Nature：生成式AI對(duì)在線知識(shí)社區(qū)的影響 The consequences of generative AI for online knowledge communities

ScientifcReports|(2024)14:10413|

/10.1038/s41598-024-61221-0

natureportfolio1

www./scientificreports

scientificreports

OPEN

TheconsequencesofgenerativeAIforonlineknowledgecommunities

GordonBurtch,DokyunLee&ZhichenChen

Generativeartifcialintelligencetechnologies,especiallylargelanguagemodels(LLMs)likeChatGPT,arerevolutionizinginformationacquisitionandcontentproductionacrossavarietyofdomains.

Thesetechnologieshaveasignifcantpotentialtoimpactparticipationandcontentproduction

inonlineknowledgecommunities.Weprovideinitialevidenceofthis,analyzingdatafromStack

OverfowandRedditdevelopercommunitiesbetweenOctober2021andMarch2023,documentingChatGPT’sinfuenceonuseractivityintheformer.Weobservesignifcantdeclinesinbothwebsite

visitsandquestionvolumesatStackOverfow,particularlyaroundtopicswhereChatGPTexcels.Bycontrast,activityinRedditcommunitiesshowsnoevidenceofdecline,suggestingtheimportance

ofsocialfabricasabuferagainstthecommunity-degradingefectsofLLMs.Finally,thedeclinein

participationonStackOverfowisfoundtobeconcentratedamongnewerusers,indicatingthatmorejunior,lesssociallyembeddedusersareparticularlylikelytoexit.

Recentadvancementsingenerativeartifcialintelligence(GenAI)technologies,especiallylargelanguagemodels(LLMs)suchasChatGPT,havebeensignifcant.LLMsdemonstrateremarkableprofciencyintasksthatinvolveinformationretrievalandcontentcreation

1

–

3

.Giventhesecapabilities,itisimportanttoconsidertheirpotentialtodriveseismicshifsinthewayknowledgeisdevelopedandexchangedwithinonlineknowledgecommunities

4

,5

.

LLMsmaydrivebothpositiveandnegativeimpactsonparticipationandactivityatonlineknowledgecom-munities.Onthepositiveside,LLMscanenhanceknowledgesharingbyprovidingimmediate,relevantresponsestouserqueries,potentiallybolsteringcommunityengagementbyhelpinguserstoefcientlyaddressawiderrangeofpeerquestions.Viewedfromthisperspective,GenAItoolsmaycomplementandenhanceexistingactivitiesinacommunity,enablingagreatersupplyofinformation.Onthenegativeside,LLMsmayreplaceonlineknowledgecommunitiesaltogether.

Ifthedisplacementefectdominates,itwouldgiverisetoseveralseriousconcerns.First,whileLLMsoferinnovativesolutionsforinformationretrievalandcontentcreationandhavebeenshowntosignifcantlyenhanceindividualproductivityinavarietyofwritingandcodingtasks,theyhavealsobeenfoundtohallucinate,i.e.,providing‘confdentlyincorrect’responsestouserqueries

6

,andtoundermineworkerperformanceoncertaintypesoftasks

3

.Second,ifindividualparticipationinonlinecommunitiesweretodecline,thiswouldimplyadeclineinopportunitiesforallmannerofinterpersonalinteraction,uponwhichmanyimportantactivitiesdepend,e.g.,collaboration,mentorship,jobsearch.Further,totheextentasimilardynamicmayemergewithinformalorganizationsandworkcontexts,itwouldraisetheprospectofanalogousdeclinesinorganizationalattachment,peerlearning,careeradvancementandinnovation

7

–

12

.

Withtheaboveinmind,weaddresstwoquestionsinthiswork.First,weexaminetheefectsthatgenerativeartifcialintelligence(AI),particularlylargelanguagemodels(LLMs),haveonindividualengagementinonlineknowledgecommunities.Specifcally,weassesshowLLMsinfuenceuserparticipationandcontentcreationinonlineknowledgecommunities.Second,weexplorefactorsthatmoderate(amplifyorattenuate)theefectsofLLMsonparticipationandcontentcreationatonlineknowledgecommunities.Byaddressingtheserelationships,weaimtoadvanceourunderstandingoftheroleLLMsmayplayinshapingthefutureofknowledgesharingandcollaborationonline.Further,weseektoprovideinsightsintoapproachesandstrategiesthatcanencourageasustainableknowledgesharingdynamicbetweenhumanusersandAItechnologies.

WeevaluateourquestionsinthecontextofChatGPT’srelease,inlateNovemberof2022.Westartbyexam-ininghowthereleaseofChatGPTimpactedStackOverfow.WeshowthatChatGPT’sreleaseledtoamarkeddeclineinwebtrafctoStackOverfow,andacommensuratedeclineinquestionpostingvolumes.Wethenconsiderhowdeclinesinparticipationmayvaryacrosscommunitycontexts.LeveragingdataonpostingactivityinRedditdevelopercommunitiesoverthesameperiod,wehighlightanotablecontrast:nodetectibledeclinesinparticipation.Weattributethisdiferencetosocialfabric;whereasStockOverfowfocusesonpureinforma-tionexchange,Redditdevelopercommunitiesarecharacterizedbystrongersocialbonds.Further,consideringheterogeneityacrosstopicdomainswithinStackOverfow,weshowthatdeclinesinparticipationvariedgreatly

QuestromSchoolofBusiness,BostonUniversity,Boston,MA02215,USA.email:gburtch@

ScientifcReports|(2024)14:10413|

/10.1038/s41598-024-61221-0

natureportfolio2

www./scientificreports/

dependingontheavailabilityofhistoricalcommunitydata,alikelyproxyforLLM’sabilitytoaddressques-tionsinadomain,giventhatdatawouldlikelyhavebeenusedintraining.Finally,weexplorewhichusersweremostafectedbyChatGPT’srelease,andtheimpactChatGPThashadonthecharacteristicsofcontentbeingposted.WeshowthatnewerusersweremostlikelytoexitthecommunityaferChatGPTwasreleased.Further,andrelatedly,weshowthatthequestionspostedtoStackOverfowbecamesystematicallymorecomplexandsophisticatedaferChatGPT’srelease.

Methods

Toaddressthesequestions,weleverageacombinationofdatasourcesandmethods(additionaldetailsarepro-videdinthesupplement).First,weemployaproprietarydatasetcapturingdailyaggregatecountsofvisitorsto,andalargesetofotherpopularwebsites.TisdatacoverstheperiodfromSeptember2022throughMarch2023.Additionally,weemploydataonthequestionsandanswerspostedtoStackOverfow,alongwithcharacteristicsofthepostingusers,fromtwocalendarperiodsthatcoverthesamespanofthecalendaryear.TetwosamplescoverOctober2021throughmid-Marchof2022,andOctober2022throughmid-Marchof2023.TesedatasetswereobtainedviatheStackExchangeDataExplorer,whichprovidesdownloadable,anonymizeddataonactivityindiferentStackExchangecommunities.Further,weemploydatafromsubred-,whichtracksaggregatedailycountsofpostingvolumestoeachsub-Reddit.Ourdatasourcesdonotincludeanypersonaluserinformation,andnoneofouranalysesmakeuseofanypersonaluserinformation. WefrstexaminedtheefectthatChatGPT’sreleaseonNovember30thof2022hadonwebtrafcarrivingatStackOverfow,leveragingthedailywebtrafcdataset.Tesample,sourcedfromSimilarWeb,includesdailytrafctothetop1000websites.Weemployavariantofthesyntheticcontrolmethod

13

,namelySyntheticControlUsingLASSO,orSCUL

14

.Takingthetimeseriesofwebvisitstoastreated,themethodidenti-fes,viaLASSO

15

,alinear,weightedcombinationofcandidatecontrolseries(websites)thatyieldsanaccuratepredictionoftrafctopriortoChatGPT’srelease.TeresultinglinearcombinationisthenusedtoimputeacounterfactualestimateoftrafcatintheperiodfollowingChatGPT’srelease,refectingpredictionsofwebtrafcvolumesthatwouldhavebeenobservedintheabsenceofChatGPT.

Second,weexaminedChatGPT’sefectsonthevolumeofquestionsbeingpostedtoStackOverfow.Weidentifedthetop50mostpopulartopictagsassociatedwithquestionsonStackOverfowduringourperiodofstudy,calculatingthedailycountofquestionsincludingeachtagoveratimewindowbracketingthedateofChatGPT’srelease.WethenfollowedtheapproachofRefs.

16

,

17

,constructingthesamesetoftopicpanelsforthesamecalendarperiod,oneyearprior,toserveasourcontrolwithinadiference-in-diferencesdesign,toestimateanaveragetreatmentefect,andtoenableevaluationbothoftheparalleltrendsassumption(whichissupportedbytheabsenceofsignifcantpre-treatmentdiferences)andtreatmentefectdynamics

18

.FigureS1inthesupplementprovidesavisualexplanationofourresearchdesign.

Tird,weconsideredwhethertheefectsmightdiferacrossonlineknowledgecommunities,dependingonthedegreetowhichacommunityisfocusedstrictlyoninformationexchange.Tatis,weconsideredthepotentialmitigatingefectofsocialfabric,i.e.socialbondsandconnections,asabuferagainstLLMsnegativeefectsonconnectionwithhumanpeers.TelogicforthistestisthatLLMs,despitebeingcapableofhigh-qualityinforma-tionprovisionaroundmanytopics,areoflessclearvalueasapuresubstituteforhumansocialconnections

19

.WethuscontrastedouraverageefectestimatesfromStackOverfowwithefectestimatesobtainedusingpanelsofdailypostingvolumesfromanalogoussub-communitiesatReddit(sub-Reddits),focusedonthesamesetsoftopics.RedditisausefulpointofcomparisonbecauseithasbeenwelldocumentedthatRedditdevelopercommunitiesarerelativelymoresocialandcommunalthanStackOverfow

20

,

21

.Wealsoexploredheterogene-ityintheStackOverfowefectsacrosstopics,repeatingourdiference-in-diferencesregressionforeachStackOverfowtopicandassociatedsub-reddit.

Lastly,weexploredshifsintheaveragecharacteristicsofusersandquestionsatStackOverfowfollowingChatGPT’srelease,specifcallyintermsofthepostingusers’accounttenure,indays,and,relatedly,theaveragecomplexityofpostedquestions.ItisreasonabletoexpectthattheindividualsmostlikelytorelyonChatGPTarejunior,newermembersofthecommunity,astheseindividualslikelyhavelesssocialattachmenttothecom-munity,andtheyarelikelytoaskrelativelysimplerquestions,whichChatGPTisbetterabletoaddress.Inturn,itisreasonabletoexpectthatthequestionsthatfailtobepostedarethosethatwouldhavebeenrelativelysimpler.Wetestedthesepossibilitiesintwoways,consideringquestion-leveldatafromStackOverfow.WebeganbyestimatingtheefectofChatGPT’sreleaseontheaveragetenure(indays)ofpostingusers’accounts.Next,weestimatedasimilarmodel,consideringtheaveragefrequencyof‘long’words(wordswith6ormorecharacters)withinpostedquestions,asaproxyforcomplexity.

Results

OverallimpactofLLMsoncommunityengagement

Figure

1

AdepictstheactualdailywebtrafctoStackOverfow(blue)alongsideourestimatesofthetrafcthatStackOverfowwouldhaveexperiencedintheabsenceofChatGPT’srelease(red).TeSyntheticControlesti-matescloselymirrorthetruetimeseriespriortoChatGPT’srelease,supportingtheirvalidityasacounterfactualforwhatwouldhaveoccurredpost.Figure

1

Bpresentsthediferencebetweenthesetimeseries.WeestimatethatStackOverfow’sdailywebtrafchasdeclinedbyapproximately1millionindividualsperday,equivalenttoapproximately12%ofthesite’sdailywebtrafcjustpriortoChatGPT’srelease.

LLMs’efectonusercontentproduction

Ourdiference-in-diferencesestimationsemployingdataonpostingactivityatStackOverfowrevealedthatquestionpostingvolumesper-topiconStackOverfowhavedeclinedmarkedlysinceChatGPT’srelease(Fig.

2

A).

ScientifcReports|(2024)14:10413|

/10.1038/s41598-024-61221-0

natureportfolio3

www./scientificreports/

Figure1.Syntheticcontrolestimatesofdeclineindailywebtrafctostackoverfow.Estimatesareobtained

viasyntheticcontrolusingLASSO(SCUL),basedondailywebtrafcestimatesaccordingtoSimilarWebforthe1000mostpopularwebsitesontheinternet.Panel(A)depictstheactualwebtrafcvolumes(inblue)recordedbySimilarWebalongsidetheSyntheticControl(inred).Panel(B)depictsthediferencebetweenthetwoseries,refectingtheestimatedcausalefectofChatGPT.

Figure2.EstimatedefectsofChatGPTonuseractivityatstackoverfowandreddit.Estimatesareobtainedviadiference-in-diferencesregression,comparingcontentpostingvolumesoveraperiodbracketingtherelease

ofChatGPT(onNovember30th,2022)withawindowofequallengthobservedonecalendaryearprior.Panel

(A)depictsefectsovertime(byweek)onStackOverfowquestionvolumespertopic.Panel(B)depictsefectsonRedditpostingvolumes,persub-reddit,forsub-redditsdealingwithanoverlappingsetoftopics.Teshadedarearepresents95%confdenceintervals.

TisresultreinforcestheideathatLLMsarereplacingonlinecommunitiesasasourceofknowledgeformanyusers.RepeatingthesameanalysisusingRedditdata,weobservednoevidencethatChatGPThashadanyefectsonuserengagementatReddit(Fig.

2

B).WereplicatetheseresultsinFig.S2ofthesupplementemployingthematrixcompletionestimatorofRef.

22

.

HeterogeneityinChatGPT’sefectonstackoverfowpostingvolumesbytopic

WeobservedagreatdealofheterogeneityacrossStackOverfowtopics,yetconsistentlynullresultsacrosssub-reddits(Fig.

3

).Ourestimatesthusindicate,again,thatRedditdevelopercommunitieshavebeenlargelyunafectedbyChatGPT’srelease.OurStackOverfowresultsfurtherindicatethatthemostsubstantiallyafectedtopicsarethosemostheavilytiedtoconcrete,self-containedsofwarecodingactivities.Tatis,themostheav-ilyafectedtopicsarealsothosewherewemightanticipatethatChatGPTwouldperformquitewell,duetotheprevalenceofaccessibletrainingdata.

ScientifcReports|(2024)14:10413|

/10.1038/s41598-024-61221-0

natureportfolio4

www./scientificreports/

Figure3.Topic-specifcefectsofChatGPTonstackoverfowandreddit.Estimatesareobtainedviadiference-in-diferencesregression,pertopic.Tefguredepictsefectestimatesforeachstackoverfowtopic(inorange)with95%confdenceintervalsandestimatesforeachsub-reddit(inred),whereavailable.Notethatdataon

sub-redditpostingvolumeswasnotavailableforthreesub-redditcommunities:javascript,jQuery,andDjango.OtherRedditestimatesareomittedduetothelackofaclearlyanalogoussub-redditaddressingthattopic.

Forexample,Python,CSS,Flutter,ReactJS,Django,SQL,Arrays,andPandasareallreferencestoprogram-minglanguages,specifcprogramminglibraries,ordatatypesandstructuresthatonemightencounterwhileworkingwithaprogramminglanguage.Incontrast,relativelyunafectedtagsappearmorelikelytorelatetotopicsinvolvingcomplextasks,requiringnotonlyappropriatesyntaxbutalsocontextualinformationthatwouldofenhavebeenoutsideofthescopeofChatGPT’strainingdata.Forexample,SpringandSpring-bootareJava-basedframeworksforenterprisesolutions,ofeninvolvingback-end(server-side)programminglogicwithprivateenterpriseknowledgebasesandsofwareinfrastructures.Questionsrelatedtothesetopicsareintuitivequestionsforwhichanautomated(i.e.cut-and-paste)solutionwouldbelessstraightforward,andlesslikelytoappearinthetextualtrainingdataavailablefortrainingtheLLM.AdditionalexampleshereincludethetagsrelatedtoAmazonWebServices,Firebase,Docker,SQLServer,andMicrosofAzure.

Toevaluatethispossibleexplanationmoredirectly,wecollecteddataonthevolumeofactiveGitHubreposi-toriesmakinguseofeachlanguageorframework,aswellasthenumberofindividualssubscribedtosub-redditsfocusedoneachlanguageorframework.WethenplottedascaledmeasureofeachvalueatoptheobservedefectsizesandobtainedFig.

4

.Tefgureindicatesaroughcorrelationbetweenavailablepublicsourcesoftrainingdataandourefectsizes.

ChatGPT’sefectonaverageuseraccountageandquestioncomplexity

Figure

5

depictsthechangeinaveragepostingusers’accounttenure,makingclearthat,uponChatGPT’srelease,asystematicrisebegantotakeplace,suchthatuserswereincreasinglylikelytobemoreestablished,olderaccounts.TeimplicationofthisresultisthatneweruseraccountsbecamesystematicallylesslikelytoparticipateintheStackOverfowcommunityaferChatGPTbecameavailable.Figure

6

depictstheefects,indicatingthatques-tionsexhibitedasystematicriseincomplexityfollowingthereleaseofChatGPT.

Tesefndings,consistentwiththeideathatmorejuniorandlessexperiencedusersbegantoexitmightbecauseforconcernifasimilardynamicisplayingoutinmoreformalorganizationandworkcontexts.Tisisbecausejuniorindividualsmaystandtolosethemostfromdeclinesinpeerinteraction—theseindividualstypi-callyaremoremarginalmembersoforganizationsandthushavelessrobustnetworksandhavethemosttoloseintermsofopportunitiesforcareeradvancement

23

.Further,theseindividualsmaybeleastcapableofrecognizingmistakesintheoutputofLLMs,whicharewellknowntoengageinhallucination,providing‘confdentlywrong’answerstouserqueries

6

.Indeed,recentworkobservesthatnon-expertsfacethegreatestdifcultydeterminingwhethertheinformationtheyhaveobtainedfromanLLMiscorrect

24

.

ScientifcReports|(2024)14:10413|

/10.1038/s41598-024-61221-0

natureportfolio5

www./scientificreports/

Figure4.Topic-specifcefectsofChatGPTonstackoverfow(blackpointswith95%confdenceintervals)

withNumberofGithubrepositories(purple)andsub-redditsubscribers(red)overlaid.Weobservearough

correlationbetweenthevolumeofGithubrepositoriesmakinguseofagivenlanguageorframework,thelevelofactivityinassociatedsub-redditcommunities,andthemagnitudeofefectsizes.TisassociatesuggestsefectsarelargerfortopicswheremorepublicdatawasavailabletotraintheLLM.

Figure5.EfectofChatGPTreleaseontheaveragetenure(indays)ofuseraccountspostingquestionsto

stackoverfow.ShortlyaferChatGPT’srelease,weseeasystematicriseintheaverageage(indays)forthe

useraccountspostingquestionstoStackOverfow.Weseethataverageaccountagerisessystematicallyonce

ChatGPTisreleased,consistentwithneweraccountssystematicallyreducingtheirparticipationandexitingthecommunity.

Discussion

WehaveshownthatChatGPTsreleasewasassociatedwithadiscontinuousdeclineinwebtrafcandquestionpostingvolumesatStockOverfow.Tisresultisconsistentwiththeideathatmanyindividualsarenowrely-ingonLLMsforknowledgeacquisitioninlieuofhumanpeersinonlineknowledgecommunities.OurresultsdemonstratethattheseefectsmanifestedforStackOverfow,yetnotforRedditdevelopercommunities.

Further,wehaveshownthattheseefectsweremorepronouncedforverypopulartopicsascomparedtolesspopulartopics,andevidencesuggeststhatthisheterogeneityderivedfromthevolumeoftrainingdataavail-ableforLLMtrainingpriortoChatGPTsrelease.Finally,ourresultsdemonstratethatChatGPT’sreleasewas

ScientifcReports|(2024)14:10413|

/10.1038/s41598-024-61221-0

natureportfolio6

www./scientificreports/

Figure6.EfectofChatGPT’sreleaseontheaveragecomplexityofquestionspostedtoStackOverfow,

refectedbytheaveragefrequencyof‘long’words(wordswith6ormorecharacters).ShortlyaferChatGPT’srelease,weseeasystematicriseintheaveragecomplexityofquestions.Tisresultisagainconsistentwiththeideathatneweraccountssystematicallyreducedtheirparticipationandexitedthecommunity.

associatedwithasignifcance,discontinuousincreaseintheaveragetenureofaccountsparticipatingonStackOverfow,andinthecomplexityofquestionsposted(asrefectedbytheprevalenceoflengthywordswithinquestions).Teseresultsareconsistentwiththeideathatthatnewer,lessexpertusersweremorelikelytobeginrelyingonChatGPTinlieuoftheonlineknowledgecommunity.

Ourfndingsbearseveralimportantimplicationsforthemanagementofonlineknowledgecommunities.Foronlinecommunities,ourfndingshighlighttheimportanceofsocialfabricasameansofensuringthesustain-abilityandsuccessofonlinecommunitiesintheageofgenerativeAI.OurfndingsthushighlightthatmanagersofonlineknowledgecommunitiescancombattheerodinginfuenceofLLMsbyenablingsocialization,asacomplementtopureinformationexchange.Ourfndingsalsohighlighthowcontentcharacteristicsandcom-munitymembershipcanshifbecauseofLLMs,observationsthatcaninformcommunitymanagerscontentmoderationstrategiesandtheiractivitiescenteredoncommunitygrowthandchurnprevention.

Beyondthepotentialconcernsaboutwhattheobserveddynamicsmayimplyforonlinecommunitiesandtheirmembers,ourfndingsalsoraiseimportantconcernsaboutthefutureofcontentproductioninonlinecommunities,whichbyallaccountshaveservedasakeysourceoftrainingdataformanyofthemostpopularLLMs,includingOpenAI’sGPT.Totheextentcontentproductiondeclinesintheseopencommunities,itwillreinforceconcernsthathavebeenraisedintheliteratureaboutlimitationsonthevolumeofdataavailableformodeltraining

25

.Ourfndingssuggestthatlong-termcontentlicensingagreementsthathaverecentlybeensignedbetweenLLMcreatorsandonlinecommunityoperatorsmaybeundermined.Iftheseissuesarelefunaddressed,thecontinuedadvancementofgenerativeAImodelsmaynecessitatethattheircreatorsidentifyalternativedatasources.

Conclusion

Ourworkisnotwithoutlimitations,someofwhichpresentopportunitiesforfutureresearch.First,forourresearchdesigntoyieldcausalinterpretations,wemustassumetheabsenceofconfoundedtreatments.Forexample,wereanotherlargeonlinecommunitytohaveemergedaroundthesametime,thepossibilityexiststhatitmayexplainthedeclineinparticipationatStackOverfow.Second,ourstudylacksanuancedanalysisofchangesincontentcharacteristics.Althoughwestudychangesinanswerqualityusingnetvotescores(seethesupplement),ourmeasuresmayrefectchangesinotheraspectsunrelatedtoinformationquality.Similarly,althoughwestudychangesinquestioncomplexity,ourmeasureofcomplexityistiedtowordlength.Futureworkcanthusrevisitthesequestionsemployingavarietyofothermeasuresofqualityandcomplexity.

Tird,althoughwehaveshownadeclineinparticipationatStackOverfow,weareunabletospeaktowhetherthesamedynamicisplayingoutinotherorganizationalsettings,e.g.workplaces.Itisalsoimportanttorecognizethatthecontextofouranalysesmaybeunique.TotheextentStackOverfowandRedditdevelopercommunitiesmightnotberepresentativeofdevelopercommunitiesmorebroadly,thegeneralizabilityoftheseresultswouldbeconstrained.Relatedly,itispossiblethattheresultsweobserveareuniquetoknowledgecommunitiesthatfocusonsofwaredevelopmentandinformationtechnology.Tedynamicsofcontentproductionmaydifermarkedlyinotherknowledgedomains.Finally,ourworkdemonstratesefectsoverarelativelyshortperiodoftime(severalmonths).Itispossiblethatthelonger-rundynamicsoftheobservedefectsmayshif.Giventhesepoints,futureworkcanandshouldendeavortoexplorethegeneralizabilityofourfndingstoothercommunities,andfutureworkshouldexaminethelonger-runefectsofgenerativeAItechnologiesoncommunityparticipa-tionandknowledgesharing.

ScientifcReports|(2024)14:10413|

/10.1038/s41598-024-61221-0

natureportfolio7

www./scientificreports/

WeanticipatethatourstudywillinspiremoresophisticatedanalysesoftheefectsthatgenerativeAItechnolo-gies,includingLLMs,butalsogenerativeimage,audio,andvideomodels,mayhaveonpatternsofknowledgesharingandcollaborationwithinorganizationsandsocietymorebroadly.Suchworkiscruciallyneeded,tobetterunderstandthenuancesofwhereandwhenindividualsmayrelyonhumanpeersversusGenerativeAItools,andthedesirableandundesirableconsequencesfororganizationsandsociety,suchthatwecanbegintoplanforandmanagethisnewdynamic.

Dataavailability

DataonStackOverfowusers,questions,andanswerswasobtainedviatheStackExchangeDataExplorerat

/stackoverfow/query/new.

Dataonsub-redditpostingvolumeswasobtainedfrom

.SimilarWebdailywebtrafcdataisnotavailableforpublicdissemination,thoughitisavailableforpurchasefrom

https://deweydata.io

.StackOverfowdata,RedditdataandanalysisscriptsareavailableinapublicrepositoryattheOSF:

https://osf.io/qs6b3/

.

Received:23October2023;Accepted:2May2024

publishedonline:06May2024

References

1.Noy,S.&Zhang,W.Experimentalevidenceontheproductivityefectsofgenerativeartifcialintelligence.Science

/

10.2139/ssrn.4375283

(2023).

2.Peng,S.,Kalliamvakou,E.,Cihon,P.,Demirer,M.TeimpactofAIondeveloperproductivity:EvidencefromGithubcopilot.Preprintat

https://arX/2302.06590

(2023).

3.Dell-Acqua,F.etal.Navigatingthejaggedtechnologicalfrontier:FieldexperimentalevidenceoftheefectsofAIonknowledgeworkerproductivityandquality.HarvardBusinessSchoolWorkingPaper,no.24-013(2023).

4.Hwang,E.H.,Singh,P.V.&Argote,L.Knowledgesharinginonlinecommunities:Learningtocrossgeographicandhierarchicalboundaries.Organ.Sci.26(6),1593–1611(2015).

5.Hwang,E.H.&Krackhardt,D.Onlineknowledgecommunities:Breakingorsustainingknowledgesilos?.Prod.Oper.Manag.29(1),138–155(2020).

6.Bang,Y.etal.Amultitask,multilingual,multimodalevaluationofChatGPTonreasoning,hallucination,andinteractivity.InProc.ofthe13thInternationalJ