




版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
Ethicsandgovernanceof
artificialintelligenceforhealth
Guidanceonlargemulti-modalmodels
worldHealthorganization
Ethicsandgovernanceof
artificialintelligenceforhealth
Guidanceonlargemulti-modalmodels
Ethicsandgovernanceofartificialintelligenceforhealth.Guidanceonlargemulti-modalmodels
ISBN978-92-4-008475-9(electronicversion)
ISBN978-92-4-008476-6(printversion)
©WorldHealthOrganization2024
Somerightsreserved.ThisworkisavailableundertheCreativeCommonsAttribution-NonCommercial-ShareAlike3.0IGOlicence(CCBY-NC-SA3.0IGO;
/licenses/by-nc-sa/3.0/igo
).
Underthetermsofthislicence,youmaycopy,redistributeandadapttheworkfornon-commercialpurposes,providedtheworkisappropriatelycited,asindicatedbelow.Inanyuseofthiswork,thereshouldbeno
suggestionthatWHOendorsesanyspecificorganization,productsorservices.TheuseoftheWHOlogois
notpermitted.Ifyouadaptthework,thenyoumustlicenseyourworkunderthesameorequivalentCreativeCommonslicence.Ifyoucreateatranslationofthiswork,youshouldaddthefollowingdisclaimeralongwiththesuggestedcitation:“ThistranslationwasnotcreatedbytheWorldHealthOrganization(WHO).WHOisnotresponsibleforthecontentoraccuracyofthistranslation.TheoriginalEnglisheditionshallbethebindingandauthenticedition”.
AnymediationrelatingtodisputesarisingunderthelicenceshallbeconductedinaccordancewiththemediationrulesoftheWorldIntellectualProperty
Organization.(
/amc/en/mediation/rules/
).
Suggestedcitation.Ethicsandgovernanceofartificialintelligenceforhealth.Guidanceonlargemulti-modalmodels.Geneva:WorldHealthOrganization;2024.Licence:
CCBY-NC-SA3.0IGO
.
Cataloguing-in-Publication(CIP)data.CIPdataareavailableat
/iris
.
Sales,rightsandlicensing.TopurchaseWHOpublications,see
/bookorders
.Tosubmitrequestsforcommercialuseandqueriesonrightsandlicensing,see
/about/licensing
.
Third-partymaterials.Ifyouwishtoreusematerialfromthisworkthatisattributedtoathirdparty,suchastables,figuresorimages,itisyourresponsibilitytodeterminewhetherpermissionisneededforthatreuseandtoobtainpermissionfromthecopyrightholder.Theriskofclaimsresultingfrominfringementofanythird-
partyownedcomponentintheworkrestssolelywiththeuser.
Generaldisclaimers.ThedesignationsemployedandthepresentationofthematerialinthispublicationdonotimplytheexpressionofanyopinionwhatsoeveronthepartofWHOconcerningthelegalstatusofany
country,territory,cityorareaorofitsauthorities,orconcerningthedelimitationofitsfrontiersorboundaries.
Dottedanddashedlinesonmapsrepresentapproximateborderlinesforwhichtheremaynotyetbefullagreement.
Thementionofspecificcompaniesorofcertainmanufacturers’productsdoesnotimplythattheyareendorsedorrecommendedbyWHOinpreferencetoothersofasimilarnaturethatarenotmentioned.Errorsand
omissionsexcepted,thenamesofproprietaryproductsaredistinguishedbyinitialcapitalletters.
AllreasonableprecautionshavebeentakenbyWHOtoverifytheinformationcontainedinthispublication.
However,thepublishedmaterialisbeingdistributedwithoutwarrantyofanykind,eitherexpressedorimplied.Theresponsibilityfortheinterpretationanduseofthemateriallieswiththereader.InnoeventshallWHObeliablefordamagesarisingfromitsuse.
Graphicsdesign:JoannaSleigh(ETHZurich,Zurich,Switzerland)
Layout:ImprimerieCentrale(Luxembourg)
Contentsiii
Contents
Acknowledgements v
Abbreviations vii
Executivesummary viii
1Introduction 1
1.1SignificanceofLMMs 3
1.2WHOguidanceonethicsandgovernanceofAIforhealth 4
I.Applications,challengesandrisksofLMMs 7
2ApplicationsandchallengesofuseofLMMsinhealth 8
2.1Diagnosisandclinicalcare 8
2.2Patient-centredapplications 12
2.3Clericalfunctionsandadministrativetasks 16
2.4Medicalandnursingeducation 17
2.5Scientificandmedicalresearchanddrugdevelopment 17
3Riskstohealthsystemsandsocietyandethicalconcernsaboutuse
ofLMMs 20
3.1Healthsystems 20
3.2Compliancewithregulatoryandlegalrequirements 23
3.3Societalconcernsandrisks 24
II.EthicsandgovernanceofLMMsinhealthcareandmedicine 31
4Designanddevelopmentofgeneral-purposefoundationmodels
(LMMs) 34
4.1Riskstobeaddressedduringthedevelopmentofgeneral-
purposefoundationmodels(LMMs) 34
4.2Measuresdeveloperscantaketoaddressriskswithgeneral-
purposefoundationmodels(LMMs) 35
4.3Governmentlaws,policiesandpublicsectorinvestments 39
4.4Open-sourceLMMs 41
ivEthicsandgovernanceofartificialintelligenceforhealth.Guidanceonlargemulti-modalmodels
5Provisionwithgeneral-purposefoundationmodels(LMMs) 45
5.1Riskstobeaddressedwhenprovidingahealth-careserviceor
applicationwithageneral-purposefoundationmodel(LMM) 45
5.2Measuresthatgovernmentscanintroducetoaddresssuchrisks
andethicalprinciplesthatshouldbeupheld 46
6Deploymentwithgeneral-purposefoundationmodels(LMMs) 53
6.1Riskstobeaddressedwhendeployingahealth-careserviceor
applicationwithageneral-purposefoundationmodel(LMM) 53
6.2On-goingresponsibilitiesofdevelopersandprovidersduring
deployment 54
6.3Responsibilitiesofdeployers 54
6.4Governmentprogrammesandpractices 55
7LiabilityforLMMs 58
8InternationalgovernanceofLMMs 60
References 62
Annex.Methods 77
Acknowledgementsv
Acknowledgements
DevelopmentofthisWorldHealthOrganization(WHO)guidancewasledbyAndreasReis
(co-leadoftheHealthEthicsandGovernanceunitintheDepartmentofResearchforHealth)andSameerPujari(DepartmentofDigitalHealthandInnovation),undertheoverallguidanceofJohnReeder(Director,ResearchforHealth),AlainLabrique(Director,DigitalHealthand
Innovation)andJeremyFarrar(ChiefScientist).
RohitMalpani(consultant,France)wastheleadwriter.Theco-chairsoftheWHOexpertgrouponethicsandgovernanceofAIforhealth,EffyVayena(ETHZurich,Switzerland)andParthaMajumder(IndianStatisticalInstituteandNationalInstituteofBiomedicalGenomics,India),providedoverallguidanceondraftingofthereportandleadershipoftheexpertgroup.
WHOisgratefultothefollowingindividualswhocontributedtodevelopmentofthisguidance.
WHOexpertgrouponethicsandgovernanceofAIforhealth
NajeebAlShorbaji,eHealthDevelopmentAssociation,Amman,Jordan;MariaPazCanales,
GlobalPartnersDigital,SantiagodeChile,Chile;ArisaEma,UniversityofTokyo,Tokyo,Japan;AmelGhouila,Bill&MelindaGatesFoundation,Seattle(WA),USA;JenniferGibson,WHO
CollaboratingCentreforBioethics,UniversityofToronto,Toronto,Canada;KennethGoodman,InstituteofBioethicsandHealthPolicy,UniversityofMiamiMillerSchoolofMedicines,Miami(FL),USA;MalavikaJayaram,DigitalAsiaHub,Singapore;DaudiJjingo,MakerereUniversity,
Kampala,Uganda;TzeYunLeong,NationalUniversityofSingapore,Singapore;AlexJohn
London,CarnegieMellonUniversity,Pittsburgh(PA),USA;ParthaMajumder,IndianStatisticalInstituteandNationalInstituteofBiomedicalGenomics,Kolkata,India;ThilidziMarwala,
UniversityofJohannesburg,Johannesburg,SouthAfrica;RoliMathur,IndianCouncilof
MedicalResearch,Bangalore,India;TimoMinssen,CentreforAdvancedStudiesinBiomedicalInnovationLaw,FacultyofLaw,UniversityofCopenhagen,Copenhagen,Denmark;Andrew
Morris,HealthDataResearchUK,London,UnitedKingdom;DanielaPaolotti,ISIFoundation,Turin,Italy;JeromeSingh,UniversityofKwa-ZuluNatal,Durban,SouthAfrica;Jeroenvan
denHoven,UniversityofDelft,Delft,Netherlands(Kingdomofthe);EffyVayena,ETHZurich,Zurich,Switzerland;RobynWhittaker,UniversityofAuckland,Auckland,NewZealand;andYiZeng,ChineseAcademyofSciences,Beijing,China.
Observers
DavidGruson,Luminess,Paris,France;LeeHibbard,CouncilofEurope,Strasbourg,France
viEthicsandgovernanceofartificialintelligenceforhealth.Guidanceonlargemulti-modalmodels
Externalreviewers
OrenAsman,TelAvivUniversity,TelAviv,Israel;I.GlennCohen,HarvardLawSchool,Boston(MA),USA;AlexandrinePirlotdeCorbion,PrivacyInternational,London,UnitedKingdom;
RodrigoLins,FederalUniversityofPernambuco,Recife,Brazil;DougMcNair,DeputyDirector,IntegratedDevelopment,Bill&MelindaGatesFoundation,Seattle(WA),USA;Keymanthri
Moodley,StellenboschUniversity,CapeTown,SouthAfrica;AmirTal,TelAvivUniversity,TelAviv,Israel;TomWest,PrivacyInternational,London,UnitedKingdom.
Externalcontributors
Box2(EthicalconsiderationsfortheuseofLMMsbychildren)oftheguidancewasdraftedbyVijaythaMuralidharan,AlyssaBurgart,RoxanaDaneshjouandSherriRose,Stanford
University,Stanford(CA),USA.Box3(EthicalconsiderationsassociatedwithLMMsandtheirimpactonindividualswithdisabilities)oftheguidancewasdraftedbyYonahWelker,independentconsultant,Geneva,Switzerland.
Allexternalreviewers,expertsandcontributorsdeclaredtheirinterestsinlinewithWHOpolicies.Noneoftheinterestsdeclaredwereassessedtobesignificant.
WHO
ShadaAl-Salamah,TechnicalOfficer,DepartmentofDigitalHealthandInnovation,Geneva;
MariamOtmaniDelBarrio,Scientist,SpecialProgrammeonTropicalDiseasesResearch,
Geneva;MarceloD'Agostino,UnitChief,InformationSystemsandDigitalHealth,WHORegionalOfficefortheAmericas,Washington(DC);JeremyFarrar,ChiefScientist,Geneva;Clayton
Hamilton,TechnicalOfficer,WHORegionalOfficeforEurope,Copenhagen,Denmark;KanikaKalra,Consultant,DepartmentofDigitalHealthandInnovation,Geneva;AhmedMohamedAminMandil,Coordinator,ResearchandInnovation,WHORegionalOfficefortheEastern
Mediterranean,Cairo;IssaT.Matta,LegalAffairs,Geneva;JoseEduardoDiazMendoza,
Consultant,DepartmentofDigitalHealthandInnovation,Geneva;MohammedHassanNour,TechnicalOfficer,DepartmentofDigitalHealthandInnovation,WHORegionalOfficefor
theEasternMediterranean,Cairo;DeniseSchalet,TechnicalOfficer,DepartmentofDigitalHealthandInnovation,Geneva;YuZhao,TechnicalOfficer,DepartmentofDigitalHealthandInnovation,Geneva.
Acknowledgementsvii
Abbreviations
AIartificialintelligence
LMMlargemulti-modalmodelUSAUnitedStatesofAmerica
viiiEthicsandgovernanceofartificialintelligenceforhealth.Guidanceonlargemulti-modalmodels
Executivesummary
ArtificialIntelligence(AI)referstothecapabilityofalgorithmsintegratedintosystems
andtoolstolearnfromdatasothattheycanperformautomatedtaskswithoutexplicit
programmingofeverystepbyahuman.GenerativeAIisacategoryofAItechniquesin
whichalgorithmsaretrainedondatasetsthatcanbeusedtogeneratenewcontent,suchastext,imagesorvideo.ThisguidanceaddressesonetypeofgenerativeAI,largemulti-modal
models(LMMs),whichcanacceptoneormoretypeofdatainputandgeneratediverseoutputsthatarenotlimitedtothetypeofdatafedintothealgorithm.IthasbeenpredictedthatLMMswillhavewideuseandapplicationinhealthcare,scientificresearch,publichealthanddrug
development.LMMsarealsoknownas“general-purposefoundationmodels”,althoughitisnotyetprovenwhetherLMMscanaccomplishawiderangeoftasksandpurposes.
LMMshavebeenadoptedfasterthananyconsumerapplicationinhistory.Theyarecompellingbecausetheyfacilitatehuman–computerinteractiontomimichumancommunicationandtogenerateresponsestoqueriesordatainputsthatmayappearhuman-likeandauthoritative.Withrapidconsumeradoptionanduptakeandinviewofitspotentialtodisruptcoresocial
servicesandeconomicsectors,manylargetechnologycompanies,start-upsandgovernmentsareinvestinginandcompetingtoguidethedevelopmentofgenerativeAI.
In2021,WHOpublishedcomprehensiveguidance
(1)
ontheethicsandgovernanceofAIfor
health.WHOconsulted20leadingexpertsinAI,whoidentifiedbothpotentialbenefitsand
potentialrisksofuseofAIinhealthcareandissuedsixprinciplesarrivedatbyconsensusforconsiderationinthepoliciesandpracticesofgovernments,developers,andprovidersthat
areusingAI.TheprinciplesshouldguidethedevelopmentanddeploymentofAIinhealthcarebyawiderangeofstakeholders,includinggovernments,publicsectoragencies,researchers,companiesandimplementers.Theprinciplesare:(1)protectautonomy;(2)promotehumanwell-being,humansafetyandthepublicinterest;(3)ensuretransparency,“explainability”andintelligibility;(4)fosterresponsibilityandaccountability;(5)ensureinclusivenessandequity;and(6)promoteAIthatisresponsiveandsustainable(Figure1).
WHOisissuingthisguidancetoassistMemberStatesinmappingthebenefitsandchallenges
associatedwithuseofLMMsforhealthandindevelopingpoliciesandpracticesfor
appropriatedevelopment,provisionanduse.Theguidanceincludesrecommendationsforgovernance,withincompanies,bygovernmentsandthroughinternationalcollaboration,alignedwiththeguidingprinciples.Theprinciplesandrecommendations,whichaccountfortheuniquewaysinwhichhumanscanusegenerativeAIforhealth,arethebasisof
thisguidance.
Executivesummaryix
Figure1:WHOconsensusethicalprinciplesforuseofAIforhealth
Protectautonomy
Ensuretransparency,
explainabilityand
intelligibility
Promotehuman
well-being,humansafetyandthepublicinterest
Fosterresponsibilityandaccountability
PromoteAIthatis
responsiveand
sustainable
Ensureinclusivenessandequity
Applications,challengesandrisksoflargemulti-modalmodels
ThepotentialapplicationsofLMMsinhealthcarearesimilartothoseofotherformsofAI,yethowLMMsareaccessedandusedisnew,withbothnovelbenefitsandrisksthatsocieties,healthsystemsandend-usersmaynotyetbepreparedtoaddressfully.Table1summarizesthemainapplicationsofLMMsandtheirpotentialbenefitsandrisks.
ThesystemicrisksassociatedwithuseofLMMsincluderisksthatcouldaffecthealthsystems(Table2).
BroaderregulatoryandsystemicriskscouldemergewithuseofLMMs.Oneconcern(being
examinedbyseveraldataprotectionauthorities)iswhetherLMMscomplywithexistinglegalorregulatoryregimes,includinginternationalhumanrightsobligations,andwithnational
dataprotectionregulations.AlgorithmsmightnotcomplywithsuchlawsbecauseofthewayinwhichdataarecollectedtotrainLMMs,themanagementandprocessingofdatathathavebeencollected(orputintoLMMsbyendusers),thetransparencyandaccountabilityofentitiesthatdevelopLMMs,andthepossibilitythatLMMs“hallucinate”.LMMscouldalsobenon-
compliantwithconsumerprotectionlaws.
BroadersocietalrisksassociatedwiththegrowinguseofLMMs(includingandbeyondtheuseofsuchalgorithmsinhealthcare)includethefactthatLMMsareoftendevelopedand
deployedbylargetechnologycompanies,duepartlytothesignificantcomputing,data,
humanandfinancialresourcerequiredfordevelopmentofLMMs.Thismayreinforcethe
dominanceofthesecompaniesvis-a-vissmallerenterprisesandgovernmentswithrespecttothedevelopmentanduseofAI,includingthefocusofAIresearchinthepublicandprivatesectors.Additionalconcernsaboutthepotentialdominanceoflargetechnologycompaniesincludeinsufficientcorporatecommitmenttoethicsandtransparency.Newvoluntary
xEthicsandgovernanceofartificialintelligenceforhealth.Guidanceonlargemulti-modalmodels
Table1.PotentialbenefitsandrisksinvarioususesofLMMsinhealthcare
Use
Potentialorproposedbenefits
Potentialrisks
Diagnosisand
clinicalcare
Assistinmanagingcomplexcasesandreviewofroutinediagnoses
Reducethecommunicationworkloadofhealth-careproviders(“keyboardliberation”)
Providenovelinsightsandreportsfromvariousunstructuredformsofhealth
data
Inaccurate,incompleteorfalseresponses
Poorqualitytrainingdata
Bias(oftrainingdataandresponses)Automationbias
Degradationofskills(ofhealth-careprofessionals)
Informedconsent(ofpatients)
Patient-guideduse
Generateinformationtoimprove
understandingofamedicalcondition(asapatientorasacaregiver)
Virtualhealthassistant
Clinicaltrialenrolment
Inaccurate,incompleteorfalse
statements
ManipulationPrivacy
Lessinteractionbetweencliniciansandpatients
Epistemicinjustice
Riskofdeliveryofcareoutsidethehealthsystem
Clericaland
administrativetasks
Assistwithpaperworkand
documentationrequiredforclinicalcareAssistinlanguagetranslation
CompletionofelectronichealthrecordsDraftclinicalnotesafterapatientvisit
Inaccuraciesanderrors
Inconsistentresponsesdependingonprompts
Medical
andnursing
education
Dynamictextssuitedtoeachstudent’sneeds
Simulatedconversationtoimprovecommunicationandtopractiseindiversesituationsandwithdiversepatients
Responsestoquestionsaccompaniedbychain-of-thoughtreasoning
Contributetoautomationbias
Errorsorfalseinformationunderminethequalityofmedicaleducation
Newburdenoflearningdigitalskills
Scientific
research
anddrug
development
Generateinsightsfromscientificdataandresearch
Generatetextforuseinscientificarticles,manuscriptsubmissionorpeer-review
AnalyseandsummarizedataforresearchProofreading
Denovodrugdesign
Cannotholdalgorithmsaccountableforcontent
Algorithmsencodebiastowardsthe
perspectivesofhigh-incomecountriesGenerateinformationand/orreferencesthatdonotexist
Underminekeytenetsofscientificresearch,suchaspeerreview
Exacerbatedifferentialaccesstoscientificknowledge
Executivesummaryxi
Table2.RiskstohealthsystemsassociatedwithuseofLMMsinhealthcare
Typeofrisk
Description
OverestimationofthebenefitsofLMMs
Theremaybeatendencyto‘technologicalsolutionism’,orover-estimationofthebenefitsofLMMswhileignoringordownplayingchallengesinitsuse,includingitssafety,efficacyandutility.
Accessibilityandaffordability
EquitableaccesstoLMMsmaybelackingforseveralreasons,includingthe“digitaldivide”andsubscriptionfeestoaccessLMMs.
System-widebiases
Useofever-largerdatasetscouldincreasebiasesencodedinLMMs,whichcouldbeautomatedthroughoutahealth-caresystem.
Impactsonlabour
UseofLMMscouldleadtojoblossesinsomecountriesandrequirehealthworkerstoretrainandadjusttouseofLMMs.Dataannotationandfilteringcanleadtolowwagesandtountreatedpsychologicaldistress.
Dependenceofhealthsystemsonill-suitedLMMs
DependenceonLMMscouldmakehealthsystemsvulnerableifLMMsarenotmaintainedor(inlow-andmiddle-incomecountries)areupdatedonlyforuseinhigh-incomecountries.Furthermore,lackofpreservationandprotectionofprivacyandconfidentialitycouldunderminetrustinhealth-caresystemsbypeoplewhoarenotconfidentthattheirprivacywillbe
protected.
Cybersecurityrisks
MaliciousattacksorhackingcouldunderminesafetyandtrustintheuseofLMMsinhealthcare.
commitmentsbysuchcompanies,withoneanotherandwithgovernments,couldmitigate
severalrisksintheshort-termbutarenotanalternativetogovernmentaloversightthatmighteventuallybeenacted.
AnothersocietalriskisthecarbonandwaterfootprintsofLMMs,whichlikeotherformsofAI,requirebothsignificantenergyandcontributetoAI’sgrowingwaterfootprint.WhileLMMsandotherformsofAIcanprovideimportantsocietalbenefits,thegrowingcarbonfootprintmaybecomeamajorcontributortoclimatechange,andincreasingwaterconsumptioncanhaveafurthernegativeimpactinwater-stressedcommunities.Anothersocietalriskassociated
withtheemergenceofLMMsisthat,byprovidingplausibleresponsesthatareincreasinglyconsideredasourceofknowledge,LMMsmayeventuallyunderminehumanepistemic
authority,includinginthedomainsofhealthcare,scienceandmedicine.
EthicsandgovernanceofLMMsinhealthcareandmedicine
LMMscanbeconsideredproductsofaseries(orchain)ofdecisionsonprogrammingand
productdevelopmentbyoneormoreactors(Figure2).DecisionsmadeateachstageofanAIvaluechainmayhavebothdirectandindirectconsequencesonthosewhoparticipateinthedevelopment,deploymentanduseofLMMsdownstream.Thedecisionscanbeinfluencedandregulatedbygovernmentsbyenactingandenforcinglawsandpoliciesnationally,regionallyandglobally.
xiiEthicsandgovernanceofartificialintelligenceforhealth.Guidanceonlargemulti-modalmodels
Figure2:Valuechainofthedevelopment,provisionanddeploymentofLMMs
Provision
Providers
•Companies
•Not-for-profitentities
•Academicinstitutions
•Healthproviders
•Developers(verticallyintegrated)
Adapt/integrateLMMs
•Finetuning
•Integrateintoso什waresystems
•Organizeintoaregulatedorverifiedformat(e.g.
withplug-ins)
CF
Deployment
Deployers
•MinistriesofHealth
•Pharmaceuticalcompanies
•Healthsystems
•Hospitals
•Individualproviders
•Developers(verticallyintegrated)
•Providers(verticallyintegrated)
UseLMMs
•Diagnosisandclinicalcare
•Patientcenteredcare
•Clericaland
administrativetasks
•Medicalandnursingeducation
•Scientificresearch
•Drugdevelopment
Development
Developers
•Largetechnologycompanies
•Smalltechnologycompanies
•Universities
•Scientificcollaborations
•Public-privateconsortiums
•Nationalhealthsystems
DevelopLMMs
•General-purpose
foundationmodels(largemulti-modalmodels)
•Healthormedical
domainspecific
foundationalmodels
•Healthsystemspecificfoundationalmodels
TheAIvaluechainoftenbeginsinalargetechnologycompany,referredtoasa“developer”
inthisguidance.Thedevelopercouldalsobeauniversity,asmallertechnologycompany,
nationalhealthsystems,public-privateconsortiumsorotherentitiesthathavetheresourcesandcapacitytouseseveralinputs,whichcomprisethe“AIinfrastructure”,suchasdata,
computingpowerandAIexpertise,todevelopgeneral-purposefoundationmodels(aterm
usedbygovernmentstodescribeLMMsinlegislationandregulation).Thesemodelscanbe
useddirectlytoperformvarious,oftenunanticipatedtasks,includingthoserelatedtohealthcare.Severalgeneral-purposefoundationmodelsaretrainedspecificallyforuseinhealthcareandmedicine.
Ageneral-purposefoundationmodelcanbeusedbyathirdparty(a“provider”)throughanactiveprogramminginterfaceforaspecificpurposeoruse.Thisinvolves(i)fine-tuninganewLMM,whichmayrequireadditionaltrainingofthefoundationmodel;(ii)integratingtheLMMintoapplicationsoralargersoftwaresystemtoprovideaservicetousers;or(iii)integrating
Executivesummaryxiii
componentsknownas“plug-ins”tochannel,filterandorganizetheLMMsintoformalorregulatedformatstogenerate“digestible”results.1
Thereafter,theprovidercanmarketaproductorservicebasedontheLMMtoacustomer(or“deployer”),suchasaministryofhealth,ahealth-caresystem,ahospital,apharmaceuticalcompanyorevenanindividual,suchasahealth-careprovider.Thecustomerwhoacquiresorlicensestheproductorapplicationmaythenuseitdirectlyforpatients,health-careproviders,otherentitiesinthehealthsystem,laypeopleorinitsownbusiness.Thevaluechaincanbe“verticallyintegrated”,sothatacompany(orotherentity,suchasanationalhealthsystem)
thatcollectsdataandtrainsageneral-purposefoundationmodelcanmodifytheLMMforaparticularuseandprovidetheapplicationdirectlytousers.
Governanceisameansforenshriningethicalprinciplesandhumanrightsobligationsthroughexistinglawsandpoliciesandthroughneworrevisedlaws,norms,internalcodesofpracticeandproceduresfordevelopersandinternationalagreementsandframeworks.
OnewayofframingthegovernanceofLMMsisinthethreestagesoftheAIvaluechain:(i)
thedesignanddevelopmentofgeneral-purposefoundationmodelsorLMMs;(ii)provisionofaservice,applicationorproductbasedonageneral-purposefoundationmodel;and(iii)deploymentofahealth-careserviceorapplication.Inthisguidance,eachphaseisexaminedwithrespecttothreeareasofenquiry:
•Whatrisks(describedabove)shouldbeaddressedateachstageofthevaluechain,andwhichactorsarebestplacedtoaddressthoserisks?
•Whatcanarelevantactordotoaddresstherisks,andwhichethicalprinciplesmustbeupheld?
•Whatistheroleofgovernment,includingrelevantlaws,policiesandregulations?
CertainriskscanbeaddressedateachphaseoftheAIvaluechain,andcertainactorsarelikelytoplaymoreimportantrolesinmitigatingeachriskandupholdingethicalvalues.Whilethereislikelytobedisagreementandtensionaboutwhereresponsibilityrestsbetweendevelopers,providersanddeployers,thereareclearareasinwhicheachactorisbestplacedoristheonlyentitywiththecapacitytoaddressapotentialoractualrisk.
Designanddevelopmentofgeneral-purposefoundationmodels(LMMs)
Duringthedesignanddevelopmentofgeneral-purposefoundationmodels,theresponsibilityrestswiththedevelopers.Governmentsbeartheresponsibilitytosetlawsandstandardstorequireorforbidcertainpractices.Section4ofthisguidanceprovidesrecommendationstohelpaddressrisksandmaximizebenefitsduringthedevelopmentofLMMs.
1CommunicationfromLeongTze-Yun,WHOexpertontheethicsandgovernanceofAIforhealth.
xivEthicsandgovernanceofartificialintelligenceforhealth.Guidanceonlargemulti-modalmodels
Provisionwithgeneral-purposefoundationmodels(LMMs)
Duringprovisionofaservic
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 2025版石材施工合同大全:范本与深度解读
- 二零二五版抖音平台短视频内容审核与合规合作协议
- 2025年度宅基地使用权转让及配套基础设施建设合同
- 二零二五年度汽车商业险融资担保协议
- 二零二五版物流仓储智能管理系统建筑工程施工框架协议
- 广告牌维修工程施工合同书(2025版)
- 家庭新房装修合同2025年
- 公司业务保密协议2025年
- 工程追加合同范本2025年
- 合伙股份协议书范本2025年
- 粤语教学课件
- 2025至2030中医医院行业市场发展分析及前景趋势与投资机会报告
- 音响售后质保合同协议
- 邮政银行笔试题目及答案
- 2025-2030年中国风电塔筒行业市场现状供需分析及投资评估规划分析研究报告
- 保底收益投资合同协议书
- AI技术在中小学心理健康教学中的实践与探索
- 《2025年普通高校在陕招生计划》
- 水手英语考试试题及答案
- 基于大语言模型的事件常识知识图谱扩展方法
- JJF2096-2024软包装件密封性试验仪校准规范
评论
0/150
提交评论