ExecutiveSummary

Inourpastpublication,“AnArgumentforHybridAIIncidentReporting,”weproposedimplementingafederated

*

andcomprehensiveartificialintelligenceincidentreportingframeworktosystematicallyrecord,analyze,andrespondtoAIincidents

.1

Thehybridreportingframeworkproposesimplementingmandatory,voluntary,andcitizen

reportingmechanisms.Thisdocumentdescribesthecriticalcontentthatshouldbe

includedinamandatoryAIincidentreportingregimeandshouldalsoinformvoluntaryandcitizenreportingefforts.

Inthispublication,wedefineasetofstandardizedkeycomponentsofAIincidentsthatcanbeusedasareportingtemplatetocollectvitalAIincidentdata.Thesecomponentsinclude,butarenotlimitedto,informationaboutthetypeofAIincident,thenatureandseverityofharm,technicaldata,affectedentitiesandindividuals,andthecontextandcircumstanceswithinwhichtheincidentunfolded.Whileintentionallyhighlevel,ourproposedsetofcomponentsdistillsinformationfromexistingAIinitiativesthattrackreal-worldevents,harms,andrisksrelatedtoAI,andincorporateslessonslearned

fromincidentreportingsystemsandpracticesinthetransportation,healthcare,andcybersecuritysectors.

Ifadoptedandusedwidelyandconsistentlybygovernments,regulators,professionalorganizations,developers,andresearchers,thesereportingcomponentscanhelp

enhanceAIsafetyandsecuritymeasuresby:

●FacilitatingconsistentdatacollectionofAIincidents

●Promotingtracking,monitoring,research,andinformationsharingofAIincidents

●EnhancingknowledgearoundAI-relatedharmsandrisks

●EnsuringthatessentialAIincidentdataiscollectedtopreventreportinggaps

●BuildingafoundationalframeworkforagileincidentreportingthatadaptstoAIadvancements

Tofullyutilizethebenefitsofthislistofcomponents,werecommendpublishingmandatoryAIincidentreportingformatsbasedonthemandestablishingan

independentinvestigativeagencytouncoverincidentdatathatmaynotbe

immediatelydiscernibleatthetimeofreporting.ThislistcanalsoserveasatemplatefordesirabledisclosureguidelinesofincidentdataforvoluntaryandcitizenAIincidentreportingsystems.

*Wedefineafederatedframeworkasacentralizedframeworkprescribedbyasingularauthoritativegovernmentbodyorthefederalgovernment.Theframeworkstipulatesasetofminimumrequirementsthatcanbeadaptedandimplementedacrossgovernmentagenciesornongovernmentalorganizations.

CenterforSecurityandEmergingTechnology|2

CenterforSecurityandEmergingTechnology|3

TableofContents

ExecutiveSummary 2

Introduction 4

KeyComponentsofAIIncidents 7

SynthesizingKeyComponentsofAIIncidents 10

TypesofEvents 10

AIIncidentsandNearMisses 10

HarmDimensions 11

TypeofHarm 11

MechanismofHarm 11

SeverityFactor 12

TechnicalData 14

Context,Circumstances,andStakeholders 14

Post-incidentData 16

PolicyRecommendations 17

PublishAIIncidentReportingFormats 17

EstablishanIndependentInvestigationAgency 17

Conclusion 18

Appendix 19

DecomposingAIIncidents 19

AIHarmEventsasaSpectrum 21

DocumentingAIHarm:TheManyImpactsofAI 22

SignificantbutLimitedInformation 23

IncidentComponentsReportedinOtherSectors 24

SharedIncidentComponents 25

AdditionalKeyComponents 25

MeasuringSeverity:AFirstGlimpse 26

Authors 27

Acknowledgments 27

Endnotes 28

Introduction

Artificialintelligenceincidentshavebeenoccurringwithwide-rangingadverse

impacts.Currently,onlyafewindependentdatabasesdocumentAIincidents

.2

Thesedatabasesprimarilyrelyuponnewsreportsandotherpubliclyavailableincident

reportsfrom“popular,trade,andacademicpress.

”3

Thereisnotyetanyfederated

*

policyframeworkestablishedtofacilitatesystematicandcomprehensiveAIincidentreportingpractices.AsAIcontinuestoexpanditscapabilities,thereisanurgentneedtosystemicallycollectAIincidentdatatoenhanceourknowledgeofAI-relatedharmandhelpusdevelopsafe,secure,andtrustworthyAIsystems

.4

Toaddressthiscriticalgap,inourpreviouspaper,“AnArgumentforHybridAIIncidentReporting,”weproposedestablishingpoliciesforafederated,comprehensive,and

standardizedAIincidentreportingframework

.5

Wefoundthatimplementingconsistentandcomprehensivereportinganddocumentationofincidentscanhelpcapturemore

completeandusefuldata.Asmoredataisgathered,itcanrevealvitaltrendsinAI

incidentsandprovidegreaterclarityontheirseverity.Overtime,thiscanallowustomoreaccuratelyassesstheimpactandeffectivenessofdifferentAIsafetyandsecuritypoliciesandmeasures.

ThefederatedAIincidentreportingframeworkweoutlinedinourpreviouspapercallsforacombinationofmandatory,voluntary,andcitizenreportingtoanindependent

externalentity(suchasagovernmentagency,professionalassociation,oroversightbody)topromotetransparencyandaccountabilityinAIincidentmanagement.

Governmentagencies,regulators,professionalassociations,andoversightbodiesthatwillbeimplementingsuchaframeworktorecord,document,andmonitorAIincidentswouldbenefitfromhavingaconsistentsetofcomponentsthatcanconveyvital

informationabouttheAIincidentathand.Inthispaper,wepresentpreciselysuchalistthatshouldbeusedwhencollectingincidentdataunderamandatoryincident

reportingregime,andthatmayalsobeusedtoinformvoluntaryandcitizenAIincidentreportingregimes.Thislistdrawsonandsynthesizeshigh-levelcomponentsfrom

variousinitiativesfocusedonAIincidents,harms,andrisks,andalsoincludesbestpracticesinincidentreportingfromthetransportation,healthcare,andcybersecuritysectors.

EstablishingafundamentallistofkeycomponentsofAIincidentscanprovide

numerousbenefitsforpolicymakers,governmentagencies,oversightcommittees,civic

*Wedefineafederatedframeworkasacentralizedframeworkprescribedbyasingularauthoritativegovernmentbodyorthefederalgovernment.Theframeworkstipulatesasetofminimumrequirementsthatcanbeadaptedandimplementedacrossgovernmentagenciesornongovernmentalorganizations.

CenterforSecurityandEmergingTechnology|4

CenterforSecurityandEmergingTechnology|5

organizations,developers,andresearchersinrecording,documenting,andmonitoringAIincidentdata,suchas:

●FacilitatingconsistentdatacollectionofAIincidents

●Promotingtracking,monitoring,research,andinformationsharingofAIincidents

●EnhancingknowledgearoundAI-relatedharmsandrisks

●EnsuringthatessentialAIincidentdataiscollectedtopreventreportinggaps

●BuildingafoundationalframeworkforagileincidentreportingthatadaptstoAIadvancements

TheoutcomesfromtheseactionscanaidinenhancingAIsafetyandsecuritymeasures.

Toachievethesebenefits,stakeholderscanusethisstandardizedlistofkey

componentsasatemplateforAIincidentreporting.Forinstance,regulatorscandesignandtailorAIincidentreportingsystems,templates,andformatstosuitthespecific

needsoftheirrespectivedomainsbasedonthelistofkeycomponents.Additionally,thelistcanserveastheminimumanddesirabledisclosurerequirementsofincidentdatainmandatoryandvoluntaryAIincidentreportingsystems,respectively.

WebeginwithanoverviewofthesynthesizedlistofkeycomponentsofAIincidents,continuewithadiscussionoftheselectionofthecomponents,andconcludewith

policyrecommendations.

CenterforSecurityandEmergingTechnology|6

Definitions:ThesedefinitionswereadoptedfromDefiningAIIncidentsandRelatedTerms,publishedbytheOrganisationforEconomicCo-operationandDevelopment(OECD)

.*

AIincident:Anevent,circumstance,orseriesofeventswherethedevelopment,use,ormalfunctionofoneormoreAIsystemsdirectlyorindirectlyleadstoanyofthe

followingharms:

a)injuryorharmtothehealthofapersonorgroupsofpeople;

b)disruptionofthemanagementandoperationofcriticalinfrastructure;

c)violationsofhumanrightsorabreachofobligationsundertheapplicablelawintendedtoprotectfundamental,labor,andintellectualpropertyrights;or

d)harmtoproperty,communities,ortheenvironment

.6

AInearmiss:Anevent,circumstance,orseriesofeventswherethedevelopment,use,ormalfunctionofoneormoreAIsystemscouldhavedirectlyorindirectlyledtoanyofthefollowingharms,butfailedtobychanceorwasintercepted:

a)injuryorharmtothehealthofapersonorgroupsofpeople;

b)disruptionofthemanagementandoperationofcriticalinfrastructure;

c)violationsofhumanrightsorabreachofobligationsundertheapplicablelawintendedtoprotectfundamental,labor,andintellectualpropertyrights;or

d)harmtoproperty,communities,ortheenvironment

.7

AIhazard:Anevent,circumstance,orseriesofeventswherethedevelopment,use,ormalfunctionofoneormoreAIsystemscouldplausiblyleadtoanyofthefollowingharms:

a)injuryorharmtothehealthofapersonorgroupsofpeople;

b)disruptionofthemanagementandoperationofcriticalinfrastructure;

c)violationstohumanrightsorabreachofobligationsundertheapplicablelawintendedtoprotectfundamental,labor,andintellectualpropertyrights;or

d)harmtoproperty,communities,ortheenvironment

.8

Forthesakeofbrevity,inthispaperwewilluse“AIincident”asanoverarchingterm

encompassingbothincidentsandnearmisses,exceptwhereadistinctionis

specificallynoted.Thisterminologyalignswiththepaper’sfocusonamandatoryreportingframework,whichdoesnotextensivelydiscussconceptssuchasAIrisks,hazards,andissues,aswedonotconsiderthempertinenttosucharegime(see

Appendix).

KeyComponentsofAIIncidents

TocreatealistofkeycomponentsofAIincidents,weanalyzedandsynthesizeda

diversearrayofhigh-levelcomponentsfrommultipleinitiativesdiscussingAIincidentsandrelatedconcepts(seeTableA1).TheprocessentailedexaminingexistingAI

incidentdatabasestoevaluatetheiremphasisondifferentcomponents,aswellastherationalebehindtheirvariedapproachestocollectionandclassificationmethods.

Additionally,weanalyzedincidentreportingsystemsfromotherhigh-impactsectors,includingtransportation,healthcare,andcybersecurity,toidentifycommonalitiesandgapsthatcouldfurtherbolsterourselectionofkeycomponentsofAIincidents(seeTableA2).

Thekeycomponentsincludedinourlistaremeanttoconveythemostrelevant,

meaningful,andusefulinformationaboutAIincidents.Pragmatic,adaptable,andefficient(ratherthansophisticated)taxonomiesaremorelikelytoencourageand

enhanceincidentdatacollection

.9

Therefore,weusedthefollowingcriteriatoguideouranalysisandselection,ensuringthatthekeycomponentsofAIincidentsare:

?Easytouse,understand,andimplementacrossawiderangeofsectorsandapplicationsbydiversestakeholders,includingpolicymakers,governmentagencies,andoversightcommittees

?AdaptabletotheemergingcapabilitiesofAIanddiversesocietalfactors

?FunctionalforprovidingdataforpolicymakerstocraftsafetyandsecuritymeasuresandforresearcherstoanalyzeAIharms

Table1providesanoverviewofthekeycomponents,alongwithelements,we

identified.ThesecomponentsindicatetheAIincidentinformationthatshouldbe

reportedanddocumented.Certainincidentdatamaynotnecessarilybeimmediately

apparentatthetimeofreportingandmayrequirefurtherinvestigationandassessment.

*Inthispaper,wedifferentiatebetweenAInearmissandAIhazard,whichtheOECDdefinitiongroupstogether.Formoredetails,seeOECD,DefiningAIIncidentsandRelatedTerms(Paris:OECD,2024),13,

/10.1787/d1a8d965-en.

CenterforSecurityandEmergingTechnology|7

CenterforSecurityandEmergingTechnology|8

Table1.KeyComponentsofAIIncidents

KeyComponents

Elements

Typeofevent

Didharmoccurornearlyoccur?

Incident

Nearmiss

Typeofharm

Whattypesofharm?

Physical

Environmental

Economicandfinancial

Reputational

Publicinterest

Humanrightsandfundamentalrights

Psychological

Mechanismofharm

Whatwerethecontributingfactors?

Technicalfactors

Otherfactors

Severityfactors

Howimpactfulwastheincident?

Remediability

Levelofseverity

Distributionofharm

Exposedpopulationsize

Duration

Optionality

Frequency

Technicalinformation

WhatwerethetechnicaldimensionsoftheimplicatedAIsystems?

AIsystemcard

AImodelcard

Datasheet

CenterforSecurityandEmergingTechnology|9

Contextandcircumstances

Goalsandapplicationpurpose

Sector

Startandenddate

Location

Reporter

Existingsafeguardsandpolicies

Entitiesandindividuals

Whowasinvolvedandaffected?

AIactors

Affectedstakeholders

Incidentresponse

Mitigation,termination,etc.

Ethicalimpact

EthicalimpactassessmentusingtheUnitedNationsEducational,ScientificandCulturalOrganization(UNESCO)ethicalimpactassessmenttoolortheOECD“AIPrinciples

”10

Theelementsoutlinedarenotexhaustiveandcanberefinedtomeetdomain-specificneeds.Regulatorscandefinethedesiredgranularityofincidentdatacollectionthroughvariousresponseformats,suchasmultiple-choiceoptions,open-textfields,andratingscales.Forexample,technicalfactorsunder“mechanismofharm”couldpresentalistofpossiblefactors,suchasbiasintrainingdata,dataqualityissues,andmodeldrift.

The"typeofharm"couldbedocumentedthroughacombinationofmultiple-choicecategoriesandacorrespondingtextualdescription.

WhilethesekeycomponentsaimtodocumentAIincidentswithinamandatory

reportingregime,theycanalsoserveasanominaltemplateforvoluntaryandcitizenreporting.Bothmandatoryandvoluntarysystems,whicharedirectedatAIactorsandoversightgroups,shouldincludethecompletesetofkeycomponents(withoptional

inputsforvoluntaryreporting).Citizenreportingformatsmayadoptanabbreviated

version,focusingonessentialcomponents,suchasincidentdescription,AIsystem,andcircumstantialdetails.

CenterforSecurityandEmergingTechnology|10

SynthesizingKeyComponentsofAIIncidents

FromourexaminationoftheAIinitiativesandofreportingsystemsfromthe

transportation,healthcare,andcybersecuritysectors,wedistilledandsynthesizedkeycomponentsofAIincidentsthatpresentedseveraloverarchingthemes:typesof

events,harmdimensions,technicaldata,contextandcircumstances,andpost-incident

data.

TypesofEvents

AIIncidentsandNearMisses

ThefirststepindocumentingAIincidentsunderanyreportingregimeistodetermine

whethertheeventmeetsthecriteriaanddefinitionsforincidentssubjecttoreportingunderthatregime.Inthecontextofamandatoryreportingregime,wepropose

documentingtwotypesofevents:incidentsandnearmisses(seeTable2).

Table2.KeyComponent:TypeofEvent

KeyComponent

Elements

Description

Typeofevent

Didharmoccurornearlyoccur?

Incident

Harmoccurred.

Nearmiss

Harmnearlyoccurredbut

eitherwasavoidedbychanceorwasintercepted.

WeadopttheOECDdefinitionofanAIincident,whichis“anevent,circumstanceor

seriesofeventswherethedevelopment,useormalfunctionofoneormoreAIsystemsdirectlyorindirectlyleadsto”harm

.11

Forinstance,adeepfakeimageofanexplosionataU.S.federalgovernmentbuildingthatbrieflycausesthestockmarkettodipwould

beclassifiedasanAIincident

.12

NearmissesaresimilartoAIincidents,exceptharmeitherwasavoidedbychanceorwasintercepted

.13

Anearmissexamplewouldbeanautonomousvehiclefailingtostopatastopsign,butfortunatelynoothercarswerepresentattheintersection,avoidingapotentialaccident.

EventhoughtheOECDintegrates“nearmiss”underitsdefinitionofAIhazards,we

believeitiscrucialtodistinguishbetweenAIhazardsandnearmisses.Nearmissesarecharacterizedaseventsthatcouldhaveledtoaharmfuloutcomebutdidnot,either

duetochanceorbecausetheywereaverted

.14

Incontrast,hazardsaredescribedas

CenterforSecurityandEmergingTechnology|11

unsafeconditionsorcircumstancesthathavethepotentialtocauseharmbuthavenotyetresultedinanactualevent,oftenreferredtoas“accidentswaitingtohappen.”ThisdistinctionisvitalindefiningthescopeofAIincidentreportingpoliciesandenhancingAIsafetyandsecurityprotocolsrelatedtohazards.

AIincidentsandnearmissesshouldbeincludedwithinthescopeofmandatory

reporting.Reportingnearmissescanenhanceincidentdatacollection,astheseeventsexhibitsimilarcharacteristicstoincidents,apartfromtheiroutcomes.Inadditionto

aidingearlydetectionofnovelAIrisks,trackingnearmissescouldrevealvital

conditionsthatpreventedharmfromoccurring,whichcanbeleveragedtostrengthensafetymeasures.

Incontrast,AIhazards—whichareunsafeconditions—donotgenerallywarrant

mandatoryreportingoradheretothesameincidentreportingpoliciesanddisclosurerequirements.Forexample,voluntaryreportingmaybemoresuitableforAIhazards,allowingdevelopersandpolicymakerstousetheinformationtoimplementmore

effectiveguardrails,therebyreducingthelikelihoodofharmmaterializing.

HarmDimensions

TheharmdimensionsofanAIincidentconsistofseveralkeycomponents:typesof

harm,mechanismsofharm,andseverityfactors.

TypeofHarm

Amongtheinitiativesexamined,theOECDoffersthemostfunctionaland

comprehensivecategorizationofAI-relatedharms,includingphysical,environmental,economicorfinancial,reputational,publicinterest,humanandfundamentalrights,andpsychological

.15

Thecategorieseffectivelycoverthediversedescriptionsofharms

foundacrosstheinitiatives.Documentingthetypesofharmusingthesecategoriescanaidresearchersandpolicymakersinconductingsystematicanalysesoftheimpact

causedbyAIincidents.Often,anincidentcouldinvolvemorethanonetypeofharm.Forinstance,afacialrecognitiontechnologythatdisproportionatelymisidentifies

women,Black,Latino,andAsianshoppersas“l(fā)ikely”shoplifterscanresultinharmtofundamentalrights,psychologicalharm,andeconomiclosses

.16

MechanismofHarm

Reportingthemechanismofharmaimstocaptureallplausiblecontributingfactorsof

AIincidentsandnearmisses.Themechanismsofharmthatweobservedfromexistingincidentsandinitiativescanbegroupedintotwocategories:technicalfactorsandother

CenterforSecurityandEmergingTechnology|12

factors.Technicalfactorsrelatetosystemvulnerabilities,modeldrift,andsystemfailures

.17

Otherfactorscanincludehumanandcontextualfactors,suchasweakgovernance,usermisuse,andintentionalabusebyuser.Whiletechnicaland

nontechnicalfactorscancontributetoAIharmsimultaneously,wesuggest

documentingthesetwofactorsseparatelytodistinguishbetweenintentionalandunintentionalharms,aswellastoenhanceourunderstandingofhuman-AI

interaction

.18

Thedatagatheredfromthiscomponentwillbecrucialforconductingrootcauseanalysis.

SeverityFactor

Collectingdataonseverityfactorswillbeessentialforseverityassessmentsthatcan

aidinprioritizingmitigationefforts,formulatingappropriatesafetyandsecurity

policies,andenhancingriskassessments

.19

Althoughextremeharm,suchasdeathorirreversibledisruptionofcriticalinfrastructure,clearlyconstitutessignificantimpact,

theseverityofharmisoftennuancedandhighlycontextdependent

.20

Theperceivedseveritycanvarybasedonsocial,legal,andtemporalcontext,leadingtodiffering

assessmentsofharmacrossindividuals,organizations,andgovernments

.21

DevelopinganAIincidentseverityratingormetricframeworkwillrequirefurtherexplorationandwillnotbediscussedinthispaper.

Table3providesanoverviewofthekeycomponentsrelatedtotheharmdimensionsofanAIincident,alongwiththeirelementsanddescriptions.

Table3.KeyComponents:HarmDimensions

KeyComponents

Elements

Description

Typeofharm

Whattypesofharm?

Physical

Physicalinjuryanddeath

Environmental

E.g.,soilcontamination,waterpollution,andairpollution

Economicandfinancial

E.g.,financiallossordamages,harmtoproperty

Reputational

Mayaffectorganizationsandindividuals

Publicinterest

Involvescriticalinfrastructureanditsfunctionandthesocialfabricof

CenterforSecurityandEmergingTechnology|13

society

Humanrights

andfundamentalrights

Includesestablisheddomesticandinternationallaws

Psychological

E.g.,affectsmentalhealth

Mechanismofharm

Whatwerethecontributingfactors?

Technicalfactors

Malfunctions,systemvulnerabilities,datapoisoning,andconceptdrift

Otherfactors

Weakgovernance,lackof

safeguards,usermisuse,anduserabuse

Severityfactors

Howimpactfulwastheincident?

Remediability

Theabilitytorestorethoseaffectedtoasituationatleastequivalenttotheirsituationbeforetheimpac

t22

Levelofseverity

Howacutelytheharmimpactedthoseaffected

Distributionofharm

Whetheranindividual,group,orpopulationwasdisproportionatelyaffectedbytheharm

Exposed

populationsize

Afullestimateoftheadverselyimpactedstakeholders

Duration

Howlongtheharmwas

experiencedbytheaffectedstakeholders

Optionality

Users’orstakeholders’abilityto

accept,challenge,correct,oroptoutofthesystem’soutput

Frequency

Therateatwhichaffected

stakeholdersexperienceharm

CenterforSecurityandEmergingTechnology|14

TechnicalData

Next,weconsolidatedalltechnicaldataunderonekeycomponentandpropose

requiringAIactorstosubmitAIsystemormodelcardsanddatasheetsaspartofmandatoryreportingobligations(seeTable4)

.23

Table4.KeyComponent:TechnicalInformation

KeyComponent

Elements

Description

Technicalinformation

Whatwerethetechnical

dimensionsoftheimplicatedAIsystems?

AIsystemcard

Data,models,code,system

AImodelcard

Modeldetails,intendeduse,evaluationdata,andtrainingdata

Datasheet

Trainingdatainformationonmotivation,composition,

collectionprocess,

recommendeduse,andmore

Atthemoment,itischallengingtofullycapturethevitaltechnicaldimensionsofAI

harmeventsbecausesuchdataisneitherpubliclyavailablenorreadilydiscernible

fromnewsreports.IfstandardizedAIsystemandAImodelcardsarerequiredina

mandatoryreportingframework,researchersandpolicymakerswillhaveaccesstovitaltechnicaldata,whichcanimprovetechnicalstandardsandrisk-mitigationpractices

throughouttheAIlifecycle.Forinstance,theNationalTelecommunicationsand

InformationAdministrationrecommendsfederalagenciestoworkwithstakeholderstoimprovestandardinformationdisclosureswithinsuchartifacts

.24

Drawingfromthe

conceptofnutritionlabelsmandatedbytheFoodandDrugAdministration,these

artifactsshouldprovideessentialdetailsaboutAIsystems,models,andtrainingdata.

Context,Circumstances,andStakeholders

Table5listskeycomponentsrelatedtothecontextualandcircumstantialdata,alongwithinformationabouttherelevantAIactorsandaffectedstakeholders.These

componentsaredesignedtocapturethebroadersituationalandpreexistingconditionssurroundinganAIincident.ItincludesinformationsuchasthepurposeoftheAImodelorsystem,thesectorinwhichtheAIwasdeployed,thestartandenddateoftheevent(ifapplicable),locations,thereportingentityorindividual,andanyexistingsafeguards

CenterforSecurityandEmergingTechnology|15

andpolicies.Theidentityofthereportermayormaynotbedocumenteddependingonthereportingrequirements;incertainschemataanonymityisprovided.Documentingexistingsafeguardsandpoliciesenablesassessmentsofsafetyandsecuritymeasuresandidentifiesgapsforimprovement.InadditiontogatheringbackgrounddetailsabouttherelevantAIactorsandaffectedstakeholders,thedatashouldindicatewhethertheaffectedstakeholderswereusersornonusersoftheimplicatedAIsystem(s)toclarifytheirrelationshipwiththesystem(s).

Table5.KeyComponents:ContextandCircumstances

Key

Components

Elements

Description

Contextand

circumstances

Goalsand

applicationpurpose

E.g.,autonomousdriving,videogeneration,facerecognition

Sector

E.g.,healthcare,financial,agriculture

Startandenddate(ifapplicable)

Dateoffirstandlastknownincidents

Location

City,state,country

Reporter(ifrelevant)

Nameandinformationofpersonreporting

Existingsafeguardsandpolicies

Safeguardsorpoliciesthatwerealreadyinplacebeforetheincidentoccurred,if

applicable

Entitiesandindividuals

Whowas

involvedandaffected?

AIactors

DetailsofAIprovider,operator,anddeployer

Affected

stakeholders

Detailsofaffectedindividualsandentities,ifavailable,includingwhethertheywereusersornonusersoftheimplicatedAI

system(s)

CenterforSecurityandEmergingTechnology|16

Post-incidentData

Certaindata,suchastheincidentresponseandethicalimpact,isonlyavailableinthe

aftermathofanAIincident.

Recordingtheactionstakenafteranincidentoccursprovidesinformationon

organizationalincidentresponseplans,whichoffersmultiplebenefits(seeTable6).

ThispracticepromotestransparencyandaccountabilityamongAIactorswhile

facilitatingtheevaluationofincidentresponsestrategies.Additionally,documentingincidentresponsescontributestothedevelopmentofbestpracticesforaddressingAI-

relatedharms.TheNationalInstituteofStandardsandTechnology(NIST)furtherrecommendstrackingincidentresponsetimeasasafetymetricinAIrisk

management

.25

Impactassessmentshaveemergedasarelevantpolicyinstrumentforpromotingsafe,secure,andtrustworthyAIdevelopmentanddeployment.Numerousregulatoryandstandard-settingframeworks,includin