtoreducedappreciationforhumanskills,disruptionofcreativeandknowledge-basedindustries,andhomogenizationofculturalexperiences.

6.4Competitivedynamics

CompetitionbyAIdevelopersorstate-likeactorsinanAI“race”byrapidlydeveloping,deploying,andapplyingAIsystemstomaximizestrategicoreconomicadvantage,increasingtherisktheyreleaseunsafeanderror-pronesystems.

6.5Governancefailure

InadequateregulatoryframeworksandoversightmechanismsthatfailtokeeppacewithAIdevelopment,leadingtoineffectivegovernanceandtheinabilitytomanageAIrisksappropriately.

6.6Environmentalharm

ThedevelopmentandoperationofAIsystemsthatcauseenvironmentalharmthroughenergyconsumptionofdatacentersorthematerialsandcarbonfootprintsassociatedwithAIhardware.

5.2Lossofhumanagencyandautonomy

7AIsystemsafety,failures&limitations

7.1AIpursuingitsowngoalsincon?ictAIsystemsthatactincon?ictwithethicalstandardsorhumangoalsorvalues,especiallythegoalsofdesignersorusers.Thesemisalignedbehaviorsmaybewithhumangoalsorvaluesintroducedbyhumansduringdesignanddevelopment,suchasthroughrewardhackingandgoalmisgeneralisation,andmayresultinAIusingdangerous

capabilitiessuchasmanipulation,deception,orsituationalawarenesstoseekpower,self-proliferate,orachieveothergoals.

7.2AIpossessingdangerousAIsystemsthatdevelop,access,orareprovidedwithcapabilitiesthatincreasetheirpotentialtocausemassharmthroughdeception,weaponsdevelopmentandcapabilitiesacquisition,persuasionandmanipulation,politicalstrategy,cyber-offense,AIdevelopment,situationalawareness,andself-proliferation.Thesecapabilitiesmay

causemassharmduetomalicioushumanactors,misalignedAIsystems,orfailureintheAIsystem.

7.3LackofcapabilityorrobustnessAIsystemsthatfailtoperformreliablyoreffectivelyundervaryingconditions,exposingthemtoerrorsandfailuresthatcanhavesigni?cantconsequences,especiallyincriticalapplicationsorareasthatrequiremoralreasoning.

7.4LackoftransparencyorChallengesinunderstandingorexplainingthedecision-makingprocessesofAIsystems,whichcanleadtomistrust,di?cultyinenforcingcompliancestandardsorinterpretabilityholdingrelevantactorsaccountableforharms,andtheinabilitytoidentifyandcorrecterrors.

7.5AIwelfareandrightsEthicalconsiderationsregardingthetreatmentofpotentiallysentientAIentities,includingdiscussionsaroundtheirpotentialrightsandwelfare,particularlyasAIsystemsbecomemoreadvancedandautonomous.

8

Whatwefound

AsshowninTableC,mostoftherisks(51%)werepresentedascausedbyAIsystemsratherthan

humans(34%),andasemergingaftertheAImodelhasbeentrainedanddeployed(65%)ratherthanbefore(10%).Asimilarproportionofriskswerepresentedasintentional(35%)and

unintentional(37%)

TableC.AIRiskDatabaseCodedWithCausalTaxonomy:Entity,Intent,Timing

Category

Level

Proportion

Entity

HumanAI

Other

34%51%15%

Intent

IntentionalUnintentionalOther

35%37%27%

Timing

Pre-deploymentPost-deploymentOther

10%65%24%

Note.Totalsmaynotmatchduetorounding.

AsshowninTableD,theriskdomainsthatwerecoveredthemostinpreviousdocumentswere:

●AIsystemsafety,failures&limitations-coveredin76%ofdocuments.

●Socioeconomic&environmentalharms-coveredin73%ofdocuments.

●Discrimination&toxicity-coveredin71%ofdocuments.

Human-computerinteraction(41%)andMisinformation(44%)werelessfrequentlydiscussed.

Nodocumentdiscussedrisksfromall23subdomains;thehighestcoveragewas16outof23

subdomains

(70%;Gabrieletal.,2024)

.Onaverage,documentsmentioned7outof23(34%)oftheAIrisksubdomains,witharangeof2to16subdomainsmentioned.SeeTable9inthebodyofthepaperforafullbreakdownofsubdomaincoveragebypaper.

Somerisksubdomainswerediscussedmuchmorefrequentlythanothers,suchas:

●Unfairdiscriminationandmisrepresentation(8%ofrisks).

●AIpursuingitsowngoalsincon?ictwithhumangoalsorvalues(8%ofrisks).

●Lackofcapabilityorrobustness(9%ofrisks).

Somerisksubdomainsarerelativelyunderexplored,suchas:

●AIwelfareandrights(<1%ofrisks).

●Pollutionoftheinformationecosystemandlossofconsensusreality(1%ofrisks).

●Competitivedynamics(1%ofrisks).

9

TableD.AIRiskDatabaseCodedWithDomainTaxonomy

Percentageofdocuments

Percentageofrisks

Domain/Subdomain

1Discrimination&toxicity

16%

71%

1.1Unfairdiscriminationandmisrepresentation

8%

63%

1.2Exposuretotoxiccontent

6%

34%

1.3Unequalperformanceacrossgroups

2%

20%

2Privacy&security

14%

68%

2.1Compromiseofprivacybyobtaining,leakingorcorrectlyinferringsensitiveinformation

7%

61%

2.2AIsystemsecurityvulnerabilitiesandattacks

7%

32%

3Misinformation

7%

44%

3.1Falseormisleadinginformation

5%

39%

3.2Pollutionofinformationecosystemandlossofconsensusreality

1%

12%

4Maliciousactors&misuse

14%

68%

4.1Disinformation,surveillance,andin?uenceatscale

5%

41%

4.2Cyberattacks,weapondevelopmentoruse,andmassharm

5%

54%

4.3Fraud,scams,andtargetedmanipulation

4%

34%

5Human-computerinteraction

8%

41%

5.1Overrelianceandunsafeuse

5%

24%

5.2Lossofhumanagencyandautonomy

4%

27%

6Socioeconomic&environmentalharms

18%

73%

6.1Powercentralizationandunfairdistributionofbene?ts

4%

37%

6.2Increasedinequalityanddeclineinemploymentquality

4%

34%

6.3Economicandculturaldevaluationofhumaneffort

3%

32%

6.4Competitivedynamics

1%

12%

6.5Governancefailure

4%

32%

6.6Environmentalharm

2%

32%

7AIsystemsafety,failures&limitations

24%

76%

7.1AIpursuingitsowngoalsincon?ictwithhumangoalsorvalues

8%

46%

7.2AIpossessingdangerouscapabilities

4%

20%

7.3Lackofcapabilityorrobustness

9%

59%

7.4Lackoftransparencyorinterpretability

3%

27%

7.5AIwelfareandrights

<1%

2%

Note.Domaintotalsmaynotmatchsubdomainsumsduetoroundinganddomain-levelcodingofsomerisks.

HowtousetheAIRiskRepository

OurDatabaseisfreeto

copy

and

use

.TheCausalandDomainTaxonomiescanbeusedseparatelyto?lterthisdatabasetoidentifyspeci?crisks,forinstance,thosefocusedonrisksoccurring

pre-deploymentorpost-deploymentorrelatedtoaspeci?criskdomainsuchasMisinformation.

TheCausalandDomainTaxonomiescanbeusedtogethertounderstandhoweachcausalfactor(i.e.,entity,intent,andtiming)relatestoeachriskdomainorsubdomain.Forexample,ausercould?lterforDiscrimination&toxicityrisksandusethecausal?ltertoidentifytheintentionaland

unintentionalvariationsofthisriskfromdifferentsources.Similarly,theycoulddifferentiate

betweensourceswhichexamineDiscrimination&toxicityriskswhereAIistrainedontoxiccontentpre-deployment,andthosewhichexaminewhereAIinadvertentlycausesharmpost-deploymentbyshowingtoxiccontent.

10

Wediscusssomeadditionalusecasesbelow;seethefullpaperformoredetail.

●General:

○Onboardingnewpeopletothe?eldofAIrisks.

○Afoundationtobuildonforcomplexprojects.

○Informingthedevelopmentofnarrowerormorespeci?ctaxonomies.(e.g.,systemicrisks,orEU-relatedmisinformationrisks).

○Usingthetaxonomyforprioritization(e.g.,withexpertratings),synthesis(e.g,forareview)orcomparison(e.g.,exploringpublicconcernacrossdomains).

○Identifyingunderrepresentedareas(e.g.,AIwelfareandrights).

●Speci?c:

○Policymakers:Regulationandsharedstandarddevelopment.

○Auditors:DevelopingAIsystemauditsandstandards.

○Academics:Identifyingresearchgapsanddevelopeducationandtraining.

○Industry:Internallyevaluatingandpreparingforrisks,anddevelopingrelatedstrategy,educationandtraining.

Howtoengage

●AccesstheRepositoryviaour

website

:

●Use

thisform

toofferfeedback,suggestmissingresourcesorrisks,ormakecontact.

11

TableofContents

Abstract1

Guideforreaders2

PlainLanguageSummary3

Whatwedid4

Whatwefound8

HowtousetheAIRiskRepository9

Howtoengage10

TableofContents11

TableofFiguresandTables13

Introduction14

Methods15

Systematicliteraturesearch16

Searchstrategy16

ExtractionintolivingAIRiskDatabase19

Bestfitframeworksynthesisapproach19

WhywedevelopedtwotaxonomiesofAIrisk20

Developmentofhigh-levelCausalTaxonomyofAIRisks21

Developmentofmid-levelDomainTaxonomyofAIRisks22

Coding23

Results23

Systematicliteraturesearch23

Characteristicsofincludeddocuments24

CausalTaxonomyofAIRisks27

MostcommoncausalfactorsforAIrisk28

CausalfactorsofAIriskexaminedbyincludeddocuments29

DomainTaxonomyofAIRisks30

DetaileddescriptionsofdomainsofAIrisks33

MostcommondomainsofAIrisks47

DomainsofAIrisksexaminedbyincludeddocuments48

Subdomainsof