世界經(jīng)濟論壇-人工智能治理聯(lián)盟：簡報系列（英）-2024.1

AI

Governance

AllianceBrie?ng

Paper

SeriesJA

N

U

A

RY

20

2

4January2024AIGovernanceAllianceBrie?ngPaperSeriesForewordPaulDaughertyJeremyJurgensManagingDirector,WorldEconomicForumChiefTechnologyandInnovationOf?cer(CTIO),AccentureCathyLiJohnGrangerSeniorVice-President,IBMConsultingHead,AI,DataandMetaverse;MemberoftheExecutiveCommittee,WorldEconomicForumOurworldisexperiencingaphaseofmulti-facetedtransformationinwhichtechnologicalinnovationplaysaleadingrole.Sinceitsinceptioninthelatterhalfofthe20thcentury,arti?cialintelligence(AI)hasjourneyedthroughsigni?cantmilestones,culminatingintherecentbreakthroughofgenerativeAI.GenerativeAIpossessesaremarkablerangeofabilitiestocreate,analyseandinnovate,signallingaparadigmshiftthatisreshapingindustriesfromhealthcaretoentertainment,andbeyond.ResponsibleApplicationsandTransformation,andResilientGovernance

andRegulation.Thesepillarsunderscore

acomprehensiveend-to-endapproachtoaddresskeyAIgovernancechallengesandopportunities.Theallianceisaglobaleffortthatunitesdiverseperspectivesandstakeholders,whichallowsforthoughtfuldebates,ideationandimplementationstrategiesformeaningfullong-termsolutions.Thealliancealsoadvanceskeyperspectivesonaccessandinclusion,drivingeffortstoenhanceaccesstocriticalresourcessuchaslearning,skills,data,modelsandcompute.Thisworkincludesconsideringhowsuchresourcescanbeequitablydistributed,especiallytounderservedregionsandcommunities.Mostcritically,itisvitalthatstakeholderswhoaretypicallynotengagedinAIgovernancedialoguesaregivenaseatatthetable,ensuringthatallvoicesareincluded.Indoingso,theAIGovernanceAllianceprovidesaforumforall.AsnewcapabilitiesofAIadvanceanddrivefurtherinnovation,itisalsorevolutionizingeconomiesandsocietiesaroundtheworldatanexponentialpace.WiththeeconomicpromiseandopportunitythatAIbrings,comesgreatsocialresponsibility.Leadersacrosscountriesandsectorsmustcollaboratetoensureitisethicallyandresponsiblydeveloped,deployedandadopted.The

World

Economic

Forum’s

AI

Governance

Alliance(AIGA)standsasapioneeringcollaborativeeffort,unitingindustryleaders,governments,academicinstitutions

and

civil

society

organizations.

The

alliancerepresentsasharedcommitmenttoresponsibleAIdevelopmentandinnovationwhileupholdingethicalconsiderationsateverystageoftheAIvaluechain,fromdevelopmenttoapplicationandgovernance.Thealliance,ledbytheWorldEconomicForumincollaborationwithIBMConsultingandAccentureasknowledgepartners,ismadeupofthreecoreworkstreams–SafeSystemsandTechnologies,Aswenavigatethedynamicandever-evolvinglandscapeofAIgovernance,theinsightsfromtheAIGovernanceAllianceareaimedatprovidingvaluableguidancefortheresponsibledevelopment,adoptionandoverallgovernance

ofgenerativeAI.Weencourage

decision-makers,

industry

leaders,

policy-makersandthinkersfrom

around

theworldtoactivelyparticipateinourcollectiveeffortstoshapeanAI-drivenfuturethatupholdssharedhumanvaluesandpromotesinclusivesocietalprogressforeveryone.AIGovernanceAlliance2Introduction

tothebrie?ngpaperseriesTheAIGovernanceAlliancewaslaunchedinJune2023withtheobjectiveofprovidingguidanceontheresponsibledesign,developmentanddeploymentofarti?cialintelligencesystems.Sinceitsinception,morethan250membershavejoinedthealliancefromover200organizationsacrosssixcontinents.Theallianceiscomprisedofasteeringcommitteealongwiththreeworkinggroups.businesstransformationforresponsiblegenerativeAIadoptionacrossindustriesandsectors.ThisincludesassessinggenerativeAIusecasesenablingneworincrementalvaluecreation,andunderstandingtheirimpactonvaluechainsandbusinessmodelswhileevaluatingconsiderationsforadoptionandtheirdownstreameffects.TheResilientGovernanceandRegulationworkinggroup,ledincollaborationwithAccenture,isfocusedontheanalysisoftheAIgovernancelandscape,mechanismstofacilitateinternationalcooperationtopromoteregulatoryinteroperability,aswellasthepromotionofequity,inclusionandglobalaccesstoAI.TheSteeringCommitteecomprisesleadersfromthepublicandprivatesectorsalongwithacademiaandprovidesguidanceontheoveralldirectionoftheallianceanditsworkinggroups.TheSafeSystemsandTechnologiesworkinggroup,ledincollaborationwithIBMConsulting,isfocusedonestablishingconsensusonthenecessarysafeguardstobeimplementedduringthedevelopmentphase,examiningtechnicaldimensionsoffoundationmodels,includingguardrailsandresponsiblereleaseofmodelsandapplications.Accountabilityisde?nedateachstageoftheAIlifecycletoensureoversightandthoughtfulexpansion.Thisbrie?ngpaperseriesisthe?rstoutputfromeachofthethreeworkinggroupsandestablishesthefoundationalfocusareasoftheAIGovernanceAlliance.Inatimeofrapidchange,theAIGovernanceAllianceseekstobuildamultistakeholdercommunityoftrustedvoicesfromacrossthepublic,private,civilsocietyandacademicspheres,united,totacklesomeofthemostchallengingandpotentiallymostrewardingissuesincontemporaryAIgovernance.TheResponsibleApplicationsandTransformationworkinggroup,ledincollaborationwithIBMConsulting,isfocusedonevaluatingAIGovernanceAlliance3ReadingguideThis

paper

series

is

composed

of

three

brie?ng

papersthathavebeengroupedintothematiccategoriesaccordingto

the

threeworkinggroupsof

the

alliance.policies,principlesandpracticesthatgoverntheethicaldevelopment,deployment,useandregulationofAItechnologies,theResilientGovernanceandRegulationbrie?ngpaperoffersguidance.Eachbrie?ngpaperofthereportcanalsobereadasastand-alonepiece.Forexample,developers,adoptersandpolicy-makerswhoare

moreinterestedinthetechnicaldimensionscaneasilyjumptotheSafeSystemsandTechnologiesbrie?ngpapertoobtainacontemporaryunderstandingoftheAIlandscape.

For

decision-makers

engaged

in

corporatestrategyandbusinessimplicationsofgenerativeAI,theResponsibleApplicationsandTransformationbrie?ngpaperoffersspeci?ccontext.Forbusinessleadersandpolicy-makersoccupiedwiththelaws,Whileeachbrie?ngpaperhasauniquefocusarea,manyimportantlessonsarelearnedattheintersectionofthesevaryingmultistakeholdercommunities,alongwiththeconsensusandknowledgethatemanatefromeachworkinggroup.Therefore,manyofthetakeawaysfromthisbrie?ngpaperseriesshouldbeviewedattheintersectionofeachworkinggroup,where?ndingsbecomeadditiveandareenhancedincontextandinterrelationwithoneanother.AI

Governance

AllianceBrie?ng

PaperSeriesJA

N

U

A

R

Y

2

0

24Theme1SafeSystemsandTechnologiesTheme2ResponsibleApplicationsTheme3ResilientGovernanceandRegulationandTransformation1/3

AIGovernanceAlliance2/3

AIGovernanceAlliance3/3

AIGovernanceAllianceBrie?ngPaperSeries2024Brie?ngPaperSeries2024Brie?ngPaperSeries2024PresidioAIFramework:Towards

Safe

GenerativeAIModelsUnlocking

ValueGenerative

AIGovernance:from

Generative

AI:Guidance

for

ResponsibleTransformationShapingaCollectiveGlobalFutureIN

C

O

L

LA

B

O

RAT

IO

NW

IT

H

A

C

C

E

N

TU

R

EIN

C

O

L

LA

B

O

RAT

IO

NW

IT

H

IB

M

C

O

N

S

U

LT

IN

GIN

C

O

L

LA

B

O

RAT

IO

NW

IT

H

IB

M

C

O

N

S

U

LT

IN

GAIGovernanceAlliance4AIGovernance

AllianceSteeringCommitteeNickCleggAndrewNgPresident,GlobalAffairs,MetaFounder,

DeepLearning.AIGaryCohnSabastianNilesVice-Chairman,IBMPresidentandChiefLegalOf?cer,

SalesforceSadieCreeseOmarSultanAlOlamaProfessorofCybersecurity,UniversityofOxfordMinisterofStateforArti?cialIntelligence,UnitedArabEmiratesOritGadieshChairman,Bain&CompanyLynne

ParkerAssociateVice-ChancellorandDirector,AITennessee

Initiative,UniversityofTennesseePaulaIngabireMinisterofInformationCommunicationTechnology

ofRwandaBradSmithVice-ChairandPresident,MicrosoftDaphneKollerFounderandChiefExecutiveOf?cer,

InsitroMustafaSuleymanCo-FounderandChiefExecutiveOf?cer,In?ectionAIXueLanProfessor;Dean,SchwarzmanCollege,Tsinghua

UniversityJosephineTeoMinisterforCommunicationsandInformationMinistryofCommunicationsandInformation(MCI)ofSingaporeAnnaMakanjuVice-President,GlobalAffairs,OpenAIDurgaMalladiKentWalkerSeniorVice-President,QualcommPresident,GlobalAffairs,GoogleAIGovernanceAlliance5GlossaryTerminology

inAIisafast-movingtopic,andthesametermcanhavemultiplemeanings.Theglossarybelowshouldbeviewedasasnapshotofcontemporaryde?nitions.Mis/disinformation:Misinformationinvolvesthedisseminationofincorrectfacts,whereindividualsmayunknowinglyshareorbelievefalseinformationwithouttheintenttomislead.DisinformationinvolvesthedeliberateandintentionalspreadofArti?cialintelligencesystem:amachine-basedsystemthat,forexplicitorimplicitobjectives,infers,fromtheinputitreceives,howtogenerateoutputssuchaspredictions,content,recommendationsordecisionsthatcanin?uencephysicalorvirtualenvironments.DifferentAIsystemsvaryintheirlevelsofautonomyandfalseinformationwiththeaimofmisleadingothers.4Modeldriftmonitoring:Theactofregularlycomparingmodelmetricstomaintainperformancedespitechangingdata,adversarialinputs,noiseandexternalfactors.adaptivenessafterdeployment.1Modelhyperparameters:Adjustableparametersofamodelthatmustbetunedtoobtainoptimalperformance(asopposedto?xedparametersofamodel,de?nedbasedonitstrainingset).CausalAI:AImodelsthatidentifyandanalysecausalrelationshipsindata,enablingpredictionsanddecisionsbasedontheserelationships.CausalinferencemodelsprovideresponsibleAIbene?ts,includingexplainabilityandbiasreductionthroughformalizationsoffairness,aswellascontextualisationformodelreasoningandoutputs.TheintersectionandexplorationofcausalandgenerativeAImodelsisanewconversation.Multi-modalAI:AItechnologycapableofprocessingandinterpretingmultipletypesofdata(liketext,images,audio,video),potentiallysimultaneously.Itintegratestechniquesfromvariousdomains(naturallanguageprocessing,computervision,audioprocessing)formorecomprehensiveanalysisandinsights.Fine-tuning:Theprocessofadaptingapre-trainedmodeltoperformaspeci?ctaskbyconductingadditionaltrainingwhileupdatingthemodel’sexistingparameters.Promptengineering:Theprocessofdesigningnaturallanguagepromptsforalanguagemodeltoperformaspeci?ctask.Foundationmodel:AfoundationmodelisanAImodelthatcanbeadaptedtoawiderangeofdownstreamtasks.Foundationmodelsaretypicallylarge-scale(e.g.billionsofparameters)generativemodelstrainedonavastarrayofdata,encompassingbothlabelledandunlabelleddatasets.Retrievalaugmentedgeneration:Atechniqueinwhichalargelanguagemodelisaugmentedwithknowledgefromexternalsourcestogeneratetext.Intheretrievalstep,relevantdocumentsfromanexternalsourceareidenti?edfromtheuser’s

query.Inthegenerationstep,portionsofthosedocumentsareincludedinthemodelprompttogeneratearesponsegroundedintheretrieveddocuments.Frontiermodel:Thistermgenerallyreferstothemostadvancedorcutting-edgemodelsinAItechnology.Frontiermodelsrepresentthelatestdevelopmentsandareoftencharacterizedbyincreasedcomplexity,enhancedcapabilitiesandimprovedperformanceoverpreviousmodels.Parameter-ef?cient?ne-tuning:Anef?cient,low-costwayofadaptingapre-trainedmodeltonewtaskswithoutretrainingthemodelorupdatingitsweights.Itinvolveslearningasmallnumberofnewparametersthatareappendedtoamodel’s

promptwhilefreezingthemodel’s

existingparameters(alsoknownasprompt-tuning).GenerativeAI:AImodelsspeci?callyintendedtoproducenewdigitalmaterialasanoutput(e.g.text,images,audio,videoandsoftwarecode),includingwhensuchAImodelsareusedinapplicationsandtheiruserinterfaces.ThesearetypicallyconstructedasmachinelearningsystemsthathavebeentrainedAIredteaming:

A

methodofsimulatingattacksbyagroupofpeopleauthorizedandorganizedtoidentifypotentialweaknesses,vulnerabilitiesandareasforimprovement.It

should

be

integral

frommodel

designtodevelopmenttodeploymentandapplication.Theredteam’s

objectiveistoimprovesecurityandrobustnessbydemonstratingtheimpactsofsuccessful

attacks

and

by

demonstrating

what

worksforthedefendersinanoperationalenvironment.onmassiveamountsofdata.2Hallucination:Hallucinationsoccurwhenmodelsproducefactuallyinaccurateoruntruthfulinformation.Often,hallucinatoryoutputispresentedinaplausibleorconvincingmanner,

makingdetectionbyendusersdif?cult.Reinforcementlearningfromhumanfeedback(RLHF):Anapproachformodelimprovementwherehumanevaluatorsrankmodel-generatedoutputsforsafety,relevanceandcoherence,andthemodelisupdatedbasedonthisfeedbacktobroadlyimproveperformance.Jurisdictionalinteroperability:Theabilitytooperatewithinandacrossdifferentjurisdictionsgovernedbydifferingpolicyandregulatoryrequirements.3AIGovernanceAlliance6Releaseaccess–Agradientcoveringdifferentlevelsofaccessgranted.evaluationtoensure

thatvaluecanberealized

andchangemanagementissuccessfullyalignedwithde?nedgoalsinaresponsibleframework.5–Fullyclosed:Thefoundationmodelanditscomponents(likeweights,dataanddocumentation)arenotreleasedoutsidethecreatorgrouporsub-sectionoftheorganization.Thesameorganizationusuallydoesmodelcreationanddownstreammodeladaptation.Externalusersmayinteractwiththemodelthroughanapplication.ResponsibleAI:AIthatisdevelopedanddeployedinwaysthatmaximizebene?tsandminimizetherisksitposestopeople,societyandtheenvironment.Itisoftendescribedbyvariousprinciplesandorganizations,includingbutnotlimitedtorobustness,transparency,explainability,fairnessandequity.6––Hosted:Creatorsprovideaccesstothefoundationmodelbyhostingitontheirinfrastructure,allowinginternalandexternalinteractionviaauserinterface,andreleasingspeci?cmodeldetails.Responsibletransformation:Theorganizationaleffortandorientationtoharnesstheopportunitiesandbene?tsofgenerativeAIwhilemitigatingtheriskstoindividuals,organizationsandsociety.Responsibletransformationisstrategiccoordinationandchangeacrossanorganization’sgovernance,operations,talentandcommunications.Applicationprogramminginterface(API):CreatorsprovideaccesstothefoundationmodelbyhostingitontheirinfrastructureandallowingadapterinteractionviaanAPItoperformprescribedtasksandreleasespeci?cmodeldetails.Traceability:Determiningtheoriginalsourceandfactsofthegeneratedoutput.Transparency:Thedisclosureofdetails(decisions,choicesandprocesses)inthedocumentationaboutthesources,dataandmodeltoenableinformeddecisionsregardingmodelselectionandunderstanding.––Downloadable:Creatorsprovideawaytodownloadthefoundationmodelforrunningontheadapters’infrastructurewhilewithholdingsomeofitscomponents,liketrainingdata.Usagerestriction:Theprocessofrestrictingtheusageofthemodelbeyondtheintendedusecases/purposetoavoidunintendedconsequencesofthemodel.Fullyopen:Creatorsreleaseallmodelcomponents,includingallparameters,weights,modelarchitecture,trainingcode,dataanddocumentation.Watermarking:Theactofembeddinginformationinto

outputs

created

by

AI

(e.g.

images,

videos,

audio,text)forthepurposesofverifyingtheauthenticityoftheoutput,identityand/orcharacteristicsofitsResponsibleadoption:TheadoptionofindividualusecasesandopportunitieswithintheresponsibleAIframeworkofanorganization.Itrequires

thoroughprovenance,modi?cationsand/orconveyance.7Endnotes1.2.3.4.5.6.7.“OECDAIPrinciplesoverview”,OrganisationforEconomicCo-operationandDevelopment(OECD)AIPolicyObservatory,2023,https://oecd.ai/en/ai-principles.OECD,G7HiroshimaProcessonGenerativeArti?cialIntelligence(AI)Towardsa

G7CommonUnderstandingonGenerativeAI,2023,/publications/g7-hiroshima-process-on-generative-arti?cial-intelligence-ai-bf3c0c60-en.htm.WorldEconomicForum,InteroperabilityIntheMetaverse,2023,/publications/interoperability-in-the-metaverse/.WorldEconomicForum,ToolkitforDigitalSafetyDesignInterventions

andInnovations:TypologyofOnlineHarms,2023,/publications/toolkit-for-digital-safety-design-interventions-and-innovations-typology-of-online-harms/.Solaiman,Irene,“TheGradientofGenerativeAIRelease:MethodsandConsiderations”,HuggingFace,2023,/abs/2302.04844.WorldEconomicForum,ThePresidioRecommendationsonResponsibleGenerativeAI,2023,/publications/the-presidio-recommendations-on-responsible-generative-ai/.TheWhiteHouse,ExecutiveOrderontheSafe,Secure,

andTrustworthy

DevelopmentandUseofArti?cialIntelligence,2023:/brie?ng-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-arti?cial-intelligence/.AIGovernanceAlliance71/3AIGovernanceAllianceBrie?ngPaperSeries

2024Presidio

AI

Framework:Towards

Safe

GenerativeAI

ModelsI

N

C

O

L

L

A

B

O

R

A

T

I

O

NW

I

T

H

I

B

M

C

O

N

S

U

L

T

I

N

GCoverimage:MidJourneyContentsExecutivesummary101112131515161617181922Introduction1IntroducingthePresidioAIFramework2ExpandedAIlifecycle3GuardrailsacrosstheexpandedAIlifecycle3.1Foundationmodelbuildingphase3.2Foundationmodelreleasephase3.3Modeladaptationphase4ShiftingleftforoptimizedriskmitigationConclusionContributorsEndnotesDisclaimerThisdocumentispublishedbytheWorldEconomicForumasacontributiontoaproject,insightareaorinteraction.The?ndings,interpretationsandconclusionsexpressedhereinarearesultofacollaborativeprocessfacilitatedandendorsedbytheWorldEconomicForumbutwhoseresultsdonotnecessarilyrepresenttheviewsoftheWorldEconomicForum,northeentiretyofitsMembers,Partnersorotherstakeholders.?2024WorldEconomicForum.Allrightsreserved.Nopartofthispublicationmaybereproducedortransmittedinanyformorbyanymeans,includingphotocopyingandrecording,orbyanyinformationstorageandretrievalsystem.1/3:PresidioAIFramework9Executive

summaryThePresidioAIFrameworkaddressesgenerativeAIrisksbypromotingsafety,ethics,andinnovationwithearlyguardrails.TheriseofgenerativeAIpresentssigni?cant1.

ExpandedAIlifecycle:Thiselementoftheframeworkestablishesacomprehensiveend-to-endviewofthegenerativeAIlifecycle,signifyingvaryingactorsandlevelsofresponsibilityateachstage.opportunitiesforpositivesocietaltransformations.Atthesametime,generativeAImodelsaddnewdimensionstoAIriskmanagement,encompassingvariousriskssuchashallucinations,misuse,lackoftraceabilityandharmfuloutput.Therefore,itisessentialtobalancesafety,ethicsandinnovation.2.

Expandedriskguardrails:TheframeworkdetailsrobustguardrailstobeconsideredatdifferentstepsofthegenerativeAIlifecycle,emphasizingpreventionratherthanmitigation.Thisbrie?ngpaperidenti?esalistofchallengestoachievingthisbalanceinpractice,suchaslackofacohesiveviewofthegenerativeAImodellifecycleandambiguityintermsofthedeploymentandperceivedeffectivenessofvaryingsafetyguardrailsthroughoutthelifecycle.Amidthesechallenges,therearesigni?cantopportunities,includinggreaterstandardizationthroughsharedterminologyandbestpractices,facilitatingacommonunderstandingoftheeffectivenessofvariousriskmitigationstrategies.3.

Shift-leftmethodology:Thismethodologyproposestheimplementationofguardrailsattheearliest

stage

possible

in

the

generative

AI

life

cycle.Whileshift-leftisawell-establishedconceptinsoftwareengineering,

its

application

in

the

contextofgenerativeAIpresentsauniqueopportunitytopromotemorewidespreadadoption.Inconclusion,thepaperemphasizestheneedforgreatermultistakeholdercollaborationbetweenindustrystakeholders,policy-makersandThisbrie?ngpaperpresentsthePresidioAIFramework,whichprovidesastructuredapproachtothesafedevelopment,deploymentanduseofgenerativeAI.Indoingso,theframeworkhighlightsgapsandopportunitiesinaddressingsafetyconcerns,viewedfromtheperspectiveoffourprimaryactors:AImodelcreators,AImodeladapters,AImodelusers,andAIapplicationusers.Sharedresponsibility,earlyriskidenti?cationandproactiveriskmanagementthroughtheimplementationofappropriateguardrailsareemphasizedthroughout.organizations.ThePresidioAIFrameworkpromotessharedresponsibility,earlyriskidenti?cationandproactiveriskmanagementingenerativeAIdevelopment,usingguardrailstoensureethicalandresponsibledeployment.Thepaperlaysthefoundationforongoingsafety-relatedworkoftheAIGovernanceAllianceandtheSafeSystemsandTechnologiesworkinggroup.Futureworkwillexpandonthecoreconceptsandcomponentsintroducedinthispaper,

includingtheprovisionofamoreexhaustivelistofknownandnovelThePresidioAIFrameworkconsistsofthreecorecomponents:guardrails,alongwithachecklisttooperationalizetheframeworkacrossthegenerativeAIlifecycle.1/3:PresidioAIFramework

10IntroductionThecurrentAIlandscapeincludesbothchallengesandopportunitiesforprogresstowardssafegenerativeAImodels.Thisbrie?ngpaperoutlinesthePresidioAIdiversity.However,

theavailabilityofallthemodelcomponents(e.g.weights,technicaldocumentationandcode)couldalsoamplifyrisksandreduceguardrails’effectiveness.ThereisaneedforcarefulanalysisofrisksandcommonconsensusamongtheuseofguardrailsFramework,providingastructuredapproachtoaddressingbothtechnicalandproceduralconsiderationsforsafegenerativearti?cialintelligence(AI)models.Theframeworkcentresonfoundationmodelsandincorporatesrisk-mitigationstrategiesthroughouttheentirelifecycle,encompassingcreation,adaptationandeventualretirement.InformedbythoroughresearchintothecurrentAIlandscapeandinputfromamultistakeholdercommunityandpractitioners,theframeworkunderscorestheimportanceofestablishedsafetyguidelinesandrecommendationsviewedthroughatechnicallens.NotablechallengesintheexistinglandscapeimpactingthedevelopmentanddeploymentofsafegenerativeAIinclude:considering

the

gradient

of

release;

that

is,

varying2levelsatwhichAImodelsareaccessibleoncereleased,fromfullyclosedtofullyopen-sourced.Simultaneously,therearesomeidenti?edopportunitiesforprogresstowardssafety,suchas:–Standardization:Bylinkingthetechnicalaspectsateachphaseofdesign,developmentandreleasewiththeircorrespondingrisksandmitigations,thereistheopportunityforbringingattentiontosharedterminologyandbestpractices.Thismaycontributetowardsgreateradoptionofnecessarysafetymeasuresandpromotecommunityharmonizationacrossdifferentstandardsandguidelines.–Fragmentation:Aholisticperspective,whichcoverstheentirelifecycleofgenerativeAImodelsfromtheirinitialdesigntodeploymentandthecontinuousstagesofadaptationanduse,iscurrentlymissing.Thiscanleadtofragmentedperceptionsofthemodel’s

creationandtherisksassociatedwithitsdeployment.–Stakeholdertrustandempowerment:Pursuingclarityandagreementontheexpectedriskmitigationstrategies,wherethesearemosteffectivelylocatedinthemodellifecycleandwhoisaccountableforimplementationpavesthewayforstakeholderstoimplementtheseproactively.Thisimprovessafety,preventsadverseoutcomesforindividualsandsociety,andbuildstrustamongallstakeholders.––Vague

de?nitions:Ambiguityandlackofcommonunderstandingofthemeaningofsafety,risks

(e.g.traceability),andgeneral1safetymeasures(e.g.redteaming)atthefrontierofmodeldevelopment.Guardrailambiguity:Whilethereisagreementontheimportanceofrisk-mitigationstrategies–knownasguardrails–clarityislackingregardingaccountability,effectiveness,actionability,applicability,limitationsandatwhatstagesoftheAIdesign,developmentandreleaselifecyclevaryingguardrailsshouldbeimplemented.Whilethisbrie?ngpaperdetailsthegenerativeAImodellifecyclealongwithsomeguardrails,itisbynomeansexhaustive.Sometopicsoutsidethispaper’s

scopeincludeadiscussionofcurrentorfuturegovernmentregulationsofAIrisksandmitigations(thisiscoveredintheResilientGovernanceworkinggroupbrie?ngpaper)orconsiderationofdownstreamimplementationanduseofspeci?cAIapplications.–Modelaccess:Anopenapproachpresentssigni?cantopportunitiesforinnovation,greateradoptionandincreasedstakeholderpopulation1/3:PresidioAIFramework

11Introducing

the1Presidio

AIFrameworkAstructuredapproachthatemphasizessharedresponsibilityandproactiveriskmitigationbyimplementingappropriateguardrailsearlyinthegenerativeAIlifecycle.Thosereleasing,adaptingorusingfoundationmodelsoftenfacechallengesinin?uencingtheoriginalmodeldesignorsettingupthenecessaryinfrastructureforbuildingfoundationmodels.Thecombi