LSE全球新聞機(jī)構(gòu)人工智能應(yīng)用調(diào)研報(bào)告（英）

POLIS

JournalismatLSE

GeneratingChange

AglobalsurveyofwhatnewsorganisationsaredoingwithAI

CharlieBeckettandMiraYaseen

Preface

Ournewsmediaworldhasbeenturnedupsidedownagain.Asalways,serioustechnologicalchangeproducesbothdystopianandutopianhype.MuchofthishasbeengeneratedonsocialmediabycorporatePRandpoliticians.Newscoverageandexpertcommentaryhasalso

veeredfromexcitedcoverageofpositivebreakthroughsinfieldssuchasmedicinetomuchmorefrighteningvisionsofnegativeforcesunleashed:GenerativeAI(genAI)isproducingatidalwaveofautomated,undetectabledisinformation;itwillamplifydiscrimination,extremespeechandinequalities.

Anditsimpactonjournalism?Again,muchofthecoveragehasfocusedontheunreliabilityofmanygenAItoolsandthecontroversyoveritsrapaciousappetiteforotherpeople’sdatatotrainitsalgorithms.Astheinitialstormofhypeturnsintomorepracticalconsiderationswehave

beentalkingtonewsorganisationsaroundtheworldaboutthisnewwaveoftechnologicalchange.WhataretheydoingwithAIandgenAI;whatmighttheydointhefuture;andwhataretheirhopesandfearsforitsimpactonthesustainabilityandqualityofthishard-pressedjournalismindustry?

WhetheryouareexcitedorappalledatwhatgenAIcando,thisreportmakesitclearthatitisvitaltolearnandengagewiththistechnology.Itwillchangetheworldwereportupon.Itneedscriticalattentionfromindependentbutinformedjournalists.Butoursurveyshowsitisalso

alreadychangingjournalism.Itbringsexcitingopportunitiesforefficiencyandevencreativity.Asonerespondenttoldus,“Freeinguptimeforjournaliststocontinuedoingtheirjobisthegreatestimpactachieved.”

Butitalsobringsspecificandgeneralhazards.Thegoodnewsfromourrespondents,atleast,isthattheyareawareoftheopportunitiesandrisksandarebeginningtoaddressthem.The

bestorganisationshavesetupstructurestoinvestigategenAIandprocessestoincludealltheirstaffinitsadoption.Theyhavewrittennewguidelinesandstartedtoexperimentwithcaution.Thisisacriticalphase(again!)fornewsmediaaroundtheworld.Journalistshaveneverbeenundersomuchpressureeconomically,politicallyandpersonally.GenAIwillnotsolvethose

problemsanditmightwelladdsome,too.

Responsible,effectivejournalismismoreneededthanever.WehopethisreportandourworkatJournalismAIcontributestothatmission.Welookforwardtohearingfromyou.Letusknow

whatyouaredoingandhowwecanhelp.

ProfessorCharlieBeckett

DirectorPolis,LSE,leaderoftheLSE’sJournalismAIproject

1

Contents

Preface1

TheJournalismAISurvey4

ExecutiveSummary&KeyFindings6

Introduction:HowDidWeGetHere?9

Chapter1:HowAIisBeingUsedinJournalismToday14

1.0HowAreNewsroomsUsingAI?14

1.1Newsgathering15

1.2NewsProduction17

1.3NewsDistribution18

1.4WhyNewsroomsUseAI21

1.5WhatisWorkingandWhatisNot22

Chapter2:AIStrategy25

2.0TheNeedforStrategy25

2.1Newsrooms’AIStrategies

24

2.2HowNewsroomProcessesandRolesareAffectedbyAI28

2.3ReadyforAI?31

2.4TheStrategicChallengestoAIAdoption32

2.5HaveNewsrooms’ApproachestoAIIntegrationChanged?36

Chapter3:EthicsandEditorialPolicy39

3.0AI’sImpactonEditorialQuality39

3.1AlgorithmicBias39

3.2NewsroomApproachestoEthicalConcerns41

3.3EthicalImplicationsforJournalismatLarge43

3.4TheRoleofTechnologyCompanies44

3.5TheRoleofUniversitiesandIntermediaryCompanies

47

2

(continued)

Contents

Chapter4:TheFutureofAIandJournalism49

4.0WhereisThisAllGoing?49

4.1TheNeedforEducationandTraining51

4.2NewsroomCollaboration53

4.3HowWillAIChangeJournalism?54

Chapter5:GenerativeAIandJournalism

57

5.0CurrentUseCases57

5.1OpportunitiesPresentedbyGenerativeAI60

5.2ChallengesPresentedbyGenerativeAI62

Chapter6:TheGlobalDisparityinAIDevelopmentandAdoption65

6.0TheGlobalNorth/SouthDivide65

6.1EconomicandInfrastructuralChallenges66

6.2LanguageandAccessibilityChallenges66

6.3PoliticalRealitiesAffectTrustinAI68

Conclusion:WhatDoesAIMeanforJournalism?71

SixStepsTowardsanAIStrategyforNewsOrganisations72

Glossary73

References77

Readings&Resources83

Acknowledgements85

3

7%

28%

16%

20%

13%

16%

TheJournalismAISurvey

Thisreportisthesecondglobalsurveythatwehaveconducted.Thesampleforthisreportisbiggerwithagreateremphasisongeographicaldiversity.Itisbasedonasurveyof105newsandmediaorganisationsfrom46differentcountriesregardingAIandassociated

technologies.In2019wesurveyed71newsorganisationsfrom32differentcountries,ofwhichonly16haveparticipatedagaininthis2023survey.

Thisyear,wemadeitapointtoreachamorediversegroupofparticipantsinterms

ofthesizeoftheirorganisations.Weinvitedsmallandlargenewsrooms,including

emergingandlegacyorganisations.Inadditiontothis,contributionscameinfromLatinAmerica,sub-SaharanAfrica,theMiddleEastandNorthAfrica(MENA),AsiaPacific,

Europe,andNorthAmerica.Thisnecessitatedanadditionalchapter,focusingonregionalchallengestoAIadoption.

NewsOrganisationsThatCompletedTheSurveyByType

30%

20%

10%

0%

BroadcasterNewspaper

Magazine

NewsAgency

PublishingGroup

Other

4

Thepurposeofthisreportisthesameasthefirst:togiveasenseofwhatishappeningwithAIandwhatrisksandopportunitiesitoffers.WeaskedparticipantshowtheyareengagingwithgenerativeAI(genAI)technologiesanditsimplicationsforthefutureofjournalism.Wehopeitinformsthedebate,helpsnewsorganisationscharttheirwayforward,andguidesustodevelopourprogrammestosupportthatprocess.

Thesurveywassupplementedwithinterviews,andconversationsatjournalism

conferences.Weareverygratefultoeveryonewhohassharedtheirthoughts

andexperienceswithus.ThesurveysandinterviewswereconductedbetweenAprilandJuly,2023.

Wedonotclaimthatthesurveyisrepresentativeoftheglobalindustry–thatwouldbe

almostimpossibleonaninternationalscale–nordoesitequallyreflectallviewpoints

withinthedifferentpartsofnewsorganisations.Butitdoesgiveanunprecedentedinsightintohowthesetechnologiesareperceivedbythosepeopleleadingtheirdevelopmentorapplicationinsidenewsorganisations.

Ourrespondentsrepresentdiverserolesandexpertisewithintheirorganisations;they

includejournalists,technologists,andmanagers.Weencouragednewsorganisationstogatherrepresentativesfromdifferentdepartmentstocompletethesurveycollaboratively.

NB:Thelistoforganisationsthatcompletedthesurveycanbefoundintheacknowledgments.

Thepublishedquoteshavegenerallybeenanonymised.Someorganisationnameswereaddedforcontextafterreceivingpermissionfromtheauthors.Somequoteswereeditedlightlytocorrectliteralsandforsense.Theeditorialresponsibilityforthereportliesentirelywiththeauthor.

5

ExecutiveSummary

&KeyFindings

1ArtificialIntelligence(AI)continuestobeunevenlydistributedamongsmallandlargenewsroomsandregionallyamongGlobalSouthandGlobalNorthcountries.

2ThesocialandeconomicbenefitsofAIaregeographicallyconcentratedintheGlobalNorth,whichenjoytheinfrastructureandresources,whilemanycountriesinthe

GlobalSouthgrapplewiththesocial,cultural,andeconomicrepercussionsofpost-independencecolonialism.

3Morethan75%ofrespondentsuseAIinatleastoneoftheareasacrossthenewsvaluechainofnewsgathering,productionanddistribution.

4IncreasingefficiencyandproductivitytofreeupjournalistsformorecreativeworkwerethemaindriversforAIintegrationformorethanhalftherespondents.

5Arounda1/3oftherespondentssaidtheyhadaninstitutionalAIstrategyorwerecurrentlydevelopingone.

6NewsroomshaveawiderangeofapproachestoAIstrategy,dependingontheirsize,mission,andaccesstoresources.Someearlyadoptersarecurrentlyfocusingon

achievingAIinteroperabilitywithexistingsystems,othershaveadoptedacase-by-caseapproach,andsomemediadevelopmentorganisationsareworkingtowardsbuildingAIcapacityinregionswithlowAIliteracy.

7Arounda1/3ofrespondentsbelievetheirorganisationsarereadytodealwiththe

challengesofAIadoptioninjournalism,whilealmosthalfsaidtheywereonlypartiallyreadyornotreadyyet.

8ManyrespondentssaidAIintegrationischangingexistingroleswithinthenewsroomthroughtrainingandupskilling.Alongthesamelines,AIischangingthenatureofajournalist’sroleandsoughtafterskills.

9Aswesawinour2019report,financialconstraintsandtechnicaldifficultiesremainthemostpressingchallengesforintegratingAItechnologiesinthenewsroom.

10Ethicalconcernsarestillsignificantforourrespondents;manyadvocatefor

explainableAIandsettingethicalguidelinestomitigatealgorithmicbias.

11Settingde-biasingtechniquesemergedasahighlychallengingareafor

mostrespondents.

6

12CulturalresistanceandfearsofjobdisplacementandscepticismofAItechnologies

cannotbediscounted.

13Acrosstheboard,respondentsnotedthatmitigatingAIintegrationchallengesrequiresbridgingknowledgegapsamongvariousteamsinthenewsroom.Similarly,cross-

departmentcollaborationwasseenasnecessaryforachievingeffectiveAIadoption.

14ThechallengeofkeepingpacewiththerapidevolutionofAIwasconsistently

mentionedthroughoutthesurvey.

15About40%ofrespondentssaidtheirapproachtoAIhasnotchangedoverthepast

fewyears,eitherbecausetheyarestillinthebeginningoftheirAIjourneyorbecauseAIintegrationremainslimitedintheirnewsrooms.Concurrently,arounda1/4said

theirorganisation’sapproachtoAIhasevolved;theyhavegainedhands-onexperiencethathelpsthemthinkmorerealisticallyaboutAI.

16Morethan60%ofrespondentsareconcernedabouttheethicalimplicationsofAI

integrationforeditorialqualityandotheraspectsofjournalism.JournalistsaretryingtofigureouthowtointegrateAItechnologiesintheirworkupholdingjournalistic

valueslikeaccuracy,fairness,andtransparency.

17RespondentscalledfortransparencyfromthedesignersofAIsystemsand

technologycompanies,andtheusers,namelynewsrooms,withtheiraudiences.

18Journalistsandmediamakerscontinuedtostresstheneedfora‘humanintheloopapproach,’inlinewiththeresultsinour2019survey.

19TherearefearsthatAItechnologieswouldfurthercommercialisejournalism,

boostingpoorqualityandpolarisingcontent,leadingtoafurtherdeclineinpublictrustinjournalism.

20TechcompaniesaredrivinginnovationinAIandothertechnologies,butsurveyparticipantsvoicedconcernsabouttheirprofit-drivennature,theconcentrationofpowertheyenjoy,andtheirlackoftransparency.

21Around80%oftherespondentsexpectalargerroleforAIintheirnewsrooms

inthefuture.

22SurveyparticipantsexpectAItoinfluencefourmainareas:

1Fact-checkinganddisinformationanalysis

2Contentpersonalisationandautomation

3Textsummarisationandgeneration

4Usingchatbotstoconductpreliminaryinterviewsandgaugepublicsentiment

onissues

7

8

23ThereareconcernsthatAIwillexacerbatesustainabilitychallengesfacingless-

resourcednewsroomswhicharestillfindingtheirfeet,inahighlydigitisedworldandanincreasinglyAI-poweredindustry.

24Almost43%ofresponsesemphasisedtheimportanceoftrainingjournalistsandotherpersonnelinAIliteracyandothernascentskillslikepromptengineering.

25Thevastmajoritywelcomedmorecollaborationbetweennewsroomsandothermediaorganisationsandacademicinstitutions,hopingitwouldhelplessenthedisparitybetweensmallandlargenewsrooms,aswellasregionallybetween

newsroomsinGlobalNorthandGlobalSouthcountries.

26Theneedforabalancingactbetweentechandjournalism,athemethatalsoemergedinour2019survey,remainsimperativetoafuturewhereAItechnologiesareleveragedtoservejournalismanditsmission.

27Thevastmajorityofrespondents,around85%haveatleastexperimentedwith

generativeAI(genAI)technologiesinarangeofwayssuchaswritingcode,imagegeneration,andauthoringsummaries.

28SomeareapprehensiveaboutusinggenAIineditorialtasks,whileothersareusing

themregularlyincoding,headlinegeneration,andsearchengineoptimisation.

29TherewasahighlevelofagreementamongparticipantsthatgenAIpresenteda

newsetofopportunitiesnotprovidedbytraditionalAI.TheyhighlightedsomeoftheaffordancesofgenAI,suchasaccessibilityandlowrequirementsforadvancedtechnicalskills.

30Respondentsweremuchmoredivided–almost1/2werenotsure–astowhethergenAIalsopresentedanewsetofchallenges.SomebelievegenAIpresentssimilarchallengestotraditionalAI,suchasalgorithmicbias,butraisestheriskceilingtoanewlevel.

31NewsroomsgloballycontendwithchallengesrelatedtoAIintegration,butthe

challengesaremorepronouncedfornewsroomsintheGlobalSouth.Respondentshighlightedlanguage,infrastructural,andpoliticalchallenges.

Introduction

HowDidWeGetHere?

ArtificialIntelligenceinjournalismhasbeensignificantforsomeyears.TheLSE

JournalismAIprojectstartedbackin2019andourfirstglobalreportpublishedinthesameyearshowedthatitwasakeyemergingsetoftechnologies.AIwasproducingefficienciesfornewsworkanditwasalsocreatingopportunitiesfornewpracticesandproductsorservices.

WeshowedinthatpreviousreportthatarangeofnewsorganisationswereusingAI

acrossthejournalismprocessfromnewsgathering,contentcreationanddistribution,tomarketingandrevenue-gathering.Avariedsetoftechnologieswerebeingused,withprogrammesbasedontrainingsoftwaretomanipulatedata.Advancesinmachine-

learningandNaturalLanguageProcessing(NLP)enablednewsroomstobuildoradapttoolsandsystemstosupporttheirjournalism.

Generally,thesewerelarge-scalebutrelativelybasicfunctionssuchasscraping

socialmediaorautomatingverysimplecontentcreation.Itwasusedbyinvestigativejournaliststocombthroughlargedocumentleaksortohelpautomatepaywallsandtopersonalisecontentinstraightforwardways.Someusesofmachinelearning–suchassearch–weresoroutineanduniversaltheyweretakenforgranted.

In2019wefoundthatnewsorganisationswerefacingvariouschallengesinadoptingAI.Therewasalackofgeneralknowledge,specificskills,andresources.Therewerealso

inequalitiesbetweenbigneworganisationsandsmallerones,especiallythoseinnon-Englishspeakingorlessdevelopedmarkets.

Workingwithnewsorganisationsoverthelastfiveyears,wecouldseethattheimpactofAIwassystemicandaccelerating,justasitwasinotherindustriesandsectors.

Themostsuccessfulorganisationswerethosethattookastrategic,holisticapproachandwhorecognisedthatthesetechnologiesrequiredfundamentalself-analysisoftheorganisation’scapabilitiesandfutureplanning.

InthewidercontextitispossibletoseeAIasathirdwaveoftechnologicalchange

forjournalism.Thefirstwavewasgoingonline,accompaniedbythedigitalisationoftoolsandtheshifttomobile.Thesecondwavewasthearrivalofsocialmediaandtheimpactthathadoncontentcreation,consumptionandcompetition.Thetechnologyplatformsnowprovidedmuchoftheinfrastructureforjournalismandthe‘user’wascentraltoitsdissemination.

9

ThearrivalofgenerativeAI(genAI)inthelastyearhasacceleratedallthesetrends

andcreatednewdisruptions.Thisreportisasurveyofhownewsorganisationshave

continuedtodevelop‘traditional’AIandhowtheyareapproachingthenewchallengesofgenAI.Clearly,itpresentsfreshopportunities,butithasspecialrisksandcharacteristics.Therearecontinuities.Mostnewsorganisationswespoketoweretakingamore

strategicapproachtogenAI,oftenbasedonthelessonsfromdealingwithAIandothertechnologybeforehand.

ItisimportanttostressthatgenAIisprobablythemostrapidlyemergingtechnologyformediainthisdigitalera.Someofthemoreextremedystopiancritiquesandover-heatedmarketinghypehavedistractedfromaproperdebateaboutimmediateconcerns.It

isgoodthatwearenowallawareofAIandabletointeractdirectlywithitandexploreitsforceandflaws.Itishopedthatwewillhaveamoreinclusivedebateaboutwhatitmeansforsocietyingeneralandjournalisminparticular.

Journalismisaspecialpractice.Ontheonehanditisaroundtheworldasectorundergreatcommercial,politicalandcompetitivepressure.Itisweakinresourcescomparedtothegiantcorporationsdevelopingthistechnology.Thepotentialfordeepstructuralthreatstojournalisminthefuturemustbepartofourthinkingnow.Ontheotherhand,newsorganisationshaveshownremarkableresilienceandinnovationinsustainingand

sometimesthrivingdespitethechallengestheyhavefaced.ItmightevenbethatinaworldwheregenAIissuchapower,forillaswellasgood,publicinterestjournalismwillbemoreimportantthanever.

.NewsroomDefinitionsofAI+GenerativeAI

WecontinuetorefertoAIasanumbrellatermforawidevarietyofrelatedtechnologiesandtoacknowledgethatmanyprocessesdescribedasAIoftenincorporatemore

conventionaltechnologies.WeborrowedthesamesimpledefinitionofAIweusedinour2019report:

Artificialintelligenceisacollectionofideas,technologies,andtechniquesthatrelatetoacomputersystem’scapacitytoperformtasksnormallyrequiringhumanintelligence.1

AsforgenerativeAI(genAI),whichwediscussindetailinChapter5:

ItisasubfieldwithinMachineLearning(ML),asubfieldofAIinitsownright,thatinvolvesthegenerationofnewdata,suchastext,images,orcode,basedonagivensetofinputdata.2

10

WewantedtoknowifourrespondentshadanoperationaldefinitionofAI.Astheydidin2019,theresponsesreflectedquitevariedunderstandingsofAI,echoingonceagainthefluidityofthetermandthecomplexityofthetopic.

SomerespondentsofferedaclearoperationaldefinitionofAIastheuseofmachinesorcomputersystemstoperformtasksthattraditionallyrequiredhumanintelligence.Manyofferedtechnicaldefinitionsthatcentredaroundtheconceptsof“automation”,“machinelearning”and“algorithms.”Almosthalftherespondentsusedoneormoreofthosetermsintheirdefinitions:

Itentailsthecreationofalgorithmsandmodelsthatallowmachinestocarryoutoperationslikespeechrecognition,visualperception,problemsolving,anddecision-makingthatordinarilyrequirehumanintelligence.

OtherrespondentsrelatedtheiroperationaldefinitionsofAItoitspotentialbenefitsandtheirmotivesforintegratingitinthenewsroom,suchasincreasingefficiency,orbetterservingthenewsroom’saudienceandmission:

Forus,AIrepresentsagroupoftechnologiesthatcanassistandempower

[ourteam]byprovidinginsightsandautomatedsupportacrossarangeof

editorial,operationalandcommunicationstasks.

Technologiesusedtoautomategatheringandanalysisofdatathatserveoureditorialnicheandmission.

SomehighlightedthecapacityofAItechnologiesto“l(fā)earn”orimprovethemselves:

AIistheuseofadvancedalgorithmsthatareabletoprocess,interpret,

classifyandfindpatternsincomplexand/orlargeamountsofdatainsuchawaythatinfers‘intelligence’or‘human-like’learning.

11

SeveralrespondentsemphasisedtheimportanceofethicalconsiderationsinAI

development,whileothersmentionedconcernsabouttheopacityofAIsystemsortheneedforhumanoversight:

Thesetoftechnologies,tools,processes...thatmakeitpossibletoemulate

humancapabilitiesinordertoautomateorimprovethem,notalwaysforan

ethicalorlawfulpurpose.

AfewrespondentssaidtheydidnothaveaworkingdefinitionofAIyet:

Wedonothaveacollectiveworkingdefinitionyet.Mine,asthepersoninchargeofexploringAIwithinthenewsroom,isthatAIisasetofprocessesthatacomputerdoestoaidandfacilitatehuman’swork,addingintelligencetoit.Bynomeansit

replaceshumanpresenceanditshouldalwaysbecheckedandaccompanied.

12

Thisreportispresentedinsevenchapters.Inordertofacilitatecomparisonsbetweenthisandthe2019report,wekeptthemajorityofthechaptersthesame,withthe

exceptionoftwonewchapters.

TheIntroductiongivesabriefbackgroundofthefindingsfromthe2019report,and

summativeoverviewofthetechnologicalchangesseeninthejournalismindustryoverthepastyears,todate.Wedefinekeyissuesandasummaryofwhatyoucanexpectinthisreport.

ChapterOnefocusesonhowAIiscurrentlybeingusedbynewsrooms.ThechapterlooksathownewsroomsareusingAIacrossthenewvaluechainaswellaswhathasbeenworkingandwhathasnotbeenworking.

ChapterTwounpackstheAIstrategyorlackof,innewsrooms.WelookatthetypesofAIapproachesnewsroomshaveundertaken,someofthekeychallengesandwhatimpactthetechnologycanhaveonthem.

ChapterThreeisalsosimilartothepreviousreportasweexpandonethicsand

editorialpolicy.

ChapterFourlookstothefutureandroleofAIinjournalism.

ChapterFivetouchesongenerativeAIandjournalism.ItisanewchapterthatlooksatthecurrentusecasesofgenAI,aswellasitsopportunitiesandchallenges.

ChapterSixreflectsontheglobaldisparityinAIdevelopmentandadoptionaswellasthechallengesfacedbythemajorityoftheworld’spopulationintheGlobalSouth.

TheConclusiontiesalltheabove-mentionedchapterstogetherandgivesabriefanalysisofwhatallthismeansforjournalism.Weconcludethemainbodyofthereportwithasix-steproadmaptowardsanAIstrategythatnewsroomscouldborrowfrom.Youwillalsofindaglossary,endnotes,referencesandalistofsuggestedreadingsandresources.

ThisworkwasfundedbytheGoogleNewsInitiativeandcarriedoutbyateamledby

ProfessorCharlieBeckett,directoroftheLSE’sinternationaljournalismthink-tank,Polis.

Wewouldliketothankallthejournalists,technologistsandresearcherswhotookpartintheproject.TheprojectwasmanagedbyTshepoTshabalalaandtheleadresearcherandco-authorwasMiraYaseen

13

100%

75%

50%

25%

0%

Chapter1

HowAIisBeingUsedinJournalismToday

1.0HowAreNewsroomsUsingAI?

WeaskednewsroomshowtheyareusingAItechnologiestoday.Theseareas–

newsgathering,newsproductionandnewsdistribution–coverallthecontentcreationstagesfromideationtopublishing.Thesethreeareasoftenintersectgiventhenatureof“contemporary‘networked’3or‘hybrid’4journalism.Forinstance,fact-checkingchatbotsareleveragedinnewsproductiontovalidateorrefutecertainclaims.Atthesametime,thedatacollectedcouldassistindetectingmisinformationtrendsandinspireatopicforafeaturearticle,thuscontributingtothenewsgatheringprocess.

90%

75%

80%

NewsgatheringNewsProductionNewsDistributi