世卫组织-人工智能健康伦理与治理：大型多模式模型指南（英）-2024.docx

IanizationEthicsandgovernanceofartificialintelligenceforhealthGuidanceonlargemulti-modalmodelsWorldHealthOiganizationEthicsandgovernanceofartificialintelligenceforhealthGuidanceonlargemulti-modalmodelsISBN978-92-4-008475-9(electronicversion)ISBN978-92-4-008476-6(printversion)©WorldHealthOrganization2024Somerightsreserved.ThisworkisavailableundertheCreativeCommonsAttribution-NonCommerciaI-ShareAIike3.0IGOlicence(CCBY-NC-SA3.0IGO;https:/creativecommons.Org/licenses/by-nc-sa/3.0/igo).Underthetermsofthislicence,youmaycopy,redistributeandadapttheworkfornon-commercialpurposes,providedtheworkisappropriatelycited,asindicatedbelow.Inanyuseofthiswork,thereshouldbenosuggestionthatWHOendorsesanyspecificorganization,productsorservices.TheuseoftheWHOlogoisnotpermitted.Ifyouadaptthework,thenyoumustlicenseyourworkunderthesameorequivalentCreativeCommonslicence.Ifyoucreateatranslationofthiswork,youshouldaddthefollowingdisclaimeralongwiththesuggestedcitation:"ThistranslationwasnotcreatedbytheWorldHealthOrganization(WHO).WHOisnotresponsibleforthecontentoraccuracyofthistranslation.TheoriginalEnglisheditionshallbethebindingandauthenticedition,AnymediationrelatingtodisputesarisingunderthelicenceshallbeconductedinaccordancewiththemediationrulesoftheWorldIntellectualPropertyOrganization,(http:/www.wipo.int/amc/en/mediation/rules/).Suggestedcitation.Ethicsandgovernanceofartificialintelligenceforhealth.Guidanceonlargemulti-modalmodels.Geneva:WorldHealthOrganization;2024.1.icence:CCBY-NC-SA3.0IGO.Cataloguing-In-Publication(CIP)data.CIPdataareavailableathttp:/apps.who.int/iris.Sales,rightsandlicensing.TopurchaseWHOpublications,seehttp:/apps.who.int/bookorders.TosubmitrequestsforCommercicensing,seehttp:/www.who.int/about/licensing.Third-partymaterials.Ifyouwishtoreusematerialfromthisworkthatisattributedtoathirdparty,suchastables,figuresorimages,itisyourresponsibilitytodeterminewhetherpermissionisneededforthatreuseandtoobtainpermissionfromthecopyrightholder.Theriskofclaimsresultingfrominfringementofanythird-partyownedcomponentintheworkrestssolelywiththeuser.Generaldisclaimers.ThedesignationsemployedandthepresentationofthematerialinthispublicationdonotimplytheexpressionofanyopinionwhatsoeveronthepartofWHOconcerningthelegalstatusofanycountry,territory,cityorareaorofitsauthorities,orconcerningthedelimitationofitsfrontiersorboundaries.Dottedanddashedlinesonmapsrepresentapproximateborderlinesforwhichtheremaynotyetbefullagreement.Thementionofspecificcompaniesorofcertainmanufacturers*productsdoesnotimplythattheyareendorsedorrecommendedbyWHOinpreferencetoothersofasimilarnaturethatarenotmentioned.Errorsandomissionsexcepted,thenamesofproprietaryproductsaredistinguishedbyinitialcapitalletters.AllreasonableprecautionshavebeentakenbyWHOtoverifytheinformationcontainedinthispublication.However,thepublishedmaterialisbeingdistributedwithoutwarrantyofanykind,eitherexpressedorimplied.Theresponsibilityfortheinterpretationanduseofthemateriallieswiththereader.InnoeventshallWHObeliablefordamagesarisingfromitsuse.Graphicsdesign:JoannaSleigh(ETHZurich,Zurich,Switzerland)1.ayout:ImprimerieCentrale(1.uxembourg)ContentsAcknowledgementsvAbbreviationsviiExecutivesummaryviii1 Introduction11.1 Significanceof1.MMs31.2 WHOguidanceonethicsandgovernanceofAlforhealth4,Applications,challengesandrisksof1.MMs72 Applicationsandchallengesofuseof1.MMsinhealth82.1 Diagnosisandclinicalcare82.2 Patient-centredapplications122.3 Clericalfunctionsandadministrativetasks162.4 Medicalandnursingeducation172.5 Scientificandmedicalresearchanddrugdevelopment173 Riskstohealthsystemsandsocietyandethicalconcernsaboutuseof1.MMs203.1 Healthsystems203.2 Compliancewithregulatoryandlegalrequirements233.3 Societalconcernsandrisks24AEthicsandgovernanceof1.MMsinhealthcareandmedicine314 Designanddevelopmentofgeneral-purposefoundationmodels(1.MMs)344.1 Riskstobeaddressedduringthedevelopmentofgeneral-purposefoundationmodels(1.MMs)344.2 Measuresdeveloperscantaketoaddressriskswithgeneral-purposefoundationmodels(1.MMs)354.3 Governmentlaws,policiesandpublicsectorinvestments394.4 Open-source1.MMs415 Provisionwithgeneral-purposefoundationmodels(1.MMs)455.1 Riskstobeaddressedwhenprovidingahealth-careserviceorapplicationwithageneral-purposefoundationmodel(1.MM)455.2 Measuresthatgovernmentscanintroducetoaddresssuchrisksandethicalprinciplesthatshouldbeupheld466 Deploymentwithgeneral-purposefoundationmodels(1.MMs)536.1 Riskstobeaddressedwhendeployingahealth-careserviceorapplicationwithageneral-purposefoundationmodel(1.MM)536.2 On-goingresponsibilitiesofdevelopersandprovidersduringdeployment546.3 Responsibilitiesofdeployers546.4 Governmentprogrammesandpractices557 1.iabilityfor1.MMs588 Internationalgovernanceof1.MMs60References62Annex.Methods77AcknowledgementsDevelopmentofthisWorldHealthOrganization(WHO)guidancewasledbyAndreasReis(co-leadoftheHealthEthicsandGovernanceunitintheDepartmentofResearchforHealth)andSameerPujari(DepartmentofDigitalHealthandInnovation),undertheoverallguidanceofJohnReeder(Director,ResearchforHealth),Alain1.abrique(Director,DigitalHealthandInnovation)andJeremyFarrar(ChiefScientist).RohitMalpani(consultant,France)wastheleadwriter.Theco-chairsoftheWHOexpertgrouponethicsandgovernanceofAlforhealth,EffyVayena(ETHZurich,Switzerland)andParthaMajumder(IndianStatisticalInstituteandNationalInstituteofBiomedicalGenomics,India),providedoverallguidanceondraftingofthereportandleadershipoftheexpertgroup.WHOisgratefultothefollowingindividualswhocontributedtodevelopmentofthisguidance.WHOexpertgrouponethicsandgovernanceofAlforhealthNajeebAlShorbaji,eHealthDevelopmentAssociation,Amman,Jordan;MariaPazCanales,GlobalPartnersDigital,SantiagodeChile,Chile;ArisaEma,UniversityofTokyo,Tokyo,Japan;AmelGhouila,Bill&MelindaGatesFoundation,Seattle(WA),USA;JenniferGibson,WHOCollaboratingCentreforBioethics,UniversityofToronto,Toronto,Canada;KennethGoodman,InstituteofBioethicsandHealthPolicy,UniversityofMiamiMillerSchoolofMedicines,Miami(F1.)zUSA;MalavikaJayaram,DigitalAsiaHubzSingapore;DaudiJjingo,MakerereUniversity,Kampala,Uganda;TzeYun1.eong,NationalUniversityOfSingaporezSingapore;AIexJohn1.ondon,CarnegieMellonUniversity,Pittsburgh(PA)zUSA;ParthaMajumder,IndianStatisticalInstituteandNationalInstituteofBiomedicalGenomics,Kolkata,India;ThilidziMarwalazUniversityOfJohannesburg,Johannesburg,SouthAfrica;RoliMathur,IndianCouncilofMedicalResearch,Bangalore,India;TimoMinssen,CentreforAdvancedStudiesinBiomedicalInnovation1.aw,Facultyof1.aw,UniversityofCopenhagen,Copenhagen,Denmark;AndrewMorris,HealthDataResearchUKz1.ondon,UnitedKingdom;DanielaPaolotti,ISIFoundation,Turin,Italy;JeromeSingh,UniversityofKwa-ZuluNatal,Durban,SouthAfrica;JeroenvandenHoven,UniversityofDelft,Delft,Netherlands(Kingdomofthe);EffyVayenazETHZurich,Zurich,Switzerland;RobynWhittaker,UniversityofAuckland,Auckland,NewZealand;andYiZeng,ChineseAcademyofSciences,Beijing,China.ObserversDavidGruson,1.uminess,Pans,France;1.eeHibbard,CouncilofEurope,Strasbourg,FranceExternalreviewersOrenAsman,TelAvivUniversity,TelAviv,Israel;I.GlennCohen,Harvard1.awSchool,Boston(MA),USA;AlexandrinePirlotdeCorbion,PrivacyInternational,1.ondon,UnitedKingdom;Rodrigo1.ins,FederalUniversityofPernambuco,Recife,Brazil;DougMcNairzDeputyDirector,IntegratedDevelopment,Bill&MelindaGatesFoundation,Seattle(WA)zUSA;KeymanthriMoodleyzStellenboschUniversity,CapeTown,SouthAfrica;AmirTaI,TelAvivUniversity,TelAviv,Israel;TomWestzPrivacyInternational,1.ondon,UnitedKingdom.ExternalcontributorsBox2(Ethicalconsiderationsfortheuseof1.MMsbychildren)oftheguidancewasdraftedbyVijaythaMuralidharan,AlyssaBurgart,RoxanaDaneshjouandSherriRose,StanfordUniversity,Stanford(CA),USA.Box3(Ethicalconsiderationsassociatedwith1.MMsandtheirimpactonindividualswithdisabilities)oftheguidancewasdraftedbyYonahWelker,independentconsultant,Geneva,Switzerland.Allexternalreviewers,expertsandcontributorsdeclaredtheirinterestsinlinewithWHOpolicies.Noneoftheinterestsdeclaredwereassessedtobesignificant.WHOShadaAl-Salamah,TechnicalOfficer,DepartmentofDigitalHealthandInnovation,Geneva;MariamOtmaniDelBarrio,Scientist,SpecialProgrammeonTropicalDiseasesResearch,Geneva;MarceloD'Agostino,UnitChief,InformationSystemsandDigitalHealth,WHORegionalOfficefortheAmericas,Washington(DC);JeremyFarrar,ChiefScientist,Geneva;ClaytonHamilton,TechnicalOfficer,WHORegionalOfficeforEurope,Copenhagen,Denmark;KanikaKalra,Consultant,DepartmentofDigitalHealthandInnovation,Geneva;AhmedMohamedAminMandilzCoordinator,ResearchandInnovation,WHORegionalOfficefortheEasternMediterranean,Cairo;IssaT.Matta,1.egalAffairs,Geneva;JoseEduardoDiazMendoza,Consultant,DepartmentofDigitalHealthandInnovation,Geneva;MohammedHassanNour,TechnicalOfficer,DepartmentofDigitalHealthandInnovation,WHORegionalOfficefortheEasternMediterranean,Cairo;DeniseSchaletzTechnicalOfficer,DepartmentofDigitalHealthandInnovation,Geneva;YuZhao,TechnicalOfficer,DepartmentofDigitalHealthandInnovation,Geneva.AbbreviationsAlartificialintelligence1.MMlargemulti-modalmodelUSAUnitedStatesofAmericaExecutivesummaryArtificialIntelligence(AI)referstothecapabilityofalgorithmsintegratedintosystemsandtoolstolearnfromdatasothattheycanperformautomatedtaskswithoutexplicitprogrammingofeverystepbyahuman.GenerativeAlisacategoryofAltechniquesinwhichalgorithmsaretrainedondatasetsthatcanbeusedtogeneratenewcontent,suchastext,imagesorvideo.ThisguidanceaddressesonetypeofgenerativeAl,largemulti-modalmodels(1.MMs),whichcanacceptoneormoretypeofdatainputandgeneratediverseoutputsthatarenotlimitedtothetypeofdatafedintothealgorithm.Ithasbeenpredictedthat1.MMswillhavewideuseandapplicationinhealthcare,scientificresearch,publichealthanddrugdevelopment.1.MMsarealsoknownas/"general-purposefoundationmodels”,althoughitisnotyetprovenwhether1.MMscanaccomplishawiderangeoftasksandpurposes.1.MMshavebeenadoptedfasterthananyconsumerapplicationinhistory.Theyarecompellingbecausetheyfacilitatehuman-computerinteractiontomimichumancommunicationandtogenerateresponsestoqueriesordatainputsthatmayappearhuman-likeandauthoritative.Withrapidconsumeradoptionanduptakeandinviewofitspotentialtodisruptcoresocialservicesandeconomicsectors,manylargetechnologycompanies,start-upsandgovernmentsareinvestinginandCompetingtoguidethedevelopmentOfgenerativeAI.In2021,WHOpublishedcomprehensiveguidance(1)ontheethicsandgovernanceofAlforhealth.WHOconsulted20leadingexpertsinAl,whoidentifiedbothpotentialbenefitsandpotentialrisksofuseofAlinhealthcareandissuedsixprinciplesarrivedatbyconsensusforconsiderationinthepoliciesandpracticesOfgovernmentszdevelopers,andprovidersthatareusingAl.TheprinciplesshouldguidethedevelopmentanddeploymentofAlinhealthcarebyawiderangeofstakeholders,includinggovernments,publicsectoragencies,researchers,companiesandimplementers.Theprinciplesare:(1)protectautonomy;(2)promotehumanwell-being,humansafetyandthepublicinterest;(3)ensuretransparency,zzexplainabilitywandintelligibility;(4)fosterresponsibilityandaccountability;(5)ensureinclusivenessandequity;and(6)promoteAlthatisresponsiveandsustainable(Figure1).WHOisissuingthisguidancetoassistMemberStatesinmappingthebenefitsandchallengesassociatedwithuseof1.MMsforhealthandindevelopingpoliciesandpracticesforappropriatedevelopment,provisionanduse.Theguidanceincludesrecommendationsforgovernance,withincompanies,bygovernmentsandthroughinternationalcollaboration,alignedwiththeguidingprinciples.Theprinciplesandrecommendations,whichaccountfortheuniquewaysinwhichhumanscanusegenerativeAlforhealth,arethebasisofthisguidance.Figure 1: WHOconsensusethicalprinciplesforuseofAlforhealthProtectautonomy。2。FosterresponsibilityandaccountabilityPromotehumanwell-being,humansafetyandthepublicinterest国EnsureinclusivenessandequityEnsuretransparency,explainabilityandintelligibility魄PromoteAIthatisresponsiveandsustainable浸Applications,challengesandrisksoflargemulti-modalmodelsThepotentialapplicationsof1.MMsinhealthcarearesimilartothoseofotherformsofAl,yethow1.MMsareaccessedandusedisnew,withbothnovelbenefitsandrisksthatsocieties,healthsystemsandend-usersmaynotyetbepreparedtoaddressfully.Table1summarizesthemainapplicationsof1.MMsandtheirpotentialbenefitsandrisks.Thesystemicrisksassociatedwithuseof1.MMsincluderisksthatcouldaffecthealthsystems(Table2).Broaderregulatoryandsystemicriskscouldemergewithuseof1.MMs.Oneconcern(beingexaminedbyseveraldataprotectionauthorities)iswhether1.MMscomplywithexistinglegalorregulatoryregimes,includinginternationalhumanrightsobligations,andwithnationaldataprotectionregulations.Algorithmsmightnotcomplywithsuchlawsbecauseofthewayinwhichdataarecollectedtotrain1.MMs,themanagementandprocessingofdatathathavebeencollected(orputinto1.MMsbyendusers),thetransparencyandaccountabilityofentitiesthatdevelop1.MMs,andthepossibilitythat1.MMszzhallucinate,1.MMscouldalsobenon-Compliantwithconsumerprotectionlaws.Broadersocietalrisksassociatedwiththegrowinguseof1.MMs(includingandbeyondtheuseofsuchalgorithmsinhealthcare)includethefactthat1.MMsareoftendevelopedanddeployedbylargetechnologycompanies,duepartlytothesignificantcomputing,data,humanandfinancialresourcerequiredfordevelopmentof1.MMs.Thismayreinforcethedominanceofthesecompaniesvis-a-vissmallerenterprisesandgovernmentswithrespecttothedevelopmentanduseofAlzincludingthefocusofAlresearchinthepublicandprivatesectors.Additionalconcernsaboutthepotentialdominanceoflargetechnologycompaniesincludeinsufficientcorporatecommitmenttoethicsandtransparency.NewvoluntaryTable1.Potentialbenefitsandrisksinvarioususesof1.MMSinhealthcareUsePotentialorproposedbenefitsPotentialrisksIDiagnosisandclinicalcareAssistinmanagingcomplexcasesandreviewofroutinediagnosesReducethecommunicationworkloadofhealth-careproviders("'keyboardliberation,>ProvidenovelinsightsandreportsfromvariousunstructuredformsofhealthdataInaccurate,incompleteorfalseresponsesPoorqualitytrainingdataBiasoftrainingdataandresponses)AutomationbiasDegradationofskills(ofhealth-careprofessionals)Informedconsent(ofpatients)Patient-guideduseGenerateinformationtoimproveunderstandingofamedicalcondition(asapatientorasacaregiver)VirtualhealthassistantClinicaltrialenrolmentInaccurate,incompleteorfalsestatementsManipulationPrivacy1.essinteractionbetweencliniciansandpatientsEpistemicinjusticeRiskofdeliveryofcareoutsidethehealthsystemClericalandadministrativetasksAssistwithpaperworkanddocumentationrequiredforclinicalcareAssistinlanguagetranslationCompletionofelectronichealthrecordsDraftclinicalnotesafterapatientvisitInaccuraciesanderrorsInconsistentresponsesdependingonpromptsMedicalandnursingeducationDynamictextssuitedtoeachstudent'sneedsSimulatedconversationtoimprovecommunicationandtopractiseindiversesituationsandwithdiversepatientsResponsestoquestionsaccompaniedbychain-of-thoughtreasoningContributetoautomationbiasErrorsorfalseinformationunderm