KPMG+生成式人工智能模型的商業(yè)風(fēng)險(xiǎn)和潛在回報(bào)

KPMGInternational

GenerativeAImodels—therisksandpotentialrewardsinbusiness

WhattheriseofChatGPT,DALL?E2,Bardetal.couldmeanforyourorganization.

April2023

1

GenerativeAImodels—therisksandpotentialrewardsinbusiness

LisaHeneghan

GlobalChiefDigitalOfficerKPMGInternational

GenerativeAImodelshighlightthepoweroftechnology.Theyhavethepotentialtomakeusmoreproductiveandcanmakewhatwedoeasierinsomerespects.However,thesemodelscomewithriskimplicationsthatallorganizationsandindividualsshouldbeawareof.Thatsaid,wecan’tignorethesemodels.They’rerapidlybecomingpartofourdailypersonalandprofessionallives.Weneedtodeterminehowtoembracethem—butsafely.”

ThecoverimageandimagerythroughoutthispaperweredesignedusingDALL?E2,anAIartgeneratorthatcreatesimagesbasedontextualdescriptions.

Imagepromptsforthecoverimagewere:fluidabstract,wavycolumnofblueandpurple,splash,drops,purplebackground.

WhileDALL?E2generatescompellingvisualcontent,ithasnotbeentrainedonKPMGbrandguidelines.Also,ithasneitherthehumanexpertisenoringenuitytounderstandKPMG’sbrandpositioning.Asaresult,theseimagesareconsideredoff-brandandareusedforillustrativepurposesonly,withspecialpermissionfromKPMG

GlobalBrand.

2

?2023CopyrightownedbyoneormoreoftheKPMGInternationalentities.KPMGInternationalentitiesprovidenoservicestoclients.Allrightsreserved.

3

Contents

Executivesummary04

Marketoverview05

WhataregenerativeAImodels?06

HowgenerativeAImodelswork06

Potentialopportunitiesandusecases07

Currentconsiderations10

Whatdoesthefuturehold?13

HowKPMGcanhelp14

Contacts15

3

GenerativeAImodels-therisksandpotentialrewardsinbusiness

?2023CopyrightownedbyoneormoreoftheKPMGInternationalentities.KPMGInternationalentitiesprovidenoservicestoclients.Allrightsreserved.

4

01

Executivesummary

Webelievethatgenerativeartificialintelligence(AI)modelshavethepotentialtotransformbusinessesthroughautomatingandexecutingcertaintaskswithunprecedentedspeedandefficiency.Thisisparticularlytruewhenhumanexpertiseandingenuityispairedwithdeepunderstandingofhowtousetheseprogramsandeffectivelyharnesstheircapabilities.

However,itwilltaketimeandhumanexpertisetounlocktheirfullpotentialinawaythat’sresponsible,trustworthy

andsafe.Ifyou’reconsideringusinggenerativeAI,it’simportanttoestablishasetofinternalprocessesandcontrolsforeveryoneinyourorganizationtofollow.

Inthisreport,wecoverpotentialusecasesandopportunities,aswellaswhattoconsiderifyou’rethinkingaboutusinggenerativeAIapplications,suchasChatGPT,withinyourorganization.

Hereare10thingsyoushouldknowaboutgenerativeAI:

1

ThemostcommongenerativeAIsolutionscanroughlybedividedintofivecategories:contentgenerators,informationextractors,smartchatbots,languagetranslatorsandcodegenerators.

2

GenerativeAImodelscansummarizearticles,draftemailsandproduceimagesandvideos.Trainedbyhumans,somegenerativeAImodelshavetheconversationalskillsto,forexample,answerfollow-upquestions,admitmistakes,challengeincorrectassumptionsandfilterorrejectinappropriaterequests.

3

ChatGPTisachatbottrainedonhumaninstructions.Itsinitialunderlyinglargelanguagemodel,GPT-3.5,had175billionparametersandwastrainedwithmorethan1milliondatasetsor500billiontokens(wordsorwordfragments).GPT-3.5wasnotconnectedtotheinternetandwastrainedondatafromuptoSeptember2021.GPT-4,OpenAI’snewlargemultimodalmodel,evolvedfromitsearlierlargelanguagemodel.

4

GenerativeAImodelshaveuseacrossvariousbusinessfunctions,fromIT,humanresourcesandoperations,tofinance,audit,legalandmarketing.Suitableapplicationsincludedraftingproposals,developingandtestingcode,andextractingandsummarizingcomplexinformation.

5

GenerativeAItakesdatainputsorparameterstolearnandbuildknowledge.Unlessyouexplicitlyrestricttheapplicationproviderfromdoingso,thatdatamaythenbeusedtoanswerapromptfromsomeoneelse,pos-siblyexposinganorganization’sproprietaryinformationtothepublic.Dependingontheapplication,youmayalsobesigningoveryourcopyrights.Referringtotherespectivetermsandconditionsmaygiveyouanideaofwhathappenswithuser-inputteddata.

6

DependingwhatyouusegenerativeAIforandhowyouimplementit,youractivitiescouldexposeintellectualpropertyortradesecretsandopenyourorganizationuptofraudrisk.It’simportanttobevigilantandmakesureyourorganizationisn’tusingAIinawaythatcontravenesapplicablelaws(includingprivacylaws),clientagreementsorprofessionalstandards.

7

CopyingAI-producedinformationorcodeintoanydeliverableorproductmayconstitutecopyrightorotherintellectualpropertyinfringement.Thiscouldpotentiallycauseyourorganizationlegalandreputationalharm.

8

WeexpectbothopensourceandboutiqueversionsofgenerativeAIwillcontinuetobeintegratedintomanycommonapplications,systemsandprocesses,rangingfrominternetbrowserstoAI-connectedtechnologythatorganizationslicense,suchascloud-basedsoftwareandinstantmessagingprograms.

9

CreatingsafeusageguidelineswithinyourorganizationiskeytohelpingensureproperandeffectiveuseofgenerativeAIapplications.Yourorganizationshouldalsoupskillitspeople,asthehumanintheloopbringsuniqueinsightsandunderstandingthatgenerativeAIalonecan’treplicate.

10

KPMGtakesaresponsibleapproachtodesigning,buildinganddeployingAIsystemsinasafe,trustworthyandethicalmanner.Thisapproachhelpscompaniesacceleratevalueforconsumers,organizationsandsociety.

5

Marketoverview

02

AccordingtoresearchandconsultingfirmGartner,by2025,30percentofoutboundmessagesfromlargeorganizationswillbesyntheticallygenerated.1Inthe

AIRiskSurvey

Report

,conductedbyKPMGintheUSinSeptember2022,85percentofrespondentsexpectanincreaseintheuseofAIandpredictiveanalyticsmodels.Additionally,inthe2022

KPMGintheUSTechnologySurvey

,halfofrespondentssaidthey’veseenROIfrominvestmentsinAItechnology.

GenerativeAImodelscapturedattentioninsummer2022whenanAI-generatedimagewonanartcontest.2InNovembertheywereinthespotlightagainfollowingthelaunchofChatGPT.However,itwasduringaJanuary2023

WorldEconomicForum

session,whenMicrosoftChairmanandChiefExecutiveOfficerSatyaNadellasaidthe“goldenageofAI”isunderway,3thatthebuzzaroundChatGPTreallybegantointensify,generatingmanyquestionsfrom—andconversationswith—KPMGmemberfirmclients.

Trainingthesemodelsrequireslargeamountsofventurecapital,humaneffortandcomputingpower.OpenAI,thecreatorofChatGPT,receivedUS$1billionfromMicrosoft4andanothermultiyear,multibilliondollarinvestmentfromthecompanyatthestartof2023,5andGoogle6andMeta7havecreatedgenerativeAImodelsoftheirown.Giventherangeofpossibleapplications,anentireindustryisbeingbuiltonmakinggenerativeAImodelsuseful.

GenerativeAIapplicationscanberoughlydividedintofivecategories:contentgenerators,informationextractors,smartchatbots,languagetranslatorsandcodegenerators:

?Contentgenerators:Wheregenerativepretrainedtransformertoolsgeneratecontentsuchasblogposts,emails,socialmediaposts,images,webcopyandads.

?Informationextractors:Theseapplicationscancreateshort-andlong-formsummariesofnewsarticles,blogposts,legaldocumentsandmore.Somecompaniesusethemtodevelopandanalyzelegaldocuments.

?Smartchatbots:Companiesareincreasinglyusingsmartchatbotsasconsumerassistants.Thechatbotsinteractinaconversationalwayandcananswerfollow-upquestions,admitmistakes,challengeincorrectideasandrejectinappropriaterequests.

?Languagetranslators:Multilingualtoolsthatcantranslatemanylanguages.Theyhavethepotentialtobuildentirewebsiteinterfaces,includingtranslationsites.

?Codegenerators:GenerativeAImodelscanconvertnaturaltextinputsintocodesnippetsorapplications.Withabasicdescriptionorsmallprogramfunctioninput,thesemodelscanproducecodeinvariousprogramminglanguages,andidentifyandfixbugs.

1Gartner,7TechnologyDisruptionsThatWillCompletelyChangeSales,October10,2022./en/articles/7-technology-disrup-tions-that-will-completely-change-sales.GARTNERisaregisteredtrademarkandservicemarkofGartner,Inc.and/oritsaffiliatesintheU.S.andinternationallyandisusedhereinwithpermission.Allrightsreserved.

2https://www.smithsonianmag.com/smart-news/artificial-intelligence-art-wins-colorado-state-fair-180980703/

3/press/2023/01/satya-nadella-says-ai-golden-age-is-here-and-it-s-good-for-humanity

4/blog/microsoft-invests-in-and-partners-with-openai

5/blog/2023/01/23/microsoftandopenaiextendpartnership/

6https://blog.google/technology/ai/bard-google-ai-search-updates/

7/

6

HowgenerativeAImodelswork

03

WhataregenerativeAImodels?

GenerativeAIreferstoartificialintelligencethatcangeneratecontentratherthansimplyanalyzeoractonexistingdata.

GenerativeAImodels,suchasGPT-4,arebuiltandtrainedonacollatedsetofdata.Theycanbegeneralistsorspecialistsbuiltonpredefineddatacollectionsandaredesignedtoproduceoutputthathelpsrealizecertainhuman-directedrequests.Somemodelscan,forexample,predictthenextwordbasedonpreviousphrasesorthenextimagebasedondescriptionsofimagesthatcamebefore.

Thistrainingenablesthequickgenerationoforiginalcontent,includingtext,images,videoandcode.Withareduced

needforhumanresources,somecompaniesexpecttobeabletoproducecontentfasterandatlowercost,givingthemopportunitiestocreatenewkindsofcontentthatwerepreviouslytooexpensiveortimeconsuming.Thisfundamentallychangeshuman-machineinteractionandopensmyriadpotentialusecases.

Thispredictivecapabilityenablesthemodelstoanalyze.Forexample,theycanbeusedtoidentifydocumentsthatareaboutatopicdescribedbyinputtedtext.

04

GenerativeAImodelsaredesignedtoproducecontentbasedonaclearsetofinputsandrules.

ThemostbuzzedaboutapplicationofagenerativeAImodeloflateisChatGPT,achatbottrainedonhumaninstructionscreatedbySanFranciscoresearchlabOpenAI.8Asof

14March2023,ChatGPTPlussubscriberswereabletouseGPT-4,alargemultimodalmodel(LMM)thatacceptsbothimageandtextinputsandgeneratestextoutputs.9On

23March2023OpenAIlaunchedpluginsforChatGPT,includingitsownwebbrowsingplugin.ThismeansChatGPTcannowaccesscertainthird-partysourcesanddatabases.10

ChatGPTstandsforChat(conversation-based)G(generative)P(pretrained)T(transformer).Itwasfinetunedusingreinforcedlearningfromhumanfeedback,wherearewardmodelrepresentinghumanpreferenceistrainedtohelpmakeoutputssoundmorehuman,preventundesirableresponsesandtrytoavoidhallucinations(makingupfacts).

ChatGPTwascreatedasalargelanguagemodel(LLM)andhassinceevolvedintoalargemultimodalgenerativeAIapplication.Thismeanstheapplicationcannow

acceptimageandtextinputs,notjusttextinputsaswaspreviouslythecase.Combinedwithaneuralnetworkmodelthatusesunsupervisedlearningtopredictoutcomes,thistypeofgenerativemodelcandeterminethemostlikelylinguisticpatternsandrelationshipsbetweencontentit’salreadyabsorbed.

The“l(fā)arge”referstotheamountofdatathemodelsarebasedon,aswellasthesizeofthemodelsthemselves,andinvolvestrainingthemwithamassivecollectionofpubliclyaccessibleelectronicdocuments.Forinstance,whenitwasreleasedin2022,ChatGPThad175billionparameters(avaluethatcontrolsthebehaviorofthemachinelearningmodel—thegreatertheparameters,thegreaterthemodel’sabilitytoanalyze).

Itwasinitiallytrainedwithmorethan1milliondatasetsor500billiontokens(wordsorwordfragments),includingfromWikipediaandTheNewYorkTimes.Toputthisintoperspective,theaveragehumanspeaks860.3millionwordsintheirlifetime,11makingthiscollection—or“corpus,”inAIterminology—equivalentto300years’worthoflanguage.

8/sites/cindygordon/2023/02/02/chatgpt-is-the-fastest-growing-ap-in-the-history-of-web-applications/?sh=7510d916678c

9/research/gpt-4

10ChatGPTplugins()

11/books/OL3502128M/The_joy_of_lex

7

Potentialopportunitiesandusecases

ThebasicversionofChatGPTisn’tconnectedtotheinternetandwastrainedononlinematerialuptoSeptember2021meaningitsknowledgeisn’tuptodate.Newerimplementationsreleasedtoasmallnumberofpremiumdevelopers,suchasaplugintoBing’ssearchengine,12areconnectedtotheinternetandcontainmorerecentcontent.

ChatGPTisanexampleofartificialnarrowintelligence(ANI).ANIsystemsaresuitableforperformingonetypeoftaskforwhichtheyhavebeentrained.Forexample,anANIsystemdesignedtogenerateimageswillunlikelybeabletosolvemathproblems.

AccordingtoOpenAI,GPT-4—thoughstillnotfullyreliable—issubstantiallymoresoandcapableofhandlingnuancedinstructionscomparedtoitspredecessor,GPT-3.5.Mostsignificantly,itpassedasimulatedbarexamaroundthetop10percentoftesttakers.Incomparison,GPT-3.5scoredinthebottom10percentofthesamesimulatedexam.OpenAIpointsoutthatitslimitationsaresimilartoearlierGPTmodels,hallucinatingfactsandmakingreasoningerrors.13

Keepreadingtolearnaboutpotentialusecasesfor

generativeAImodels.

05

ChatGPT’smeteoricriseinpopularityisinpartbecauseanyonecanuseit,eventhosewhodon’thavetechnicalbackgrounds.Itsrapidgrowthinusers—100millionasofFebruary202314—isasignofpeople’seagernesstousethetechnology.Andthemoreusersachatbothas,thebettertraineditsunderlyingAIcanget.

ChatGPThasthepotentialtotransformbusinessesbyautomatingandexecutinglanguage-basedtaskswithunprecedentedspeedandefficiency.LMMscanbedeployedtohelpwithawiderangeoftasks.Theycanbemodifiedtosummarizeandclassifylegaldocuments,respondtoconsumerquestions,assistexpertadvisors,andgenerateengineeringandarchitecturaldrawings.

Theycanactasastartingpointforhumaninspiration,providingideasthatcanbetransformedintofreshandcreativethinking.Thismakesthemsuitabletohelpgeneratebusinessreports,marketingpitchesandcodeforsoftwareapplications.

GenerativeAImodelsmayalsohaveapplicationsacrossIT,audit,humanresources,operationsandmanymorebusinessfunctionsbesides.Asyouexploretheseusecases,keepinmindthatdespitethemanyopportunitiesgenerativeAIhastooffer,theyaren'triskfree.

12/new

13/research/gpt-4

14/technology/2023/feb/02/chatgpt-100-million-users-open-ai-fastest-growing-app

8

IntheIToperationsspace,theycouldbeusedfor:

LMM-basedknowledgemanagementsystems

Gatheringinformationfromvariousdatasourceformats.Thisinformationcanthenbequeriedtosearchforspecificitems.

Self-serveITsupport

HelpingemployeesaddressITsystemerrorsthroughsupportinstructionsgeneratedbyconversationalAIchatbots.

Codingortestingcode

Convertingcodefromonefunctiontoanother,forinstancefromSQLtoPython,ortestingcodetoensureitworks.

Ontheaudit/compliancefront,theycouldhelpwith:

Automatingauditsubmissionfactfindinganddetailedauditreviewsbasedonqueryformats.

Automatingauditreview

Evaluatingindependencerequirements

Evaluatingauditengagementindependencerequirementstohelpsimplifytheapprovalsprocessforcertifying

independence.

Potentialuseinhumanresourcesincludes:

TraininggenerativeAImodelsonjobdescriptionandrelevantskillsdatatohelpidentifysuitablejobcandidates.

Self-serviceapplications

Candidateselection

DeployingchatbotsthatcanbothshareknowledgeinahumanwayandresolveHRqueries.

Ontheoperationsfronttheycanhelpwith:

ContextualizingESGdataandsupportingreportingoperations,includingcreatingplain-languagestatementsthatoutline

SustainabilityandESGreporting

ESGinitiatives.

Coordinatingeventmanagementthroughdraftinginvitations,schedulingsessionsandansweringattendee

Virtualeventmanagement

questions.

Simplifyingbusinessoperations

Fromdraftingemailsandpreparingforrequestforproposalstorunningacompetitiveanalysisandresearchingtoensuremarketunderstanding.

Inthefinance/logisticsspace,theycouldassistby:

Helpingorganizationsmaketaxcontributionspubliclyavailablebysortingthroughmassivevolumesofdata.

Draftingandreviewingcontractterms

Categorizingandvalidatingpayments

thecontractsprocess.

Reviewingcontractsandhighlightingpotentialconflictofinterestclausesanddraftingclausesandtermstohasten

9

Legalandorganizationalgovernanceoptionsinclude:

Enablingorganizationstoprovidepersonalizedresponsestoindependence-relatedqueriesviachatbots.

Makingpersonalizedindependencerecommendationsforinvestments

Surfacinglegalcitationsandsourcelinks

Searchingforrelevantlegalcitationsandcaseexamples,helpingtoidentifyreputablesources.

Potentialmarketingapplicationsinclude:

Findingalternatewordchoicesthattranslatewellacrossavarietyoflanguages.

Simplifyingcampaignlanguage

Helpinglocalizeglobalcampaignsbysharinglocalconversationdatawiththemodel.

Distillingcomplexinformation

Localizingmarketingcommunicationsatscale

Learningthebasicsof,forexample,financialduediligence,todigestandstructurecontenttohelpbuildastrong

marketingcampaign.

AtKPMG,we’rerecognizedleadersinAI,machinelearning,anddata&analytics,andhavedeepexpertiseinriskconsiderationswithrespecttoemergingtechnology.KPMGcanhelpassesstheethics,governanceandsecurityinplacearoundclients’AIandmachinelearningtechnologies.

KPMGfirmshavelongbeenattheforefrontofexploringandharnessingnewtechnologiesandcananswerquestionsabouthowgenerativeAIapplicationsmayhelpyourorganizationgrow.

06

Currentconsiderations

You’venowseenhowgenerativeAImodelsmayhelpconsumers,streamlineorganizationalprocessesandfreeuptimeforemployeestotakeonhigher-valueorganizationaltasks.Thatsaid,theuseofgenerativeAIhasmanylimitsandpotentialpitfalls.

Aswementioninpart4,GPT-3.5,ChatGPT’sinitialunderlyingLLM,wastrainedonmaterialuptoSeptember2021andwasn’tconnectedtotheinternet.AndthoughOpenAIhasmadeitpossibleforChatGPTtobrowsetheinternetincertaincases,it’sstillcrucialtoensurehumanreviewandexpertiseisbuiltintothegenerativeAIuseprocess.GenerativeAImodelscanbethecoreofanAIapplicationbutrequireadditionalanalytics,technologyandhumanprocessaroundthemtosolveproblems.

Inthissection,wediscusstherisksofusinggenerativeAImodelsandapplicationsandhowtomanagethem,includingaroundclientandcompanyconfidentiality,employeemisuseandphishing.

LargemultimodalmodelslikeChatGPTgeneratehuman-likeresponses.However,theylackhuman-likereasoningskills.Forthemtobeconsideredtrusted,usersareresponsibleforapplyingtheirAIcapabilitiestosuitableusecases,andyourorganizationshouldeducateemployeesonusingsuchprograms.Equallyimportant,developersshouldusereliabledatasetstotraintheAImodelsandapplyrelevantbiasandcontentfilters.

Riskmanagement

GenerativeAI’sgrowingpopularityisanotherreasonwerecommenddevelopinganddeployingAIinaresponsiblewayifyourorganizationwantstoprotectitselfagainstmisuse.ThefollowingaresomeoftheriskmanagementchallengesofgenerativeAImodels.

Internalrisksandconsiderations

Breakingconfidentialityandintellectualproperty

ManygenerativeAImodelsarebuilttoabsorbuser-inputteddatatoimprovetheunderlyingmodelsovertime,inessencehelpingthemlearnandbuildknowledge.Thatdata,inturn,couldbeusedtoanswerapromptfromsomeoneelse,possiblyexposingprivateorproprietaryinformationtothepublic.Themoreyourbusinessusesthistechnology,themorelikelyitisotherscouldaccessyoursensitiveorconfidentialinformation.Thus,yourorganizationneedstofigureouthow

toprotectitsintellectualpropertywhilestillbeingabletoenjoythebenefitsofgenerativeAIapplications.

Employeemisuseandinaccuracies

EvenlegitimateuseofgenerativeAIcomeswithrisk.Themodelsgenerateresponsesbasedoninputreceived,meaningthere’sarisktheymayprovidefalseormaliciouscontent.Asyouremployeesuseit,theyneedtobecautiousandreviewAI-generatedcontentwithacriticaleyeandemphasisonqualityassurance.

IfgenerativeAIcontentcontainsinaccuraciesthatarenotcaught,thiscouldimpactyourbusiness’outcomesorcreateliabilityissues.Forinstance,Meta’sgenerativeAIbotGalacticawascreatedtocondensescientificinformationtohelpacademicsandresearchersquicklyfindpapersandstudies.Instead,itproducedvastamountsofmisinformationthatincorrectlycitedreputablescientists.15AnotherMetabot,BlenderBot3,wascaughtmakingfalseandbiasedclaims16shortlyafteritsreleaseinAugust2022.Aswell,Google’schatbotBardcausedparentcompanyAlphabettoloseUS$100billioninmarketvalueafteritsharedincorrectinformationduringitsfirstdemo.17ChatGPThallucinatingfactsisalsowelldocumented,18,19withdeveloperOpenAIacknowledgingitsongoingshortcomings.20

OtherrisksaroundgenerativeAIincludethepossibilitythatthetechnologycouldgeneratesensitiveinformation,suchaspersonaldata,thatcouldbeusedforidentitytheftortoinvadeone’sprivacy.Evenadisgruntledemployeeorangrycustomercouldcreatefakematerialtoharmyourcompany’sreputationorthatofoneofyouremployeesorexecutives.

GenerativeAIevolves

Astheworld’sunderstandingofAIcontinuestoevolve,wearealreadyseeingarisingnumberofglobalregulations.It’simportanttostayabreastofthese,evenifyoudon’tplanonusinggenerativeAIintentionally.

WeexpectthatgenerativeAIwillcontinuetobeintegratedintomanycommonapplications,systemsandprocesses,rangingfrominternetbrowserstoAI-connectedtechnologythatyourorganizationmaylicense.Thus,it’skeytobevigilantandmakesureyoudon’tuseAIprofessionallyinawaythatcontravenesapplicablelaws(includingprivacylaws),clientagreementsorprofessionalstandards.

15/meta-ai-bot-galactica-1849813665

16/2022/08/11/tech/meta-chatbot-blenderbot/index.html

17/2023/02/09/1155650909/google-chatbot--error-bard-shares

18/2023/03/17/1164383826/heres-what-the-latest-version-of-chatgpt-gets-right-and-wrong

19/factual-errors-inflated-bios-aside-100209244.html

20/research/gpt-4

10

Questionstoconsider:

1.HowcanyouensureconfidentialityandaccuracyaremaintainedwhileusinggenerativeAImodels?

2.HowcanyouensureyourgenerativeAImodelscomplywithgrowingglobalregulations?

3.Howcanyouautomatereviewingandmanagingcompliancepolicies?

4.WhatshouldyourworkforceknowaboutgenerativeAIintermsofitsrisksandbenefits?

Talentimplications

High-quality,expertoutputcanonlybeachievedwithhigh-quality,expertqueries.Therefore,yourorganizationwillneedtoupskillitsworkforceandretainproprietaryknowledgetocontextualizethequeryandprovidetherightprompts.AtKPMG,forexample,we’vemadegenerativeAItrainingavailabletoallourpeoplethroughourDigitalandDataFoundationsprogram,whichprovidesfoundationalcontentontheevolutionofAIandhowtobuild,implementandengagewithtrustworthyAI.

Professionalsneedtobemadeawarethatthey’renotjustusingasolution—they’retrainingandevolvingit.

Inagenerativefuture,weanticipatethattheroleofprofessionalswillshiftfromproblemsolvingtoproblemdefiningasteamsworkalongsidemachinestocreatenewapproaches.GenerativeAItoolsareaninterface,notanoracle.

ThehumanintheloopbringsuniqueinsightsandunderstandingtotheprocessthatgenerativeAIalonecan’treplicate.Theyprovidecriticalfeedbacktorefineandimprovethemodelovertimeandensuretheoutput’saccurate,fairandmeetsthedesiredgoals.

Greatthingscanhappenwhenpeopleandtechnologyareinharmony,andwestronglybelievetherecanbenolastingchangewithouthumaningenuity.

Externalrisksandconsiderations

Misinformation,biasanddiscrimination

Aswediscussabove,LLMsandLMMshavesharedfalse,