生成式人工智能教育與研究應(yīng)用指南(英)

GuidanceforgenerativeAIGuidanceforgenerativeAIineducationandresearch20302030UNESCO–agloballeaderineducationEducationisUNESCO’stopprioritybecauseitisabasichumanrightandthefoundationforpeaceandsustainabledevelopment.UNESCOistheUnitedNations’specializedagencyforeducation,providingglobalandregionalleadershiptodriveprogress,strengtheningtheresilienceandcapacityofnationalsystemstoservealllearners.UNESCOalsoleadseortstorespondtocontemporaryglobalchallengesthroughtransformativelearning,withspecialfocusongenderequalityandAfricaacrossallactions.TheGlobalEducation2030AgendaUNESCO,astheUnitedNations’specializedagencyforeducation,isentrustedtoleadandcoordinatetheEducation2030Agenda,whichispartofaglobalmovementtoeradicatepovertythrough17SustainableDevelopmentGoalsby2030.Education,essentialtoachieveallofthesegoals,hasitsowndedicatedGoal4,whichaimsto“ensureinclusiveandequitablequalityeducationandpromotelifelonglearningopportunitiesforall.”TheEducation2030FrameworkforActionprovidesguidancefortheimplementationofthisambitiousgoalandcommitments.20302030PubliclyavailablegenerativeAI(GenAI)toolsarerapidlyemerging,andthereleaseofiterativeversionsisoutpacingtheadaptationofnationalregulatoryframeworks.TheabsenceofnationalregulationsonGenAIinmostcountriesleavesthedataprivacyofusersunprotectedandeducationalinstitutionslargelyunpreparedtovalidatethetools.UNESCO’sfirstglobalguidanceonGenAIineducationaimstosupportcountriestoimplementimmediateactions,planlong-termpoliciesanddevelophumancapacitytoensureahuman-centredvisionofthesenewtechnologies.TheGuidancepresentsanassessmentofpotentialrisksGenAIcouldposetocorehumanisticvaluesthatpromotehumanagency,inclusion,equity,genderequality,andlinguisticandculturaldiversities,aswellaspluralopinionsandexpressions.ItproposeskeystepsforgovernmentalagenciestoregulatetheuseofGenAItoolsincludingmandatingtheWhileChatGPTWhileChatGPTagelimitfortheiruse.ItoutlinesreforGenAIproviderstoenabletheirethicalandeffectiveuseineducation.onecountryhadreleasedonecountryhadreleasedongenerativeAIinstitutionstovalidateGenAIsystemsontheirethicalandpedagogicalappropriatenessforeducation.Itcallsontheinternationalcommunitytoreflectontheirlong-termimplicationsforknowledge,teaching,learningandassessment.Thepublicationoffersconcreterecommendationsforpolicy-makersandeducationalinstitutionsonhowtheusesofGenAItoolscanbedesignedtoprotecthumanagencyandgenuinelybenefitlearners,teachersandresearchers.Generativeartificialintelligence(GenAI)burstintothepublicawarenessinlate2022withthelaunchofChatGPT,whichbecamethefastestgrowingappinhistory.Withthepowertoimitatehumancapabilitiestoproduceoutputssuchastext,images,videos,musicandsoftwarecodes,theseGenAIapplicationshavecausedastir.MillionsofpeoplearenowusingGenAIintheirdailylivesandthepotentialofadaptingthemodelstodomain-specificAIapplicationsseemsunlimited.Suchwide-rangingcapacitiesforinformationprocessingandknowledgeproductionhavepotentiallyhugeimplicationsforeducation,astheyreplicatethehigher-orderthinkingthatconstitutesthefoundationofhumanlearning.AsGenAItoolsareincreasinglyabletoautomatesomebasiclevelsofwritingandartworkcreation,theyareforcingeducationpolicy-makersandinstitutionstorevisitwhy,whatandhowwelearn.Thesearenowcriticalconsiderationsforeducationinthisnewphaseofthedigitalera.Thispublicationaimstosupporttheplanningofappropriateregulations,policiesandhumancapacitydevelopment,toensurethatGenAIbecomesatoolthatgenuinelybenefitsandempowersteachers,learnersandresearchers.ItproposeskeystepsforgovernmentalagenciestoregulatetheuseofgenerativeAI.Italsopresentsframeworksandconcreteexamplesforpolicyformulationandinstructionaldesignthatenableethicalandeffectiveusesofthistechnologyineducation.Finally,itcallsontheinternationalcommunitytoconsidertheprofoundlonger-termimplicationsofgenerativeAIforhowweunderstandknowledgeanddefinelearningcontent,methodsandoutcomes,aswellasthewayinwhichweassessandvalidatelearning.BuildingonUNESCO’s2021RecommendationontheEthicsofArtificialIntelligence,theguidanceisanchoredinahumanisticapproachtoeducationthatpromoteshumanagency,inclusion,equity,genderequality,andculturalandlinguisticdiversity,aswellaspluralopinionsandexpressions.Furthermore,itrespondstothecallofthe2021reportoftheInternationalCommissionontheFuturesofEducation,Reimaginingourfuturestogether:contractforeducationtoredefineourrelationshipwithtechnology,asanintegralpartofoureffortstorenewthesocialcontractforeducation.AImustnotusurphumanintelligence.Rather,itinvitesustoreconsiderourestablishedunderstandingsofknowledgeandhumanlearning.Itismyhopethatthisguidancewillhelpusredefinenewhorizonsforeducationandinformourcollectivethinkingandcollaborativeactionsthatcanleadtohuman-centreddigitallearningfuturesAcknowledgementsAcknowledgementsUndertheleadershipofStefaniaGiannini,Assistant-DirectorforEducation,andtheguidanceofSobhiTawil,DirectoroftheFutureofLearningandInnovationDivisionatUNESCO,thedraftingofthepublicationwasledbyFengchunMiao,ChiefofUnitforTechnologyandAIinEducation.ParticularthanksgotoWayneHolmes,AssociateProfessoratUniversityCollegeLondon,whoco-draftedthepublication.ThispublicationisthefruitofacollectiveeffortofeducationleadersandexpertsinthefieldofAIandeducation.Itbenefitedfromtheinsightsandinputsofmanyexpertsincluding:MutluCukurova,ProfessoratUniversityCollegeLondon;ColindelaHiguera,UNESCOChairinTechnologiesfortheTrainingofTeacherswithOpenEducationalResourcesatNantesUniversity;ShafikaIsaacs,ResearchAssociateattheUniversityofJohannesburg;NatalieLao,ExecutiveDirectoroftheAppInventorFoundation;QinNi,AssociateProfessoratShanghaiNormalUniversity;CatalinaNicolin,ICTinEducationExpertattheEuropeanDigitalEducationHubinRomania;JohnShaw-Taylor,UNESCOChairinAIandProfessorofComputationalStatisticsandMachineLearningatUniversityCollegeLondon;KellyShirohira,ExecutiveManageratJetEducationServices;Ki-SangSong,ProfessoratKoreaNationalUniversityofEducation;andIlkkaTuomi,ChiefScientistatMeaningProcessingLtdinFinland.ManycolleaguesacrossUNESCOalsocontributedinvariouswaysincluding:DafnaFeinholz,ChiefofSectionforBioethicsandtheEthicsofScienceandTechnology;FrancescPedró,DirectoroftheInternationalInstituteforHigherEducationinLatinAmericaandtheCaribbean;PrateekSibal,ProgrammeSpecialist,SectionforDigitalPoliciesandDigitalTransformation;SaurabhRoy,SeniorProjectOfficerattheSectionforTeacherDevelopment,DivisionforPoliciesandLifelongLearningSystems;BenjaminVergelDeDios,ProgrammeSpecialistinICTinEducation,SectionforEducationalInnovationandSkillsDevelopmentintheBangkokOffice;thecolleaguesintheDiversityofCulturalExpressionsEntityintheCultureSector;andMarkWest,ProgrammeSpecialist,FutureofLearningandInnovationDivision.AppreciationisalsoduetoGlenHertelendy,LuisaFerraraandXiangleiZheng,UnitforTechnologyandAIinEducation,FutureofLearningandInnovation,forcoordinatingtheproductionofthepublication.GratitudeisalsoextendedtoJennyWebsterforcopy-editingandproofreadingthetext,andtoNgoc-ThuyTranfordesigningthelayout.4TableTableofcontentsForeword 2Acknowledgements 3Listofacronymsandabbreviations 6Introduction 71.WhatisgenerativeAIandhowdoesitwork? 8 81.2HowdoesgenerativeA 81.2.1HowtextGenAImodelswork 91.2.2HowimageGenAImodelswork 1.3Prompt-engineeringtogeneratedesiredoutputs 111.4EmergingEdGPTanditsimplications 132.ControversiesaroundgenerativeAIandtheirimplicationsforeducation 142.1Worseningdigitalpoverty 142.2Outpacingnationalregulatoryadaptation 142.3Useofcontentwithoutconsent 152.4Unexplainablemodelsusedtogenerateoutputs 152.5AI-generatedcontentpollutingtheinternet 162.6Lackofunderstandingoftherealworld 16 172.8Generatingdeeperdeepfakes 173.RegulatingtheuseofgenerativeAIineducation 183.1Ahuman-centredapproachtoAI 183.2StepstoregulateGenAIineducation 183.3RegulationsonGenAI:Keyelements 203.3.1Governmentalregulatoryagencies 203.3.2ProvidersofGenAI 21 23 234.TowardsapolicyframeworkfortheuseofgenerativeAIineducationandresearch 244.1Promoteinclusion,equity,andlinguisticandculturaldiversity 244.2Protecthumanagency 2454.3MonitorandvalidateGenAIsystemsforeducation 254.4DevelopAIcompetenciesincluding 254.5BuildcapacityforteachersandresearcherstomakeproperuseofGenAI 264.6Promotepluralopinionsandpluralexpressionsofideas 264.7Testlocallyrelevantapplicationmodelsandbuildacumulativeevidencebase 274.8Reviewlong-termimplicationsinanintersectoralandinterdisciplinarymanner 275.FacilitatingcreativeuseofGenAIineducationandresearch 285.1InstitutionalstrategiestofacilitateresponsibleandcreativeuseofGenAI 285.2A‘human-centredandpedagogicallyappropriateinteraction’approach 295.3Co-designingtheuseofGenAIineducationandresearch 29 295.3.2GenerativeAItofacilitateteaching 305.3.3GenerativeAIasa1:1coachfortheself-pacedacquisitionoffoundationalskills 315.3.4GenerativeAItofacilitateinquiryorproject-basedlearning 335.3.5GenerativeAItosupportlearnerswithspecialneeds 346.GenAIandthefutureofeducationandresearch 366.1Unchartedethicalissues 366.2Copyrightandintellectualproperty 366.3Sourcesofcontentandlearning 366.4Homogenizedresponsesversusdiverseandcreativeoutputs 376.5Rethinkingassessmentandlearningoutcomes 376.6Thinkingprocesses 37Concludingremarks 38References 39ListoftablesTable1.TechniquesusedingenerativeAI 8Table2.OpenAIGPTs 9Table3.Co-designingusesofGenAIforresearch 30Table4.Co-designingusesofGenAItosupportteachersandteaching 31Table5.Co-designingusesofGenAIasa1:1coachfortheself-pacedacquisitionoffoundationalskillsinlanguagesandthearts 32Table6.Co-designingusesofGenAItofacilitateinquiryorproject-basedlearning 33Table7.Co-designingusesofGenAItosupportlearnerswithspecialneeds 346ListListofacronymsandabbreviationsAGIArtificialgeneralintelligenceAIArtificialintelligenceAPIApplicationprogramminginterfaceANNArtificialneuralnetworkDAIDistributedartificialintelligenceGANGenerativeadversarialnetworksGBGigabytesGDPRGeneralDataProtectionRegulationGenAIGenerativeartificialintelligenceGPTGenerativepre-trainedtransformerICTInformationandcommunicationtechnologyLaMDALanguagemodelfordialogueapplicationsLLMLargelanguagemodelMLMachinelearningVAEVariationalautoencodersAGCCAIGovernmentCloudCluster(Singapore)CACCyberspaceAdministrationofChinaEUEuropeanUnionOECDOrganisationforEconomicCo-operationandDevelopmentUNCTADUnitedNationsConferenceonTradeandDevelopmentUNESCOUnitedNationsEducational,ScientificandCulturalOrganizationIntroductionIntroductionThereleaseofChatGPTinlate2022,thefirsteasy-to-usegenerativeartificialintelligence(GenAI)toolmadewidelyavailabletothepublic,1followedbyiterativelymoresophisticatedversions,sentshockwavesworldwide,andisfuellingtheraceamonglargetechnologycompaniestopositionthemselvesinthefieldofGenAImodeldevelopment.2Acrosstheworld,theinitialconcernineducationwasthatChatGPTandsimilarGenAItoolswouldbeusedbystudentstocheatontheirassignments,thusunderminingthevalueoflearningassessment,certificationandqualifications(Anders,2023).WhilesomeeducationalinstitutionsbannedtheuseofChatGPT,otherscautiouslywelcomedthearrivalofGenAI(Tlili,2023).Manyschoolsanduniversities,forinstance,adoptedaprogressiveapproachbelievingthat‘ratherthanseektoprohibittheiruse,studentsandstaffneedtobesupportedinusingGenAItoolseffectively,ethicallyandtransparently’(RussellGroup,2023).ThisapproachacknowledgesthatGenAIiswidelyavailable,islikelyonlytobecomemoresophisticated,andhasbothspecificnegativeanduniquepositivepotentialforeducation.Indeed,GenAIhasamyriadofpossibleuses.Itcanautomateinformationprocessingandthepresentationofoutputsacrossallkeysymbolicrepresentationsofhumanthinking.Itenablesthedeliveryoffinaloutputsbyfurnishingsemi-finishedknowledgeproducts.Byfreeinghumansfromsomecategoriesoflower-orderthinkingskills,thisnewgenerationofAItoolsmighthaveprofoundimplicationsforhowweunderstandhumanintelligenceandlearning.ButGenAIalsoraisesmultipleimmediateconcernsrelatedtoissuessuchassafety,dataprivacy,copyright,andmanipulation.SomeofthesearebroaderrisksrelatedtoartificialintelligencethathavebeenfurtherexacerbatedbyGenAI,whileothershavenewlyemergedwiththislatestgenerationoftools.Itisnowurgentthateachoftheseissuesandconcernsbefullyunderstoodandaddressed.ThisGuidanceisdesignedtorespondtothisurgentneed.However,athematicsetofguidanceonGenAIforeducationshouldnotbeunderstoodasaclaimthatGenAIisthesolutiontoeducation’sfundamentalchallenges.Despitethemediahyperbole,itisunlikelythatGenAIalonewillsolveanyoftheproblemsfacingeducationsystemsaroundtheworld.Inrespondingtolong-standingeducationalissues,itiskeytoupholdtheideathathumancapacityandcollectiveaction,andnottechnology,isthedeterminingfactorineffectivesolutionstofundamentalchallengesfacedbysocieties.ThisGuidancethereforeaimstosupporttheplanningofappropriateregulations,policiesandhumancapacitydevelopmentprogrammes,toensurethatGenAIbecomesatoolthatgenuinelybenefitsandempowersteachers,learnersandresearchers.BuildingonUNESCO’sRecommendatiofArtificialIntelligence,theGuidanceisanchoredinahuman-centredapproachthatpromoteshumanagency,inclusion,equity,genderequality,andculturalandlinguisticdiversity,aswellaspluralopinionsandexpressions.TheGuidancefirstlooksintowhatGenAIisandhowitworks,presentingthediversetechnologiesandmodelsavailable(Section1),beforeidentifyingarangeofcontroversialethicalandpolicyissuesaroundbothAIingeneral,andGenAIspecifically(Section2).ThisisfollowedbyadiscussionofthestepsandkeyelementstobeexaminedwhenseekingtoregulateGenAIbasedonahuman-centredapproach–onethatensuresethical,safe,equitableandmeaningfuluse(Section3).Section4thenproposesmeasuresthatcanbetakentodevelopcoherent,comprehensivepolicyframeworkstoregulatetheuseofGenAIineducationandresearch,whileSection5looksintothepossibilitiesforcreativelyusingGenAIincurriculumdesign,teaching,learningandresearchactivities.Section6concludestheGuidancewithconsiderationsaroundthelong-termimplicationsofGenAIforeducationandresearch.81.WhatisgenerativeAI1.1WhatisgenerativeAI?GenerativeAI(GenAI)isanartificialintelligence(AI)technologythatautomaticallygeneratescontentinresponsetopromptswritteninnatural-languageconversationalinterfaces.Ratherthansimplycuratingexistingwebpages,bydrawingonexistingcontent,GenAIactuallyproducesnewcontent.Thecontentcanappearinformatsthatcompriseallsymbolicrepresentationsofhumanthinking:textswritteninnaturallanguage,images(includingphotographs,digitalpaintingsandcartoons),videos,musicandsoftwarecode.GenAIistrainedusingdatacollectedfromwebpages,socialmediaconversationsandotheronlinemedia.Itgeneratesitscontentbystatisticallyanalysingthedistributionsofwords,pixelsorotherelementsinthedatathatithasingestedandidentifyingandrepeatingcommonpatterns(forexample,whichwordstypicallyfollowwhichotherwords).WhileGenAIcanproducenewcontent,itcannotgeneratenewideasorsolutionstoreal-worldchallenges,asitdoesnotunderstandreal-worldobjectsorsocialrelationsthatunderpinlanguage.Moreover,despiteitsfluentandimpressiveoutput,GenAIcannotbetrustedtobeaccurate.Indeed,eventheproviderofChatGPTacknowledges,‘WhiletoolslikeChatGPTcanoftengenerateanswersthatsoundreasonable,theycannotbereliedupontobeaccurate(OpenAI,2023).Mostoften,theerrorswillgounnoticedunlesstheuserhasasolidknowledgeofthetopicinquestion.1.2HowdoesgenerativeAIworks?ThespecifictechnologiesbehindGenAIarepartofthefamilyofAItechnologiescalledmachinelearning(ML)whichusesalgorithmstoenableittocontinuouslyandautomaticallyimproveitsperformancefromdata.ThetypeofMLwhichhasledtomanyoftheadvancesinAIthatwehaveseeninrecentyears,suchastheofAIforfacialrecognition,isknownasartificialneuralnetworks(ANNs),whichareinspiredbyhowthehumanbrainworksanditssynapticconnectionsbetweenneurons.TherearemanytypesofANNs.BothtextandimagegenerativeAItechnologiesarebasedonasetofAItechnologiesthathavebeenavailabletoresearcinstance,usesagenerativepre-trainedtransformer(GPT),whileimageGenAItypicallyuseswhatareknownasgenerativeadversarialnetworks(GANs)Table1.Table1.TechniquesusedingenerativeAI491.2.1.HowtextGenAImodelsworkTextgenerativeAIusesatypeofANNknownasageneral-purposetransformer,andatypeofgeneral-purposetransformercalledalargelanguagemodel.ThisiswhyAITextGenAIsystemsareoftenreferredtoaslargelanguagemodels,orLLMs.ThetypeofLLMusedbytextGenAIisknownasagenerativepre-trainedtransformer,orGPT(hencethe‘GPT’in‘ChatGPT’).ChatGPTisbuiltonGPT-3whichwasdevelopedbyOpenAI.ThiswasthethirditerationoftheirGPT,thefirstbeinglaunchedin2018andthemostrecent,GPT-4,inMarch2023(seeTable2).EachOpenAIGPTiterativelyimproveduponthepreviousthroughadvancesinAIarchitectures,trainingmethodsandoptimizationtechniques.Onewell-knownfacetofitscontinuousprogressistheuseofgrowingamountsofdatatotrainitsexponentiallyincreasingnumberof‘parametersParametersmightbethoughtofasmetaphoricalknobsthatcanbeadjustedtofine-tunetheGPT’sperformance.Theyincludethemodel’s‘weights’,numericalparametersthatdeterminehowthemodelprocessesinputandproducesoutput.InadditiontotheadvancementsinoptimizingAIarchitecturesandtrainingmethods,thisrapiditerationhasbeenmadepossiblealsoduetothemassiveamountsofdata5andimprovementsincomputingcapabilitiesavailabletothebigcompanies.Since2012,computingcapabilitiesusedfortrainingGenAImodelshavebeendoublingevery3-4months.Bycomparison,Moore’sLawhadatwo-yeardoublingperiod(OpenAI,Table2.OpenAIGPTsTable2.OpenAIGPTs6OncetheGPThasbeentrained,generatingatextresponsetoapromptinvolvesthefollowingsteps:1.Thepromptisbrokendownintosmallerunits(calledtokens)thatareinputtedintotheGPT.2.TheGPTusesstatisticalpatternstopredictlikelywordsorphrasesthatmightformacoherentresponsetotheprompt.>TheGPTidentifiespatternsofandphrasesthatcommonlyco-occurinitsprebuiltlargedatamodel(whichcomprisestextscrapedfromtheinternetandelsewhere).>Usingthesepatterns,theGPTesttheprobabilityofspecificwordsorphrasesappearinginagivencontext.>Beginningwitharandomprediction,theGPTusestheseestimatedprobabilitiestopredictthenextlikelywordorphraseinitsresponse.3.Thepredictedwordsorphrasesareconvertedintoreadabletext.4.Thereadabletextisfilteredthroughwhatareknownas‘guardrails’toremoveanyoffensivecontent.5.Steps2to4arefinished.Theresponseisconsideredfinishedwhenitreachesamaximumtokenlimitormeetspredefinedstoppingcriteria.6.Theresponseispost-processedtoimprovereadabilitybyapplyingformatting,punctuationandotherenhancements(suchasbeginningtheresponsewithwordsthatahumanmightuse,suchas‘Sure’,‘Certainly’or‘I’msorry’).WhileGPTsandtheirabilitytoautomaticallygeneratetexthavebeenavailabletoresearcherssince2018,whatmadethelaunchofChatGPTsonovelwasitsfreeaccessviaaneasy-to-useinterface,meaningthatanyonewithinternetaccesscouldexplorethetool.ThelaunchofChatGPTsetoffshockwavesaroundtheworld,andquicklyledtootherglobaltechcompaniesplayingcatch-up,alongsidenumerousstart-upcompanies,eitherbylaunchingtheirownsimilarsystemsorbybuildingnewtoolsontop.ByJuly2023,someofthealternativestoChatGPTincludedthefollowing:●Alpaca:7Afine-tunedversionofMeta’sLlama,fromStanfordUniversity,whichaimstoaddressLLMs’falseinformation,socialstereotypesandtoxiclanguage.●Bard:8AnLLMfromGoogle,basedonitsLaMDAandPaLM2systems,thathtotheinternetinrealtime,whichmeansitcanprovideup-to-dateinformation.●Chatsonic:9MadebyWritesonic,itbuildsonChatGPTwhilealsocrawlingdatadirectlyfromGoogle.Accordingly,ithaslesschanceofproducingfactuallyincorrectanswers.●Ernie(alsoknownasWenxinYiyan文心一言):10AbilingualLLMfromBaidu,stillindevelopment,whichintegratesextensiveknowledgewithmassivedatasetstogeneratetextandimages.●HuggingChat:11MadebyHuggingFace,whoemphasizedethicsandtransparencythroughoutitsdevelopment,traininganddeployment.Inaddition,alldatausedtotraintheirmodelsareopensource.●Jasper:12AsuiteoftoolsandAPIsthat,forexample,canbetrainedtowriteinauser’sparticularpreferredstyle.Itcanalsogenerateimages.●Llama:13Anopen-sourceLLMfromMetathatrequireslesscomputingpowerandfewerresourcestotestnewapproaches,validateothers’workandexplorenewusecases.●OpenAssistant:14Anopen-sourceapproachdesignedtoenableanyonewithsufficientexpertisetodeveloptheirownLLM.Itwasbuiltontrainingdatacuratedbyvolunteers.●TongyiQianwen(通義千問):15AnLLMfromAlibabathatcanrespondtopromptsinEnglishorChinese.ItisbeingintegratedintoAlibaba’ssuiteofbusinesstools.●YouChat:16AnLLMthatincorporatesreal-timesearchcapabilitiestoprovideadditionalcontextandinsightsinordertogeneratemoreaccurateandreliableresults.whilesomeareopen-source.ManyotherproductsarebeinglaunchedthatarebasedoneoftheseLLMs.Examplesincludethefollowing:●ChatPDF:17SummarizesandanswersquestionsaboutsubmittedPDFdocuments.●Elicit:TheAIResearchAssistant:18Aimstoautomatepartsofresearchers’workflows,identifyingrelevantpapersandsummarizingkeyinformation.●Perplexity:19Providesa‘knowledgehub’forpeopleseekingquick,accurateanswerstailoredtotheirneeds.Similarly,LLM-basedtoolsarebeingembeddedintootherproducts,suchaswebbrowsers.Forexample,extensionsfortheChromebrowserthatarebuiltonChatGPTincludethefollowing:●WebChatGPT:20GivesChatGPTinternetaccesstoenablemoreaccurateandup-to-dateconversations.●ComposeAI:21Autocompletessentencesinemailsandelsewhere.●TeamSmartAI:22Providesa‘teamofvirtualassistants●Wiseone:23Simplifiesonlineinformation.Inaddition,ChatGPThasbeenincorporatedintosomesearchengines,24andisbeingimplementedacrosslargeportfoliosofproductivitytools(e.g.MicrosoftWordandExcel),makingitevenmoreavailableinofficesandeducationalinstitutionsworldwide(MurphyFinally,asaninterestingtransitiontoimageGenAI,themostrecentGPTfromOpenAI,GPT-4,isabletoacceptimagesaswellastextinitsprompts.Inthissense,itismultimodal.Accordingly,somearguethatthename‘largelanguagemodel’(LLM)isbecominglessappropriate,whichisonereasonwhyresearchersatStanfordUniversi1.2.2.HowimageGenAImodelsworkImageGenAIandmusicGenAItypicallyuseadifferenttypeofANNknownasgenerativeadversarialnetworks(GANs)whichcanalsobecombinedwithvariationalautoencoders.SomeimageGenAImodelslikeDall·EandStableDiffusionuseDiffusionModels,adifferentgenerativeANN.TakingGANsasexampletoexplainhowimageGenAImodelswork:GANshavetwoparts(two‘a(chǎn)dversaries’),the‘generator’andthe‘discriminatorInthecaseofimageGANs,thegeneratorcreatesarandomimageinresponsetoaprompt,andthediscriminatortriestodistinguishbetweenthisgeneratedimageandrealimages.Thegeneratorthenusestheresultofthediscriminatortoadjustitsparameters,inordertocreateanotherimage.Theprocessisrepeated,possiblythousandsoftimes,withthegeneratormakingmoreandmorerealisticimagesthatthediscriminatorislessandlessabletodistinguishfromrealimages.Forexample,asuccessfulGANtrainedonadatasetofthousandsoflandscapephotographsmightgeneratenewbutunrealimagesoflandscapesthatarealmostindistinguishablefromrealphotographs.Meanwhile,aGANtrainedonadatasetofpopularmusic(orevenmusicbyasingleartist)mightgenerate