2024先進人工智能安全性評估國際科學(xué)報告中期報告+（英文版）-132正式版

InternationalScienti?cReportontheSafetyofAdvancedAIINTERIMREPORTMay2024ContributorsCHAIRProf.YoshuaBengio,UniversitédeMontréal/Mila-QuebecAIInstituteEXPERTADVISORYPANELProf.BronwynFox,TheCommonwealthScienti?candIndustrialResearchOrganisation(CSIRO)(Australia)Prof.HaroonSheikh,Netherlands’Scienti?cCouncilforGovernmentPolicy(Netherlands)Dr.GillJolly,MinistryofBusiness,InnovationAndréCarlosPoncedeLeonFerreiradeCarvalho,InstituteofMathematicsandComputerSciences,UniversityofS?oPaulo(Brazil)andEmployment(NewZealand)Dr.OlubunmiAjala,InnovationandDigitalEconomy(Nigeria)DominicLigot,CirroLytix(Philippines)Dr.MonaNemer,ChiefScienceAdvisorofCanada(Canada)Prof.KyoungMuLee,DepartmentofElectricalandComputerEngineering,SeoulNationalUniversity(RepublicofKorea)RaquelPezoaRivera,FedericoSantaMaríaTechnicalUniversity(Chile)AhmetHalitHatip,TurkishMinistryofIndustryandTechnology(RepublicofTurkey)Dr.YiZeng,InstituteofAutomation,ChineseAcademyofSciences(China)CrystalRugege,NationalCenterforAIandInnovationPolicy(Rwanda)JuhaHeikkil?,DGConnect(EuropeanUnion)GuillaumeAvrin,GeneralDirectorateofEnterprises(France)Dr.FahadAlbalawi,SaudiAuthorityforDataandArti?cialIntelligence(KingdomofSaudiArabia)Prof.AntonioKrüger,GermanResearchCenterforArti?cialIntelligence(Germany)DeniseWong,DataInnovationandProtectionGroup,InfocommMediaDevelopmentAuthority(IMDA)(Singapore)Prof.BalaramanRavindran,IndianInstituteofTechnology,Madras(India)Prof.HammamRiza,KORIKA(Indonesia)Dr.NuriaOliver,ELLISAlicante(Spain)Dr.CiaránSeoighe,ScienceFoundationIreland(Ireland)Dr.ChristianBusch,FederalDepartmentofEconomicAffairs,EducationandResearch(Switzerland)Dr.ZivKatzir,IsraelInnovationAuthority(Israel)OleksiiMolchanovskyi,ExpertCommitteeontheDevelopmentofArti?cialintelligenceinUkraine(Ukraine)Dr.AndreaMonti,UniversityofChieti-Pescara(Italy)MarwanAlserkal,MinistryofCabinetAffairs,PrimeMinister’sOf?ce(UnitedArabEmirates)Dr.HiroakiKitano,SonyGroup(Japan)[Interim]MaryKerema,MinistryofInformationCommunicationsTechnologyandDigitalEconomy(Kenya)SaifM.Khan,U.S.DepartmentofCommerce(UnitedStates)DameAngelaMcLean,GovernmentChiefScienti?cAdviser(UnitedKingdom)Dr.JoséRamónLópezPortillo,QElement(Mexico)AmandeepGill,UNTechEnvoy(UnitedNations)SCIENTIFICLEADS?renMindermann,Mila-QuebecAIInstituteWRITINGGROUPDanielPrivitera(leadwriter),KIRACenterTamayBesiroglu,EpochAIShayneLongpre,MassachusettsInstituteofTechnologyRishiBommasani,StanfordUniversityStephenCasper,MassachusettsInstituteofTechnologyVasiliosMavroudis,AlanTuringInstituteMantasMazeika,UniversityofIllinoisatUrbana-ChampaignYejinChoi,UniversityofWashington/A12DanielleGoldfarb,Mila-QuebecAIInstituteHodaHeidari,CarnegieMellonUniversityLeilaKhalatbari,HongKongUniversityofScienceandTechnologyKwanYeeNg,ConcordiaAIChinasaT.Okolo,Ph.D,TheBrookingsInstitutionDeborahRaji,MozillaTheodoraSkeadas,HumaneIntelligenceFlorianTramèr,ETHZürichSENIORADVISERSJohnA.McDermidOBEFREng,UniversityofYorkBayoAdekanmbi,DataScienceNigeriaPaulChristiano,contributedasaSeniorAdviserpriortotakinguphisroleattheUSAISafetyInstituteDavidDalrymple,AdvancedResearch+InventionAgency(ARIA)ThomasG.Dietterich,OregonStateUniversityEdwardFelten,PrincetonUniversityPascaleFung,HongKongUniversityofScienceandTechnology,contributedasaSeniorAdviserpriortotakingupherroleatMetaPierre-OlivierGourinchas,InternationalMonetaryFund(IMF)ArvindNarayanan,PrincetonUniversityAlondraNelson,InstituteforAdvancedStudyAliceOh,KAISTSchoolofComputingGopalRamchurn,RAIUK/UKRITASHub/UniversityofSouthamptonStuartRussell,UniversityofCalifornia,BerkeleyMarietjeSchaake,StanfordUniversityDawnSong,UniversityofCalifornia,BerkeleyAlvaroSoto,Ponti?ciaUniversidadCatólicadeChileNickJenningsCBFREngFRS,UniversityofLoughboroughAndreasKrause,ETHZurichPercyLiang,StanfordUniversityTeresaLudermir,FederalUniversityofPernambucoVidushiMarda,REALMLHelenMargettsOBEFBA,UniversityofOxford/AlanTuringInstituteLeeTiedrich,DukeUniversityGa?lVaroquaux,TheNationalInstituteforResearchinDigitalScienceandTechnology(Inria)AndrewYao,InstituteforInterdisciplinaryInformationSciences,TsinghuaUniversityYa-QinZhang,TsinghuaUniversitySECRETARIATUKGovernmentSecretariathostedbytheAISafetyInstituteBenjaminPrud’homme,Mila-QuebecAIInstituteACKNOWLEDGEMENTSTheSecretariatappreciatethehelpfulsupport,comments,andfeedbackfromthefollowingUK-basedorganisations:AdaLovelaceInstitute,TheAlanTuringInstitute,TheCentreforLong-TermResilience,CentrefortheGovernanceofAI,andUKAISafetyInstitute.AlsoaspecialthankstoDanHendrycks,DylanHad?eld-Menell,andPamelaSamuelson.?Crowncopyright2024ThispublicationislicensedunderthetermsoftheOpenGovernmentLicencev3.0exceptwhereotherwisestated.Toviewthislicence,visit.uk/doc/open-government-licence/version/3orwritetotheInformationPolicyTeam,TheNationalArchives,Kew,LondonTW94DU,oremail:psi@.ukWherewehaveidenti?edanythird-partycopyrightinformationyouwillneedtoobtainpermissionfromthecopyrightholdersconcerned.Anyenquiriesregardingthispublicationshouldbesenttousat:secretariat.AIStateofScience@.ukDisclaimerThereportdoesnotrepresenttheviewsoftheChair,anyparticularindividualinthewritingoradvisorygroups,noranyofthegovernmentsthathavesupporteditsdevelopment.ThisreportisasynthesisoftheexistingresearchonthecapabilitiesandrisksofadvancedAI.TheChairofthereporthasultimateresponsibilityforit,andhasoverseenitsdevelopmentfrombeginningtoend.Researchseriesnumber:DSIT2024/009ForewordsThisreportisthebeginningofajourneyonAISafetyIamhonouredtobechairingthedeliveryoftheinauguralInternationalScienti?cReportonAdvancedAISafety.IamproudtopublishthisinterimreportwhichistheculminationofhugeeffortsbymanyexpertsoverthesixmonthssincetheworkwascommissionedattheBletchleyParkAISafetySummitinNovember2023.WeknowthatadvancedAIisdevelopingveryrapidly,andthatthereisconsiderableuncertaintyoverhowtheseadvancedAIsystemsmightaffecthowweliveandworkinthefuture.AIhastremendouspotentialtochangeourlivesforthebetter,butitalsoposesrisksofharm.Thatiswhyhavingthisthoroughanalysisoftheavailablescienti?cliteratureandexpertopinionisessential.Themoreweknow,thebetterequippedwearetoshapeourcollectivedestiny.Ourmissionisclear:todriveashared,science-based,up-to-dateunderstandingofthesafetyofadvancedAI,andtocontinuetodevelopthatunderstandingovertime.ThereportrightlyhighlightsthatthereareareasofconsensusamongexpertsandalsodisagreementsoverthecapabilitiesandrisksofadvancedAI,especiallythoseexpectedtobedevelopedinthefuture.Inordertomeetourmissioneffectively,wehaveaimedtoaddressdisagreementamongsttheexpertcommunitywithintellectualhonesty.Bydissectingthesedifferences,wepavethewayforinformedpolicy-makingandstimulatetheresearchneededtohelpclearthefogandmitigaterisks.IamgratefultoourinternationalExpertAdvisoryPanelfortheirinvaluablecomments,initiallyshapingthereport’sscopeandlaterprovidingfeedbackonthefulldraft.Theirdiverseperspectivesandcarefulreviewhavebroadenedandstrengthenedthisinterimreport.Equallydeservingofrecognitionaremydedicatedteamofwritersandsenioradvisers.Theircommitmentoverthepastfewmonthshascreatedaninterimproductthathassurpassedmyexpectations.MythanksalsogototheUKGovernmentforstartingthisprocessandofferingoutstandingoperationalsupport.ItwasalsoimportantformethattheUKGovernmentagreedthatthescientistswritingthisreportshouldhavecompleteindependence.Thisinterimreportisonlythebeginningofajourney.Therearenodoubtperspectivesandevidencethatthisreporthasfailedtocaptureinthis?rstattempt.Inascienti?cprocesssuchasthis,feedbackisprecious.Wewillincorporateadditionalevidenceandscienti?cviewpointsasweworktowardthe?nalversion.ProfessorYoshuaBengioUniversitédeMontréal/Mila-QuebecAIInstitute&Chair7InternationalScienti?cReportontheSafetyofAdvancedAI:InterimReportAISafetyisasharedglobalissueIamdelightedtopresentthisinterimupdateonthe?rstInternationalScienti?cReportontheSafetyofAdvancedAI,akeyoutcomeofthegroundbreakingAISafetySummitheldatBletchleyParkinNovember2023.Thislandmarkreportrepresentsanunprecedentedglobalefforttobuildashared,science-basedunderstandingoftheopportunitiesandrisksposedbyrapidadvancementsinAI,andisatestamenttothe"BletchleyEffect"-thepowerofconveningbrilliantmindstotackleoneofhumanity'sgreatestchallenges.WebelievethatrealisingtheimmensepotentialofAItobene?thumanitywillrequireproactiveeffortstoensurethesepowerfultechnologiesaredevelopedanddeployedsafelyandresponsibly.Noonecountrycantacklethischallengealone.ThatiswhyIwassopassionateaboutbringingtogetheradiversegroupofworld-leadingexpertstocontributetheirknowledgeandperspectives.IwanttoespeciallythankProfessorYoshuaBengioforhisleadershipasChairinskilfullyshepherdingthiscomplexinternationaleffort.Cruciallythereportalsoshinesalightonthesigni?cantgapsinourcurrentknowledgeandthekeyuncertaintiesanddebatesthaturgentlyrequirefurtherresearchanddiscussion.Itismysincerehopethatthisreport,andthecooperativeprocessbehindit,canserveasacatalystfortheresearchandpolicyeffortsneededtoclosecriticalknowledgegapsandavaluableinputforthechallengingpolicychoicesthatlieahead.Westillhavemuchtolearn,butthisreportmarksanimportantstart.TheUKlooksforwardtocontinuingtoworkwithinternationalpartnerstopromotearesponsible,human-centricapproachtoAIdevelopment-onethatharnessesthesepowerfultoolstoimprovelivesandlivelihoodswhilevigilantlysafeguardingagainstdownsiderisksandharms.Together,wecanworktobuildafutureinwhichallofhumanitycanbene?tfromthewondersofAI.TheRtHonMichelleDonelanMP,SecretaryofState,DepartmentforScience,Innovation,andTechnologyAcriticalstepforwardandaCalltoActiononAISafetyTherapidadvancementofAIstandspoisedtoreshapeourworldinwaysbothprofoundandunforeseen.Fromrevolutionisinghealthcareandtransportationtoautomatingcomplextasksandunlockingscienti?cbreakthroughs,AI'spotentialforpositiveimpactisundeniable.However,alongsidethesenotablepossibilitiesliesigni?cantchallengesthatnecessitateaforward-lookingapproach.Concernsrangefromunintendedbiasesembeddedinalgorithmstothepossibilityofautonomoussystemsexceedinghumancontrol.Thesepotentialriskshighlighttheurgentneedforaglobalconversationtoensurethesafe,andresponsibleadvancementofAI.Inthiscontext,theInternationalAISafetyReportwillprovidevitalgroundworkforglobalcollaboration.Thereportrepresentsaconvergenceofknowledgefromexpertsacross30countries,theEuropeanUnion,andtheUnitedNations,providingacomprehensiveanalysisofAIsafety.Byfocusingontheearlyscienti?cunderstandingofcapabilitiesandrisksfromgeneralpurposeAIandevaluatingtechnicalmethodsforassessingandmitigatingthem,thereportwillsparkongoingdialogueandcollaborationamongmulti-stakeholders.Ihopethatbasedonthisreport,expertsfrom30countries,theEU,andtheUNcontinuetoengageinbalanceddiscussions,achievingAIriskmitigationthatisacceptableandtailoredtothespeci?ccontextofbothdevelopedanddevelopingcountries,therebycreatingafuturewhereinnovationandresponsibleAIcoexistharmoniously.LeeJong-Ho,MinisterofMSIT,RepublicofKorea8ExecutiveSummaryAboutthisreport?Thisistheinterimpublicationofthe?rst‘InternationalScienti?cReportontheSafetyofAdvancedAI’.Adiversegroupof75arti?cialintelligence(AI)expertscontributedtothisreport,includinganinternationalExpertAdvisoryPanelnominatedby30countries,theEuropeanUnion(EU),andtheUnitedNations(UN).??LedbytheChairofthisreport,theindependentexpertswritingthisreportcollectivelyhadfulldiscretionoveritscontent.AtatimeofunprecedentedprogressinAIdevelopment,this?rstpublicationrestrictsitsfocustoatypeofAIthathasadvancedparticularlyrapidlyinrecentyears:General-purposeAI,orAIthatcanperformawidevarietyoftasks.Amidrapidadvancements,researchongeneral-purposeAIiscurrentlyinatimeofscienti?cdiscoveryandisnotyetsettledscience.??Peoplearoundtheworldwillonlybeabletoenjoygeneral-purposeAI’smanypotentialbene?tssafelyifitsrisksareappropriatelymanaged.Thisreportfocusesonidentifyingtheserisksandevaluatingtechnicalmethodsforassessingandmitigatingthem.Itdoesnotaimtocomprehensivelyassessallpossiblesocietalimpactsofgeneral-purposeAI,includingitsmanypotentialbene?ts.Forthe?rsttimeinhistory,thisinterimreportbroughttogetherexpertsnominatedby30countries,theEU,andtheUN,andotherworld-leadingexperts,toprovideasharedscienti?c,evidence-basedfoundationfordiscussionsanddecisionsaboutgeneral-purposeAIsafety.Wecontinuetodisagreeonseveralquestions,minorandmajor,aroundgeneral-purposeAIcapabilities,risks,andriskmitigations.Butweconsiderthisprojectessentialforimprovingourcollectiveunderstandingofthistechnologyanditspotentialrisks,andformovingclosertowardsconsensusandeffectiveriskmitigationtoensurepeoplecanexperiencethepotentialbene?tsofgeneral-purposeAIsafely.Thestakesarehigh.Welookforwardtocontinuingthiseffort.Highlightsoftheexecutivesummary?Ifproperlygoverned,general-purposeAIcanbeappliedtoadvancethepublicinterest,potentiallyleadingtoenhancedwellbeing,moreprosperity,andnewscienti?cdiscoveries.However,malfunctioningormaliciouslyusedgeneral-purposeAIcanalsocauseharm,forinstancethroughbiaseddecisionsinhigh-stakessettingsorthroughscams,fakemedia,orprivacyviolations.?Asgeneral-purposeAIcapabilitiescontinuetoadvance,riskssuchaslarge-scalelabourmarketimpacts,AI-enabledhackingorbiologicalattacks,andsocietylosingcontrolovergeneral-purposeAIcouldemerge,althoughthelikelihoodofthesescenariosisdebatedamongresearchers.Differentviewsontheserisksoftenstemfromdifferingexpectationsaboutthestepssocietywilltaketolimitthem,theeffectivenessofthosesteps,andhowrapidlygeneral-purposeAIcapabilitieswillbeadvanced.??Thereisconsiderableuncertaintyabouttherateoffutureprogressingeneral-purposeAIcapabilities.Someexpertsthinkaslowdownofprogressisbyfarmostlikely,whileotherexpertsthinkthatextremelyrapidprogressispossibleorlikely.Therearevarioustechnicalmethodstoassessandreducerisksfromgeneral-purposeAIthatdeveloperscanemployandregulatorscanrequire,buttheyallhavelimitations.Forexample,currenttechniquesforexplainingwhygeneral-purposeAImodelsproduceanygivenoutputareseverelylimited.9InternationalScienti?cReportontheSafetyofAdvancedAI:InterimReport?Thefutureofgeneral-purposeAItechnologyisuncertain,withawiderangeoftrajectoriesappearingpossibleeveninthenearfuture,includingbothverypositiveandverynegativeoutcomes.ButnothingaboutthefutureofAIisinevitable.ItwillbethedecisionsofsocietiesandgovernmentsthatwilldeterminethefutureofAI.Thisinterimreportaimstofacilitateconstructivediscussionaboutthesedecisions.Thisreportsynthesisesthestateofscienti?cunderstandingofgeneral-purposeAI–AIthatcanperformawidevarietyoftasks–withafocusonunderstandingandmanagingitsrisksThecapabilitiesofsystemsusingAIhavebeenadvancingrapidly.ThishashighlightedthemanyopportunitiesthatAIcreatesforbusiness,research,government,andprivatelife.IthasalsoledtoanincreasedawarenessofcurrentharmsandpotentialfuturerisksassociatedwithadvancedAI.ThepurposeoftheInternationalScienti?cReportontheSafetyofAdvancedAIistotakeasteptowardsasharedinternationalunderstandingofAIrisksandhowtheycanbemitigated.This?rstinterimpublicationofthereportrestrictsitsfocustoatypeofAIwhosecapabilitieshaveadvancedparticularlyrapidly:general-purposeAI,orAIthatcanperformawidevarietyoftasks.Amidrapidadvancements,researchongeneral-purposeAIiscurrentlyinatimeofscienti?cdiscoveryandisnotyetsettledscience.Thereportprovidesasnapshotofthecurrentscienti?cunderstandingofgeneral-purposeAIanditsrisks.Thisincludesidentifyingareasofscienti?cconsensusandareaswheretherearedifferentviewsoropenresearchquestions.Peoplearoundtheworldwillonlybeabletoenjoythepotentialbene?tsofgeneral-purposeAIsafelyifitsrisksareappropriatelymanaged.Thisreportfocusesonidentifyingrisksfromgeneral-purposeAIandevaluatingtechnicalmethodsforassessingandmitigatingthem,includingthebene?cialuseofgeneral-purposeAItomitigaterisks.Itdoesnotaimtocomprehensivelyassessallpossiblesocietalimpactsofgeneral-purposeAI,includingwhatbene?tsitmayoffer.General-purposeAIcapabilitieshavegrownrapidlyinrecentyearsaccordingtomanymetrics,andthereisnoconsensusonhowtopredictfutureprogress,makingawiderangeofscenariosappearpossibleAccordingtomanymetrics,general-purposeAIcapabilitiesareprogressingrapidly.Fiveyearsago,theleadinggeneral-purposeAIlanguagemodelscouldrarelyproduceacoherentparagraphoftext.Today,somegeneral-purposeAImodelscanengageinmulti-turnconversationsonawiderangeoftopics,writeshortcomputerprograms,orgeneratevideosfromadescription.However,thecapabilitiesofgeneral-purposeAIaredif?culttoestimatereliablyandde?neprecisely.Thepaceofgeneral-purposeAIadvancementdependsonboththerateoftechnologicaladvancementsandtheregulatoryenvironment.Thisreportfocusesonthetechnologicalaspectsanddoesnotprovideadiscussionofhowregulatoryeffortsmightaffectthespeedofdevelopmentanddeploymentofgeneral-purposeAI.AIdevelopershaverapidlyadvancedgeneral-purposeAIcapabilitiesinrecentyearsmostlybycontinuouslyincreasingresourcesusedfortrainingnewmodels(atrendcalled‘scaling’)andre?ningexistingalgorithms.Forexample,state-of-the-artAImodelshaveseenannualincreasesofapproximately4xincomputationalresources(‘compute’)usedfortraining,2.5xintrainingdatasetsize,and1.5-3xinalgorithmicef?ciency(performancerelativetocompute).Whether‘scaling’hasresultedinprogressonfundamentalchallengessuchascausalreasoningisdebatedamongresearchers.10本報告來源于三個皮匠報告站（）,由用戶Id:673421下載,文檔Id:464666,下載日期:2025-01-24InternationalScienti?cReportontheSafetyofAdvancedAI:InterimReportThepaceoffutureprogressingeneral-purposeAIcapabilitieshassubstantialimplicationsformanagingemergingrisks,butexpertsdisagreeonwhattoexpecteveninthenearfuture.Expertsvariouslysupportthepossibilityofgeneral-purposeAIcapabilitiesadvancingslowly,rapidly,orextremelyrapidly.Thisdisagreementinvolvesakeyquestion:willcontinued‘scaling’ofresourcesandre?ningexistingtechniquesbesuf?cienttoyieldrapidprogressandsolveissuessuchasreliabilityandfactualaccuracy,orarenewresearchbreakthroughsrequiredtosubstantiallyadvancegeneral-purposeAIabilities?Severalleadingcompaniesthatdevelopgeneral-purposeAIarebettingon‘scaling’tocontinueleadingtoperformanceimprovements.Ifrecenttrendscontinue,bytheendof2026somegeneral-purposeAImodelswillbetrainedusing40xto100xmorecomputethanthemostcompute-intensivemodelspublishedin2023,combinedwithtrainingmethodsthatusethiscompute3xto20xmoreef?ciently.However,therearepotentialbottleneckstofurtherincreasingbothdataandcompute,includingtheavailabilityofdata,AIchips,capitalexpenditure,andlocalenergycapacity.Companiesdevelopinggeneral-purposeAIareworkingtonavigatethesepotentialbottlenecks.Severalresearcheffortsaimtounderstandandevaluategeneral-purposeAImorereliably,butouroverallunderstandingofhowgeneral-purposeAImodelsandsystemsworkislimitedApproachestomanagingrisksfromgeneral-purposeAIoftenrestontheassumptionthatAIdevelopersandpolicymakerscanassessthecapabilitiesandpotentialimpactsofgeneral-purposeAImodelsandsystems.Butwhiletechnicalmethodscanhelpwithassessment,allexistingmethodshavelimitationsandcannotprovidestrongassurancesagainstmostharmsrelatedtogeneral-purposeAI.Overall,thescienti?cunderstandingoftheinnerworkings,capabilities,andsocietalimpactsofgeneral-purposeAIisverylimited,andthereisbroadexpertagreementthatitshouldbeaprioritytoimproveourunderstandingofgeneral-purposeAI.Someofthekeychallengesinclude:?Developersstillunderstandlittleabouthowtheirgeneral-purposeAImodelsoperate.Thisisbecausegeneral-purposeAImodelsarenotprogrammedinthetraditionalsense.Instead,theyaretrained:AIdeveloperssetupatrainingprocessthatinvolvesalotofdata,andtheoutcomeofthattrainingprocessisthegeneral-purposeAImodel.Thesemodelscanconsistoftrillionsofcomponents,calledparameters,andmostoftheirinnerworkingsareinscrutable,includingtothemodeldevelopers.Modelexplanationandinterpretabilitytechniquescanimproveresearchers’anddevelopers’understandingofhowgeneral-purposeAImodelsoperate,butthisresearchisnascent.?General-purposeAIismainlyassessedthroughtestingthemodelorsystemonvariousinputs.Thesespotchecksarehelpfulforassessingstrengthsandweaknesses,includingvulnerabilitiesandpotentiallyharmfulcapabilities,butdonotprovidequantitativesafetyguarantees.Thetestsoftenmisshazardsandoverestimateorunderestimatecapabilitiesbecausegeneral-purposeAIsystemsmaybehavedifferentlyindifferentcircumstances,withdifferentusers,orwithadditionaladjustmentstotheircomponents.??Independentactorscan,inprinciple,auditgeneral-purposeAImodelsorsystemsdevelopedbyacompany.However,companiesoftendonotprovideindependentauditorswiththenecessarylevelofdirectaccesstomodelsortheinformationaboutdataandmethodsusedthatareneededforrigorousassessment.Severalgovernmentsarebeginningtobuildcapacityforconductingtechnicalevaluationsandaudits.Itisdif?culttoassessthedownstreamsocietalimpactofageneral-purposeAIsystembecauseresearchintoriskassessmenthasnotbeensuf?cienttoproducerigorousandcomprehensiveassessmentmethodologies.Inaddition,general-purposeAIhasawiderangeofusecases,whichareoftennotprede?nedandonlylightlyrestricted,complicatingriskassessmentfurther.Understandingthepotentialdownstreamsocietalimpactsofgeneral-purposeAImodelsandsystemsrequiresnuancedandmultidisciplinaryanalysis.Increasingtherepresentationofdiverse11InternationalScienti?cReportontheSafetyofAdvancedAI:InterimReportperspectivesingeneral-purposeAIdevelopmentandevaluationprocessesisanongoingtechnicalandinstitutionalchallenge.General-purposeAIcanposesevereriskstoindividualandpublicsafetyandwellbeingThisreportclassi?esgeneral-purposeAIrisksintothreecategories:malicioususerisks,risksfrommalfunctions,andsystemicrisks.Italsodiscussesseveralcross-cuttingfactorsthatcontributetomanyrisks.Malicioususe.Likeallpowerfultechnologies,general-purposeAIsystemscanbeusedmaliciouslytocauseharm.Possibletypesofmalicioususerangefromrelativelywell-evidencedones,suchasscamsenabledbygeneral-purposeAI,toonesthatsomeexpertsbelievemightoccurinthecomingyears,suchasmalicioususeofscienti?ccapabilitiesofgeneral-purposeAI.?Harmtoindividualsthroughfakecontentgeneratedbygeneral-purposeAIisarelativelywell-documentedclassofgeneral-purposeAImalicioususe.General-purposeAIcanbeusedtoincreasethescaleandsophisticationofscamsandfraud,forexamplethrough‘phishing’attacksenhancedbygeneral-purposeAI.General-purposeAIcanalsobeusedtogeneratefakecompromisingcontentfeaturingindividualswithouttheirconsent,suchasnon-consensualdeepfakepornography.??Anotherareaofconcernisthemalicioususeofgeneral-purposeAIfordisinformationandmanipulationofpublicopinion.General-purposeAIandothermoderntechnologiesmakeiteasiertogenerateanddisseminatedisinformation,includinginanefforttoaffectpoliticalprocesses.Technicalcountermeasureslikewatermarkingcontent,althoughuseful,canusuallybecircumventedbymoderatelysophisticatedactors.General-purposeAImightalsobemaliciouslyusedforcyberoffence,upliftingthecyberexpertiseofindividualsandmakingiteasierformalicioususerstoconducteffectivecyber-attacks.General-purposeAIsystemscanbeusedtoscaleandpartiallyautomatesometypesofcyberoperations,suchassocialengineeringattacks.However,general-purposeAIcouldalsobeusedincyberdefence.Overall,thereisnotyetanysubstantialevidencesuggestingthatgeneral-purposeAIcanautomatesophisticatedcybersecuritytasks.?Someexpertshavealsoexpressedconcernthatgeneral-purposeAIcouldbeusedtosupportthedevelopmentandmalicioususeofweapons,suchasbiologicalweapons.Thereisnostrongevidencethatcurrentgeneral-purposeAIsystemsposethisrisk.Forexample,althoughcurrentgeneral-purposeAIsystemsdemonst