美國(guó)國(guó)土安全部發(fā)布《公共部門(mén)生成人工智能部署手冊(cè)》

DHSPlaybookforPublicSectorGenerativeArtificialIntelligence

Deployment

January2025

DHSPlaybookforPublicSectorGenAIDeployment(January2025)

Contents

LetterfromtheSecretaryandtheChiefArtificialIntelligenceOfficer3

Introduction:TheDHSGenAIPilots4

Pilot1:StrengtheningInvestigativeLeadswithLLM-EnhancedSearchandSummarization 4

Pilot2:HelpingLocalGovernmentsCreateHazardMitigationPlans 5

Pilot3:CreatingNovelTrainingOpportunitiesforImmigrationOfficers 5

Plays

Mission-Enhancing

GenAIUseCases

7

CoalitionBuildingandEffectiveGovernance

8

ToolsandInfrastructure

10

ResponsibleUseandTrustworthinessConsiderations1

1

MeasurementandMonitoring1

3

TrainingandTalentAcquisition1

4

UsabilityTestingandOtherFeedbackMechanisms1

5

Appendices

AppendixA:Resources1

6

AppendixB:Glossary17

AppendixC:PilotProposalSubmissionTemplate19

AppendixD:SampleMetricsDashboard20

2

3

DHSPlaybookforPublicSectorGenAIDeployment(January2025)

LetterfromtheSecretaryandtheChiefAIOfficer

TherapidadvancementandcommercializationofGenerativeArtificialIntelligence(GenAI)presentsbothsignificant

opportunitiesandchallengesfortheDepartmentofHomelandSecurity(DHS)andpublicsectororganizationsatalllevels.InalignmentwithExecutiveOrder14110,whichemphasizestheresponsibledeploymentofAItechnologies,DHShas

proactivelyinitiatedGenAIpilotprogramstoexploretheirpotentialinenhancingourmissioncapabilities.

ThesepilotinitiativeshaveprovidedvaluableinsightsintothepracticalapplicationsofGenAIwithinouroperations.TheyhaveunderscoredtheimportanceofameasuredandthoughtfulapproachtoAIintegration,ensuringthatourdeploymentsareresponsible,trustworthy,andeffectivewhileprotectingprivacy,civilrights,andcivilliberties.

TheDHSGenAIPlaybookencapsulatesthelessonslearnedfromourpilotprogramsandoffersastructuredframeworkfortheresponsibleadoptionofGenAItechnologies.ItisintendedforanddesignedtoguidesectororganizationsinimplementingGenAIsolutionsalignedwiththeiruniquemissionsandoperationalcontexts.

Theplaybookalsodiscusseshoworganizationsshouldconsiderbuildingholistic,moderntechnologydeliverycapabilitiesincludingGenAI.SuccessfulGenAIdeploymentsdependonastrongfoundationofAI-enablingskillsandtechnologies,

includinghuman-centereddesign,agilesoftwaredelivery,anddatamanagement.WhileGenAIispowerfulandexciting,itisnottheonlytypeofAI.DHScontinuestoleveragePredictiveAIacrossouroperations,andorganizationsmustconsiderwhichtypeofAIisappropriateforeachtask.

WerecognizethatthesuccessfulintegrationofGenAIrequirescollaborationacrossgovernment,industry,academia,andcivilsociety.WeinviteyoutobetterunderstandtheinsightspresentedinthisPlaybookandtojoinusinfosteringacultureof

innovationthatisbothresponsibleandmission-focused.

Together,wecanharnessthetransformativepotentialofGenAItoenhanceourcapabilitiesandbetterservetheAmericanpeople.

AlejandroN.Mayorkas

U.S.DepartmentofHomelandSecuritySecretary

GM-

EricHysen

U.S.DepartmentofHomelandSecurityChiefArtificialIntelligenceOfficer

4

DHSPlaybookforPublicSectorGenAIDeployment(January2025)

Introduction:TheDHSGenAIPilots

DHS’sgenerativeartificialintelligence(GenAI)pilotsadvancetheDepartment’sgoalofusingAIinserviceofthehomelandsecuritymission.DHShasimplementedmoretraditionalformsofAIandmachinelearning(ML)forover15yearsandwasearlyamonggovernmentagenciestoinvestinGenAI.DHSmanagedthiseffortduetotwoprimaryfactors:first,afocusonandenthusiasmforleveragingAIintheDepartment’smissionfromseniorleaders,includingtheSecretary,whichhasbeenkeyforunitingtheDepartmentandleveragingitsresources;andsecond,stronginterestandinitiativefromtheDepartment’scomponents

1

tofindwaystoresponsiblyleverageAItosupporttheiroperationalwork.

OnMarch18,2024,DHSreleasedafirst-of-its-kind

ArtificialIntelligenceRoadmap

todescribetheDepartment’splansto

guideitsresponsibleuseofAIwhileensuringindividuals’privacyrights,civilrights,andcivillibertiesareprotected.The

roadmaplaidoutDHS’sAIinitiatives,describedthepotentialofAItechnologiesacrosstheDepartment,andprovidedvisibilityintoDHS’sapproachtoAI,whileunderscoringtheDepartment’scommitmenttoAI’sresponsibleandtrustworthyuse.

Aspartofthisroadmap,DHSpubliclyannounceditsgoaltocompletethreeGenAIpilotsoverthecourseof2024,describedbelow.

2

Inselectingthesepilots,DHSprioritizedGenAIapplicationsthatwouldhelpemployeesdotheirwork,notreplacethem.DHSalsoscopedpilotssothattheycouldbebroadlyapplicableinotherpartsoftheDepartmentandthefederal

governmentifsuccessful.Thesepilotstargetedmission-enhancingprocessesthatworkerscouldincorporateeasilyintotheirworkstreams.

Pilot1:StrengtheningInvestigativeLeadswithLLM-EnhancedSearchand

Summarization

HomelandSecurityInvestigations(HSI)testedAI’sabilitytostrengthentheirinvestigativeprocesses.ThepilothadHSIinvestigatorsusealargelanguagemodel(LLM)-basedsystemtoenhancetheefficiencyandaccuracyofthe

summariestheyrelyupontoidentifyinvestigatoryleads.Thesystemwasdesignedtoleverageopen-sourcetechnologiestoallowinvestigatorsto

morequicklysummarizeandsearchforcontextuallyrelevantinformationwithininvestigativereports.Whendeployed,thepilotcould,amongotheruses,leadtoincreaseddetectionoffentanyl-relatednetworks,aidintheidentificationofperpetratorsandvictimsofchildexploitationcrimes,andsurfacekeypatternsandtrendstoadvanceHSI’svitalwork.

AnHSIinvesigatorassessesevidence.

1OfficesandagenciescontainedwithDHSarereferredtoas“components”withinthisdocument.Theseinclude,forexample,U.S.

CustomsandBorderProtection,theFederalEmergencyManagementAgency,UnitedStatesCitizenshipandImmigrationServices,andtheCybersecurityandInfrastructureSecurityAgency.

2Asoftimeofpublication,thepilotshavebeencompletedandtherespectiveteamsareconductinganalysistodeterminewhethernext-stagedeploymentisappropriate.Thisplaybookwillbecomplementedbyadditionaltechnicaldocumentationonthepilots.

5

DHSPlaybookforPublicSectorGenAIDeployment(January2025)

Pilot2:HelpingLocalGovernmentsCreateHazardMitigationPlans

TheFederalEmergencyManagementAgency(FEMA)testedLLM

capabilitiestohelplocalgovernmentsdevelophazardmitigationplans

forcommunitiesacrossthecountry.Hazardmitigationplansareessentialtobuildingresilientcommunities,buttheyarealsolengthy,complex

documentsthatcanbedifficultformanycommunitiestoproduce.Thepilotwasdesignedtoenablestateandlocalgovernmentstoefficiently

identifyandunderstandtheircommunities’risksandcorresponding

mitigationstrategies.Thepilotcreateddraftplanningelementsfrom

publiclyavailable,well-researchedsourcesthatcouldbecustomized

tomeeteachcommunity’suniqueneeds.Thegoalistoenablemore

communitiestosubmitgrantapplicationsforhazardmitigationfunding,becomemoreresilient,andreducedisasterrisks.

AFEMAemployeeassistsadisastersurvivor.

Pilot3:CreatingNovelTrainingOpportunitiesforImmigrationOfficers

UnitedStatesCitizenshipandImmigrationServices(USCIS)testedAI’sabilitytoimproverefugeeandasylumimmigrationofficertraining.

USCISusedGenAItoprovidedynamic,personalizedinterviewtrainingthatadaptedtoofficers’specificneedsandensuredthebestpossible

knowledgeandtrainingonawiderangeofscenarios,currentpolicies,andlaws.Theenhancedtrainingmaterialsgaveeachofficerthe

opportunitytopracticeandprepareforinterviewswithrefugeesand

asylumapplicantsontheirown,asoftenasneeded,enablingcontinuous

learningandreducingtheirrelianceonlimitedinstructor-ledtraining

sessions.Thisdynamic,adaptabletrainingresourceaimedtoenhance

trainees’interviewskills,increasecomprehensionandretentionofmission-criticalinformation,improvetheaccuracyoftheirdecision-makingprocess,andlimittheneedforretrainingovertime.

AUSCISemployeeassistsapplicants

atacontactcenter.

Acrosseachofthesepilots,andotherGenAIintegrations,DHScarefullyconsideredtherisksandbenefitsofeach

ofthesetechnologies.UsingGenAIandothernewtechnologycomeswithrisks:risksposedbyadversarialuseofGenAI,risksposedtoGenAIsystems,andrisksassociatedwiththeimplementationofGenAI.

3

Thesemayincludeoperational,security,data,privacy,andinappropriatebiasrisks.Whilesomeofthesedangersarecontext-specific

–forexample,therisksassociatedwithdeployingGenAIforofficertrainingprogramslookdifferentthanfor

documentsearches–forwardplanningcanhelpincorporateriskmitigationsbydesign.TheDepartmentproactivelyincorporatedriskmitigationsthroughrobust,cross-functionalgovernanceprocessesthatincludedongoing

monitoringandtesting.

3SeeDHS’sSafetyandSecurityGuidelinesforCriticalInfrastructureOwnersandOperators,April2024,https://www.dhs.

gov/sites/default/files/2024-04/24__0426_dhs_ai-ci-safety-security-guidelines-508c.pdf

6

DHSPlaybookforPublicSectorGenAIDeployment(January2025)

StartNow,WhereverYouAre

DHSdesignedthisplaybooktomeetorganizationswherevertheyareintheirjourneytounderstandandincorporateAItechnologyintheirwork.Anypublicsectororganizationcanstarttodaytoassessandgatherresources,cultivateinternalbuy-in,andlaythegroundworkforeffectivedeploymentsofGenAI.Layingthisfoundationwillpaydividendsinfuturedevelopmentsandcreateorganizationalmomentum.

EachofthefollowingstepsdescribeactionsthatorganizationscantaketoadvancetheirownusesofGenAI.Thesestepsaddresstechnical,policy,andadministrativeconsiderations;therefore,somerecommendationsmaybemorerelevanttospecificexpertswithinanorganization,suchastechnicalorpolicypersonnel.

Mission-EnhancingGenAIUseCases:PublicsectororganizationsmustensurethatGenAIdeploymentsalignwiththeirmission.Narrowlyscoped,mission-enhancingpilotsareusefultoolsforexploringhowanorganizationcanuseGenAI.

CoalitionBuildingandEffectiveGovernance:OrganizationsshouldcultivatesupportforGenAIapplicationsfromtopleadershipandacrossfunctionalteamstogiveGenAIthegreatestchanceforsuccessfuldeploymentandeffectiveoversight.

ToolsandInfrastructure:OrganizationsshouldevaluatethetechnicaltoolsandinfrastructuretheyalreadypossessandconsiderwhattechnicalcapabilitiestheyrequiretodeployGenAIapplications.

ResponsibleUseandTrustworthinessConsiderations:Fromtheverybeginning,organizationsshouldconsider

howtomakesureGenAIuseisresponsibleandtrustworthyandhowtoaddresspotentialriskslikeprivacy,security,bias,andsafety.

MeasurementandMonitoring:TeamsthataredevelopingGenAIapplicationsshouldmeasureprogresswithappropriatemetricsandreportonthatprogresstoleadershipandotherstakeholders.

TrainingandTalentAcquisition:OrganizationsshouldtraintheirstaffonresponsibleandeffectiveGenAIuseandhireskilledemployeeswhocansupportGenAIdevelopment.

UsabilityTestingandOtherFeedbackMechanisms:OrganizationsshouldincorporateiterativefeedbackfromusersandotherstakeholderstodevelopandimproveGenAIapplications.

7

DHSPlaybookforPublicSectorGenAIDeployment(January2025)

ACTIONABLESTEPS

Mission-EnhancingGenAIUseCases

DevelopGenAIPilotsthataddressmissionneedandtargetsupportingmission-enhancingprocesses.

?

?

?

?

?

AlignanyGenAIdeployment’s

potentialmissionandvaluewiththeorganization’spriorities.

ScopeaGenAIpilotthatimprovesaspecificmission-enhancingprocessandispotentiallyscalabletosimilarprocessesorusefulinotherpartsoftheorganization.

EnlistanexecutivesponsorwhoseorganizationwillbenefitfromtheGenAIpilot.

Assessresources,includingstaff,

funding,data,andtechnology.

Definethepilot’sminimumviableproduct(i.e.,whenitwillbereadyforevaluation)andmetricsfor

evaluatingsuccess.

Organizationsshouldcarefullydeterminethescopeofanypotential

GenAIdeploymentandconsiderdesigningaGenAIpilotbeforedeployingafunctionalitybroadly.Well-designedGenAIpilotsprovidealow-stakes

environmenttotakerisksandlearnaboutresourceandgovernanceneeds.Teamsshoulddesignpilotsthatsolveaspecificprobleminawaythat

servestheorganization’smission.Importantly,pilotsshouldfirstfocusonsupportingmission-enhancingprocessesbeforeaddingGenAI

tomission-criticaloperations,wherefailurecouldhavemoreserious

consequences.Organizationsshouldalsoconsiderthelong-termpotentialofapilotproduct,includingwhetherotherdepartmentsorprocesses

couldmakeuseofthetool.Finally,organizationsshouldenlistanexecutivesponsorearlytosupportandendorsepiloteffort(s).

Pilotplansshouldassessresourceneeds—includingfunding,staff,

data,andtechnology—andcriteriaforevaluatingthepilot,whetherobjective(e.g.,performance,accuracy,andadoption)orsubjective(e.g.,userexperienceandsentiment).Datamaybeaparticularchallenge,andteamsshouldidentifydataneeds,datasets,potentialprivacyissues,andresponsibledatausebeforelaunchingapilot.

THEDHSSTORY

InApril2023,theSecretaryofHomelandSecuritystoodupanAITaskForce(AITF)andchargeditwithidentifyingwaysAIcouldenhancetheDHSmissionwhileprotectingprivacy,civilrights,andcivillibertiesacrosstheDepartment.TheAITFsolicitedandreceivednearlythreedozenproposalsforpilotprojectsfromacrosstheDepartmenttobeconsideredfor

short-termfundingfromDHSHeadquarters.TheAITFchosepilotstofundbasedontheprojects’alignmentwiththeDHSandcomponentmission,technicalandfundingfeasibility,potentialcross-organizationalapplications,accesstodataandinfrastructure,andreasonableprojectduration.TheAITFalsostrategicallychosepilotsthatrepresenteddifferentusecasesandtypesoftechnologyapplications,sothatDHScouldcollectivelylearnasmuchaspossibleaboutGenAIapplications.AsamplepilotproposaltemplatecanbefoundinAppendixC.

ScopewasanimportantconsiderationfortheAITFinselectingpilots.Tofacilitatemeasurementandtracking,theAITF

fundedpilotprojectsthatfocusedondiscrete,tangibleoutcomes,suchasreducinghoursofficersspentworkingonspecificcases.Also,becauseDHSpolicyprohibitsusingGenAIasthesolebasisfordecision-making,pilotsinsteadtargetedprocessimprovements,suchasenhancingtrainingorassistingresearch.Eachpilotteamidentifieditscriteriaforaminimumviableproductandbenchmarksthatwouldindicatewhenpilotresultswouldbereadyforevaluation.

Tofinancethepilots,DHSdrewfromafundforITmodernizationthatwasreadilyavailableforone-timeusebutcouldnotbeusedonarecurringbasis.ThesefundsgaveDHStheflexibilitytodevelopandexperimentwithGenAIwithoutimpactingtheDepartment’sbudget.

ExecutivesateachcomponentrunningaGenAIpilotprovidedintegralsupportfortheprojects.WhileDHSdidnotrequirepilotteamstoidentifycomponentexecutivesassponsorsattheonset,DHSrecommendsdoingsointhefuture.Component(orsub-agencyhead)executivescanhelpmarshalresourcesmoredirectlyandrespondtoissuesastheycomeup.

8

DHSPlaybookforPublicSectorGenAIDeployment(January2025)

ACTIONABLESTEPS

CoalitionBuildingandEffectiveGovernance

Seeksponsorshipfromthemostseniorexecutive,thenbuildcross-organizationalcoalitionstooverseeGenAIdeployments.

?Prioritizeearlybuy-infromthemostseniorleadershippossibleandgaininputfromkeystakeholdersacrossallpartsoftheorganization.

?Evaluatecurrentgovernance

structuresandpoliciestoidentifypotentialgaps.

?Designateanexistinggovernance

bodythatincludescross-functionalrepresentationorstandoneupfor

managementandoversight.Includetechnicalexperts,cybersecurity

professionals,privacy,civilrights

andcivillibertiesprofessionals,

anddedicatedresponsibleuseand

trustworthinessexpertsinteams

developinganddevelopingtheGenAItool.

Organizationsshouldsolicitbuy-inforGenAIprojectsfromkeyinternalstakeholders,fromthemostseniorleadershippossible(uptoand

includingtheheadoftheorganization).Itisespeciallyimportantto

includeriskmanagement,compliance,andoversightleads(suchascivilrightsandcivillibertiesexperts),gaintheirsupport,andincorporate

theirinsightsearlyintheplanningprocess.Identifyingtheright

stakeholdersmaytakesomeworkinlargerbureaucracies,buthavingacross-organizationalcoalitionworkingonGenAIcanleadtobetterdecision-making,fasterdevelopment,andstrongerresponsibleuse

practices.

Organizationsshouldassesstheircurrentgovernancestructures

andconsiderhowtheycouldbeleveraged,updated,oraddedupon

toaddressAIgovernance.Sometimesanorganizationcansimplyadd

responsibilitiestoanexistingbody;atothertimes,itwillneedtocreateanewbodytooverseeAIandanyGenAIpilots.Thesestructuresmay

changeovertimeaspolicyandthetechnologyevolves,butAIgovernancestructuresshouldsimilarlyrepresentawiderangeofstakeholdersandteamsacrosstheorganization,suchasprivacyprofessionals,civilrightsandcivillibertiesexperts,

legalcounsel,andcybersecurityanddataexperts.Thisapproachservestoreinforcecoalitionsandensurethesebodiesarewell-positionedtoidentifythefullrangeofpotentialrisks,supportimpactassessments,designmitigationstrategies,andimplementremediationsinrealtime.

THEDHSSTORY

DHSbenefittedfromtheSecretaryofHomelandSecurity’sdirect

sponsorshipofGenAIapplicationadoption,whichunlockedresourcesandbuy-inforGenAIacrosstheDepartment.Inaddition,manyDHS

componentsalreadyhadlong-standingexperienceincorporatingAIintotheirmissions.Thiscombinationoftop-levelleadershipandcomponent-levelexperienceandsupporthelpedDHSbuildcoalitionsacrossthe

DepartmenttosupportAIgovernance,development,anddeployment.

DHSreinforcedthisapproachonthegroundwithintegratedprojectteams

(IPTs)thatincludedcybersecurityexpertsandlegal,privacy,andcivilrights

andcivillibertiesprofessionals.TheIPTsoversawthepilots’development

anddeployment,helpingtheGenAIprojectsestablishlegitimacyandearnbuy-inacrosstheDepartment.

SecretaryMayorkasparticipatesinafiresidechataboutAI.

Cross-functionalcollaborationwaskeyforDHS’sAIgovernanceefforts,bothfortheGenAIpilotsandotherAIdeployments.WhenDHSbeganitsGenAIpilots,theAITaskForce—anew,limited-termentity—wasinitiallythe

Department’sonlygoverningbodyfocusedonAI.TheAITFwasledbyseniorrepresentativesfromtheOfficeoftheChief

9

DHSPlaybookforPublicSectorGenAIDeployment(January2025)

InformationOfficer/ChiefAIOfficer,DHSScienceandTechnologyDirectorate,andtheDHSOfficeforCivilRightsandCivilLiberties.MembershipincludedrepresentativesfromkeyorganizationsthroughouttheDepartment,includingtheDHSOfficeofStrategy,Policy,andPlansandtheDHSPrivacyOffice.ThemultidisciplinaryIPTsalsoprovidedcritical

on-the-groundgovernanceforAIrisks,asteammemberscouldaddressissueslikecybersecurityconcernsandprivacy

considerationsastheyarose.ThepilotteamscouldthenelevateproblemsandlessonslearnedfromDHScomponent-levelAIworkinggroupstotheAITaskForce,informingDHS’sbroaderAIgovernanceefforts.

InparallelwithitsGenAIpilots,DHStookotherstepstodevelopandmatureitspoliciesandpracticestoguideresponsibleAIuse.Critically,theChiefAIOfficerissuedaDepartment-widepolicythatprovidesinitialguidanceforresponsibleGenAIusewithinDHS.ThepolicydetailedwhatdataemployeescansharewithaGenAItool,requiredemployeeswhowantedtouseGenAItoundergoatrainingandsecuretheapprovalofasupervisor,requiredhumanreviewofGenAIoutputsbeforethoseoutputscouldbereliedupon,andsetforthappropriateusecases.TheChiefPrivacyOfficerpublishedaPrivacy

ImpactAssessmentontheDepartment’suseofcommercialGenAItools.

4

TheAITFalsosetupcommunitiesofpractice,

suchasaResponsibleUseGroup,whichcollaboratedonresponsibleAIbestpracticesandpriorities.DHSwillcontinuetoassessandupdatepoliciestoguideresponsibleGenAIuseatDHSastheDepartmentandthetechnologyevolves.

4SeeDHS/ALL/PIA-097UseofConditionallyApprovedCommercialGenerativeArtificialIntelligenceTools,November2023,

/publication/dhsallpia-097-use-conditionally-approved-commercial-generative-artificial-intelligence

10

DHSPlaybookforPublicSectorGenAIDeployment(January2025)

ACTIONABLESTEPS

ToolsandInfrastructure

BuildorganizationalcapacitytodevelopanddeployGenAIbyusingexistingtoolsandinfrastructureforpilots.

?Assessexistingtools,infrastructure,andtechnicalcapabilitiesagainsttheneedsandgoalsoftheGenAIpilotordeployment,includingsecuringsensitivedata.

?Considertheuseofcommercial,open-source,oropen-weight

models,andwhetherusingmorethanonetypeofmodelwouldbeappropriate.

?Determineifanyadditionaltoolingorconfigurationsareneeded,

keepinginmindyourteam’stechnicalcapacity.

OrganizationsshouldassessexistingtoolsandinfrastructureavailabletothemforGenAIandplanaroundtheirstrengthsandlimitations.

Organizationscanadapttheirexistingprocessesforsecuritycompliance,scalability,andintegrationwithexistingITinfrastructureanddata

managementpractices.

Organizationsshouldconsultwiththeirtechnicalexpertsand

considerwhethercommercial

5

,open-source

6

,oropen-weight

7

AI

modelsaremoreappropriatefortheirneeds,especiallyiftheyface

resourceconstraintsorhandlesensitivedata.Open-sourcemodels

offerorganizationsmorecontrolandvisibilityintothemodel’s

functioningandsecurity,oftenwithalowertotalcostofownership,

whilecommercialandopen-weightmodelsmayhavefewerlong-termmaintenancecosts.Anorganization’scloudenvironmentcouldaffectitsmodelchoice,assomelargemodelsarelessexpensivetousethrough

existingcloudcontractAPIs,whileopen-sourcemodelsrequireaserviceproviderororganizationalinfrastructure.The

purposeofanorganization’sGenAIpilotcanalsoinfluencetoolingandinfrastructuredecisions.Thisisespeciallytrueforlawenforcementapplications,whichmayrequireofflinemodelsthatcanrunlocally.Open-weightAImodelscansupportsuchofflineuses.

Exploringnoveland/oradditionalAItoolsshouldcomeafterbuildingstrongsoftwaredesign,development,and

cybersecurityfundamentals.Withoutsufficientbaselineexpertiseortooling,aGenAIpilot’sresourceneedscouldquicklyoverwhelmanagency’sbudgetandinfrastructure.

THEDHSSTORY

DHS’sGenAIpilotteamsbegantheirworkbyassessingtheiravailabletools,infrastructure,andtechnicalskills,andthenplannedaccordingly.Theteamsfoundthatusingavailabletoolsandlicensessimplifiedpilotdevelopment.

TheGenAIpilotteamshadaccesstocommercialGenAImodelsthroughapprovedcloudservicesthatwerealreadyavailabletofederalemployees.Theyalsoexploredvariousopen-sourceandopen-weightmodels.Pilotteamsfoundthatusing

modelslinkedtopre-approvedcloudservicesstreamlinedcompliancewithsecurityneeds,scalability,andintegrationprocessesthatotherwisemayhaveproventobesignificantobstacles.

Thepilotteamsalsoidentifiedandusedtoolsthatmatchedtheteams’skillsets.Forteamswithlimitedresources,low-codeplatformssuchascommercialAIstudiosproveduseful.

5‘CommercialGenAItools’aredefinedasgenerativeAIproductsorservicesavailableforlicense,subscription,orpurchasebythegeneralpublicorthroughcommercialagreements.ThisdefinitiondoesnotincludesoftwareorservicesdevelopedorcustomizedspecificallyforthegovernmentthroughITacquisitionprocess.

6Anopen-sourceAImodelisanartificialintelligencemodelwhosesourcecode,modelweights,andtrainingdataarepublicallyavailableforanyonetouse,modify,study,andshare.

7Anopen-weightAImodelisanartificialintelligencemodelwhosemodelweightsandinformationaboutthetrainingdataare

publiclyavailableandthathasapermissiveuseofthecapability.Incontrasttoopen-sourcemodels,open-weightmodelsma