




版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
UsersFavorLLM-GeneratedContent—UntilTheyKnowIt’sAI
PetrParshakov
HSEUniversityandMoscowSchoolof
IuliiaNaidenova
HSEUniversity
SofiaPaklina
HSEUniversity
NikitaMatkin
HSEUniversity
CornelNesseler
UniversityofStavanger
ManagementSKOLKOVO
Abstract
arXiv:2503.16458v1[cs.HC]23Feb2025
Inthispaper,weinvestigatehowindividualsevaluatehumanandlargelanguemodelsgen-eratedresponsestopopularquestionswhenthesourceofthecontentiseitherconcealedordisclosed.Throughacontrolledfieldex-periment,participantswerepresentedwithasetofquestions,eachaccompaniedbyare-sponsegeneratedbyeitherahumanoranAI.Inarandomizeddesign,halfoftheparticipantswereinformedoftheresponse’soriginwhiletheotherhalfremainedunaware.Ourfindingsindicatethat,overall,participantstendtopre-ferAI-generatedresponses.However,whentheAIoriginisrevealed,thispreferencedimin-ishessignificantly,suggestingthatevaluativejudgmentsareinfluencedbythedisclosureoftheresponse’sprovenanceratherthansolelybyitsquality.TheseresultsunderscoreabiasagainstAI-generatedcontent,highlightingthesocietalchallengeofimprovingtheperceptionofAIworkincontextswherequalityassess-mentsshouldbeparamount.
1Introduction
Therapidevolutionoflargelanguagemodels(LLMs)overrecentyearshasfundamentallytrans-formedthelandscapeoftextgenerationandhuman-computerinteraction.Thesemodelsarenowin-tegraltovariousapplications,rangingfromcus-tomerserviceandcontentcreationtopersuasivemessagingandpersonalizedcommunication.Re-centadvancementsinLLMshavedemonstratedsignificantpotentialtoenhancelaborproductiv-ityacrossvariousbusinessapplications,includingcustomercommunicationandcontentcreation(
Ay-
ersetal.,
2023;
Brynjolfssonetal.,
2025;
Zhang
andGosline,
2023
).LLMsarealsousefulincreat-ingpersuasivepublicmessages(
Karinshaketal.,
2023
)ormaintainingpersonalizedin-depthcon-versationstochangeindividual’sbeliefs(
Costello
etal.,
2024
).Understandinghowindividualsper-ceiveAI-generatedresponsesisessentialforen-
suringtheeffectivenessandacceptanceofthesetechnologiesinbusinessandsocialsettings.
AsLLM-generatedcontentincreasinglymirrorshuman-authoredtextinfluencyandcoherence,thechallengeofdistinguishingbetweenthetwosourceshasbecomemorecomplex.Priorresearchhashighlightedmeasurabledifferencesinlinguis-ticfeaturesandsentimentexpressionbetweenhu-manandAItexts,whilealsonotingthattheper-ceivedqualityofresponsesmayshiftwhentheoriginofthecontentisdisclosed.Thisdynamicisparticularlyimportantinsettingswheretrustandcredibilityareparamount,suchasincustomerin-teractionsorpubliccommunications.Manystud-iesfocusonspecificareas(e.g.,healthorpub-licannouncements)andneglecttheimportanceofgeneral-interesttexts.
OurstudyinvestigateshowresponsestopopularquestionsonplatformssuchasQuoraandStackOverflowareperceivediftheresponsecomesfromahumanorfromanLLM.Weexaminetheseper-ceptionsacrossavarietyofdomains,includingPhysicalSciences,LifeSciences,HealthSciences,SocialSciences,andHumanities,usingadiversesetofpopularquestions.Incorporatingrespon-dentcharacteristicssuchasgender,age,educa-tionalbackground,andprogrammingskills,ourresearchaimstoofferanuancedunderstandingofthefactorsthatdrivetrustandpreferenceincontentgeneration.
2Literaturereview
2.1HumanandLLMgeneratedtexts
Recentresearchhassystematicallyexaminedthedistinctionsbetweenhuman-generatedandAI-generatedtexts,revealingmeasurabledifferencesinsentencestructure,emotionexpression,andotherlinguisticfeatures(
Muñoz-Ortizetal.,
2024;
NituandDascalu,
2024
).Earlystudiesdemon-stratedthattraditionalmachinelearningclassifiers
couldeffectivelydifferentiatebetweenhumanandAI-generatedcontent;however,theadventofad-vancedlargelanguagemodels(LLMs)hassignifi-cantlycomplicatedthistask(
Hayawietal.,
2024
).Infact,AI-generatedtextshaveattimesmatchedorevenexceededhuman-writtentextsinspecificapplications,suchaspersuasivemessaging(
Karin-
shaketal.,
2023
)andprovidingwritingfeedbackineducation(
Escalanteetal.,
2023
).Theincreas-ingsophisticationofLLMshasledtoasignifi-cantconvergencebetweenAI-generatedtextandhuman-writtencontent,renderingthedistinctionbetweenthetwoincreasinglychallenging(
Ollivier
etal.,
2023;
Hayawietal.,
2024
).Asthesemodelsevolve,theyproducetextthatnotonlymimicshu-manwritingstylesbutalsoadherestothenuancesoflanguage,context,andcoherencethatcharacter-izeauthentichumancommunication.EvaluationsofexistingLLM-generatedtextdetectorshavere-portedinconsistencies(
Weber-Wulffetal.,
2023
)andhighfalsepositiverateswhenthesesystemsareappliedtohuman-authoredtexts(
Elkhatatetal.,
2023
).Inaddition,theobjectivityofAI-generatedcontentisalsoquestionable.TheliteraturehasrevealedinherentbiaseswithinoutputsproducedbyLLMs.Studieshavedocumentedsignificantgenderandracialbiases,notablyindepictionsofhealthcareprofessionalsandsurgeons,wheremalerepresentationsarefrequentlyfavored(
Menz
etal.,
2024;
Ceviketal.,
2024
).Politicalbiashasalsobeenobserved,withcertainplatformssuchasChatGPTexhibitingatendencytowardleft-leaningperspectives(
Motokietal.,
2024;
Rozado,
2024
).Moreover,LLMstendtomanifesthuman-likecon-tentbiases,asdemonstratedbytransmissionex-periments(
AcerbiandStubbersfield,
2023
)andlinguisticanalyses(
Fangetal.,
2024
).
2.2PerceptionofLLM-generatedcontent
ComparativestudiesrevealthathumanevaluatorsoftenstruggletoreliablydifferentiatebetweenAI-generatedandhuman-authoredcontent(
Boutadjine
etal.,
2024
).
ZhangandGosline
(2023
)demon-stratedthatgenerativeandaugmentedAIcontentisfrequentlyperceivedassuperiortothatproducedbyhumanexperts,evenwhenhumansutilizeAItools.However,disclosingthesourceofcontentnarrowstheperceivedqualitygap,suggestingabiasfavor-inghumancontributionsoverAI.Participantsratedcontentmorefavorablywhenattributedtohumanexperts,whereasawarenessofAIinvolvementhad
minimalimpactonperceptions.
Ayersetal.
(2023
)examinedtheabilityofanAIchatbot(ChatGPT)todeliverqualityandempatheticresponsestopatientquestionscomparedtophysicians.Theirfindingsrevealedthatchatbotresponseswerepreferredinthemajorityofevaluations,ratedhigherinqual-ity,anddeemedmoreempatheticthanthoseofphysicians.Notably,chatbot-generatedtextswerealsosignificantlylongerthanphysicianresponses.
Karinshaketal.
(2023
)highlightedthatlargelan-guagemodels(LLMs),particularlyGPT-3,canproducehigh-qualitypersuasivecontent;however,individualstendtopreferpublichealthmessagesoriginatingfromhumaninstitutionsratherthanAIsources.Somestudieshighlightthenuancedper-ceptionsandpreferencessurroundingAI-generatedcontentacrossvariousdomains.Forexample,
Es-
calanteetal.
(2023
)comparedhumantutorfeed-backwithAI-generatedfeedbackineducationalsettings,revealingmixedresults.Whileface-to-faceinteractionswithtutorsenhancedstudenten-gagement,AI-generatedfeedbackwasfavoredforitsclarityandspecificity.Theresearchby
Chen
etal.
(2024
)demonstratesthatconsumerspreferAI-generatedadswithagenticappeals,whilefa-voringhuman-createdadswithcommunalappeals.Thesefindingsunderscorethatcontextualfactorsandtheawarenessoftextoriginplayacriticalroleinshapinguserpreferences.
3Methodologyanddesign
3.1Datacollection
Thisstudyaimstoevaluateresponsesthatareofinteresttoabroadaudience.Toachievethis,theanalysisfocusesonpopularquestionsfromQuora,aplatformwhereuserscanpostquestionsandpro-videanswers,withthemostpopularresponsesprominentlydisplayed.Additionally,forthedo-mainofPhysicalSciencesandEngineering,ques-tionsfromStackOverflow,awidelyusedplatformforprogramming-relatedqueries,areincorporated.Theselectedquestionsarebroadlycategorizedintofivescientificareas:PhysicalSciencesandEngi-neering,LifeSciences,HealthSciences,SocialSciences,andHumanities.Thiscategorizationen-suresbalancedrepresentationacrossdomainsandpreventsthedominanceofanysinglearea.Whilethequestionsarenotstrictlyscientific,theyarede-signedtoappealtoageneralaudiencewithdiversebackgrounds.Examplesofsuchquestionsinclude,"WhatstartedWWII?","Whatwasthebestteamin
thehistoryofsports?",and"Whatroledoestherapyplayintreatinganxiety?".
FromQuoraandStackOverflow,fivequestionsareselectedforeachofthefivescientificareas,re-sultinginatotalof25questions.Thesequestionsarethenposedtofourprominentlargelanguagemodels(LLMs)—ChatGPT,Claude,Gemini,andLlama—selectedfortheirsuperiorperformanceintextgenerationatthetimeofthestudy.Eachques-tiongeneratesfiveresponses:fourfromtheLLMsandonefromahumanrespondent.Thecompletelistofquestions,organizedbyfield,ispresentedintheAppendix
A.1.
Theaveragelengthofhumanresponsesis1,515charactersor265words,whileresponsesfromtheLLMsvarybetween1,854and2,265characters.Detailedsummarystatisticsre-gardingresponselengths,measuredinbothchar-actersandwords,areprovidedintheAppendix
A.2.
Wecreatedasurveyinwhichparticipantsarerandomlypresentedwith5questions,eachaccom-paniedbytworesponses:onegeneratedbyanLLMandonebyahuman.Participantsarealsoaskedtoprovidedemographicinformation,includingage,gender,andeducationalbackground.Priortocom-mencingthesurvey,participantsareinformedthattheywillberequiredtochoosebetweentwore-sponsesforeachquestion.Importantly,arandompartoftheparticipantisnotinformedabouttheoriginoftheresponses(i.e.,whethertheyaregen-eratedbyanLLMorahuman)tomitigatepotentialbiases,suchastheHawthorneeffect(
Sedgwickand
Greenwood,
2015
).TheexperimentisregisteredontheSocialScienceRegistry
1.
Allparticipantswereinformedabouthowtheirdatawouldbeusedinthisstudy,andexplicitconsentwasobtainedpriortoparticipation.Wecontactedourinstitu-tion’sEthicalCommitteeandwereinformedthat,aswedonotstoreanypersonaldata,additionalethicalapprovalwasnotrequired.
Thesurveyisimplementedintwoformats:aTelegrambotandawebapplicationdevelopedus-ingStreamlit.Bothplatformsutilizethesamealgo-rithmicstructureandshareanidenticaldatabaseofquestions.Intotal,thestudyinvolves993partici-pants,with507respondentsusingtheTelegrambotand486thewebapplication.Fromthisinitialpool,130respondentsareexcludedfromthefinalsampleduetotheirfailuretoprovideanyresponses.
ThesurveyisdistributedthroughP,a
1SocialScienceRegistry
platformdesignedtomatchsurveyswithappropri-aterespondents.Participantsarecompensatedatarateof9GBP(approximately11USD)perhour.Thefinalsampleconsistsof846participants,whotakeanaverageof6.6minutestocompletethesur-vey.Thedemographiccompositionofthesampleis35%maleand45%female,withanaverageageof30years.Additionaldetailedsummarystatistics,includingbreakdownsbyawarenessgroups,areprovidedintheAppendix
B.1
andtheAppendix
B.2.
3.2Estimationstrategy
Toexaminewhetherindividualspreferhuman-orAI-generatedanswersbasedontheirawarenessofthesource,weemployedalogisticregressionanal-ysis.Specifically,werestructuredourdatasetsuchthattheunitofobservationisarespondent-answerpair.Wethenintroducedadummyvariable,hu-man,whichtakesthevalueof1iftherespondentselectedthehuman-generatedanswer.Thisvari-ableservesasthedependentvariable.
Amongtheindependentvariables,theprimaryvari-ableofinterestisaware,whichindicateswhetherarespondentcanseethesourceoftheanswer(hu-manorAImodel).Thecoefficientassociatedwithawarereflectswhetherknowledgeoftheanswer’ssourceinfluencestheprobabilityofselectingahuman-generatedresponse.Additionally,wein-cludedrespondentcharacteristicssuchasgender,age,levelofeducation,fieldofstudy,andprogram-mingskills.Wealsoaccountedforthedurationofthesurveyandthefieldofthequestion.Further-more,thebasicmodelspecificationispresentedbelow:
Pr(humani=1)=β0+β1awarei+β2femalei
+β3agei+β4durationi
+β5education_leveli
+β6education_fieldi+
+β7programming_skillsi
+β8modeli+β9question_fieldi
(1)
Toensurerobustness,weclusteredthestandarderrorsbygender.Toassessthevaryingeffectsofawarevariableacrossdifferentrespondentandquestioncharacteristics,wealsointroducedinter-actiontermswithgender,programmingskills,andquestionfield.
Figure1:Distributionofchosenanswersbygenderandsourceawareness
4Empiricalresults
4.1Exploratoryanalysis
Toillustratethepossiblefactorsthatmightinflu-encethechoicesbetweenanswersgeneratedbyhumansandAImodels,weperformedagraphi-calanalysis.Themostnoticeableandinterestingresultswerefoundinthedifferencesinchoicesacrossgenderandprogrammingskills.
Figure
1
presentsthedistributionofchosenan-swersbasedonwhetherparticipantswereawareofthesourceoftheresponseandtheirgender.Acrossallconditions,aconsistentpatternemerges,indicat-ingahigherpreferenceforAI-generatedresponsescomparedtohuman-generatedones.Aninterestingresultisthatwhenrespondentsareawareofthesourceoftheanswers,theyslightlytendtochoosehuman-generatedanswers,especiallyamongmalerespondents.
Figure
2
illustratesthedistributionofselectedanswersbasedonparticipants’programmingskillsandtheirawarenessofthesourceoftheresponses.TheresultsalsoindicateaconsistentpreferenceforAI-generatedresponsesacrossallconditions.Whenparticipantswereinformedabouttheoriginoftheanswers,thepreferenceforAI-generatedresponsesremaineddominant,althoughaslightincreaseintheselectionofhuman-generatedre-sponsescanbeobserved,particularlyamongthosewithprogrammingskills.However,thisdifferenceappearsmarginalandwouldrequiremorein-depth
Figure2:Distributionofchosenanswersbyprogram-mingskillsandsourceawareness
hypothesistesting.
Takentogether,thesefindingsindicatethatpriortechnicalknowledgedoesnotsignificantlyimpactthelikelihoodofselectingAI-generatedresponses.Evenwhenparticipantsaremadeawareofthesource,thepreferenceforAI-generatedanswersremainsstrong,reinforcingthenotionthatsuchresponsesareperceivedasequallyormorereliablecomparedtohuman-generatedalternatives.
4.2Regressionanalysis
Table
1
presentstheregressionresultsexamin-ingindividuals’preferencesforhuman-generatedanswersbasedontheirawarenessoftheanswersource.Thedependentvariableinallmodelsistheselectionofahuman-generatedresponse.Controlvariablesforeducationlevel,fieldofstudy,pro-grammingskills,AImodel,andquestionfieldareincludedinallmodels.
Acrossallspecifications,awarenessofthean-swersourcehasasignificantimpactonindividuals’preferences.InModels(1),(2),and(4),thein-dividualcoefficientforawarenessispositiveandstatisticallysignificant,suggestingthatwhenre-spondentsknowwhetherananswerisgeneratedbyahumanoranAI,theyaremorelikelytopreferhuman-generatedresponses.
Thevariablefemaleisconsistentlypositiveandsignificantacrossallmodels,indicatingthatfe-malerespondentsaremorelikelytopreferhuman-generatedresponsescomparedtomalerespondents.
Dependentvariable:
Aigeneratedanswer(=0),humangeneratedanswer(=1)
(1)(2)(3)(4)
AwareofAI-answer
0.109**
0.146***
—0.112
0.206***
(0.051)
(0.017)
(0.071)
(0.024)
Female
0.047***
0.090***
0.046***
0.047***
(0.015)
(0.024)
(0.013)
(0.015)
Age
—0.012**
—0.012**
—0.012**
—0.012**
(0.005)
(0.005)
(0.005)
(0.005)
Duration
0.0005
0.001
0.002
0.0004
(0.003)
(0.003)
(0.004)
(0.003)
AwareofAI-answer·Humanities
—0.111***
(0.017)
AwareofAI-answer·LifeSciences
—0.130**
(0.061)
AwareofAI-answer·PhysicalSciencesandEngineering
—0.068
(0.244)
AwareofAI-answer·SocialSciences
—0.184
(0.125)
AwareofAI-answer·Female
—0.089***
(0.023)
AwareofAI-answer·Programmingskills
0.322**
(0.126)
EducationLevel
Yes
Yes
Yes
Yes
EducationField
Yes
Yes
Yes
Yes
ProgrammingSkills
Yes
Yes
Yes
Yes
AIModel
Yes
Yes
Yes
Yes
QuestionField
Yes
Yes
Yes
Yes
Observations
3,206
3,206
3,206
3,206
LogLikelihood
—1,951.209
—1,951.050
—1,949.411
—1,950.902
Note:*p<0.1;**p<0.05;***p<0.01
Table1:Logisticregressionresults
Agehasasmallbutsignificantnegativeeffect,im-plyingthatolderindividualsareslightlylesslikelytopreferhuman-generatedanswers.
TheinteractioneffectsinModels(2)–(4)pro-videfurtherinsights.InModel(2),theinteractionbetweenawarenessofthesourceandfemaleisneg-ativeandsignificant,indicatingthattheeffectofknowingthesourcediffersbygender.Sincethejointeffectoffemaleanditsinteractionwithaware-nessofthesourceisnotstatisticallysignificant,weobservethepositiveeffectofknowingthesourceonlyformalerespondents.Specifically,formalerespondents,knowingwhetherananswerisgener-atedbyahumanorAIincreasestheprobabilityofselectingahuman-generatedresponse.
Model(3)includesaninteractionbetweenaware-nessofthesourceandprogrammingskills,whichispositiveandstatisticallysignificant,indicatingthatindividualswithprogrammingexpertisearemorelikelytopreferhuman-generatedresponseswhentheyareawareofthesource.
Model(4)explorestheroleofthefieldofthequestion.Theinteractiontermsshowthatforques-tionsinHumanitiesandLifeSciences,respon-dentsexhibitasignificantnegativepreferenceforhuman-generatedanswerswhentheyareawareofthesource.However,nosignificanteffectsareobservedforquestionsinPhysicalSciencesandEngineeringorSocialSciences.
Overall,thesefindingshighlightthatknowingthesourceofananswerinfluencesindividuals’pref-erences,withvariationsbygender,programmingexpertise,andthefieldofthequestion.
5Conclusion
Thisstudyhasinvestigatedtheinfluenceofsourceawarenessonindividuals’preferencesbetweenhuman-generatedandAI-generatedresponses.Ourempiricalresultsindicatethat,althoughrespon-dentsgenerallyshowatendencytofavorAI-generatedcontent,thedisclosureofananswer’sorigininducesameasurableshifttowardhuman-generatedresponses.Moreover,theextentofthisshiftiscontingentuponrespondentcharacteristicssuchasgenderandprogrammingexpertise,aswellasthecontextualdomainofthequestion.Thesefindingsnotonlyextendtheliteratureontheper-ceptualdifferencesbetweenhumanandAItexts
(Boutadjineetal.,
2024;
ZhangandGosline,
2023
)butalsocomplementpreviousstudiesthathavenotedtheimpactofsourcedisclosureoncontent
perception(
Ayersetal.,
2023;
Karinshaketal.,
2023
).
Interestingly,womendemonstrateastrongerpreferenceforhumanresponsescomparedtomen,potentiallyduetodifferencesinstyleortextlengthpreferences,asevidencedinAppendix
A.2.
Whilethesepreferencesremainconsistentregardlessofknowledgeabouttheanswer’sorigin,menexhibitasignificantresponsetothedisclosureofwhetherahumanoranLLMauthoredthetext.Thefind-ingsontext-lengthdifferencesbetweenhumanandAI-generatedtextsalignwiththosereportedby
Karinshaketal.
(2023
).
Incontrasttoearlierresearchthatreportedaper-vasivedifficultyamongevaluatorsindistinguishingbetweenAIandhuman-authoredtexts,ouranal-ysisrevealsthatsourceawarenesscanleadtoanuancedreallocationoftrustdependingonbothindividualanddisciplinaryfactors.WhilepreviousstudieshavelargelyemphasizedthechallengesindetectingAI-generatedcontent(
Weber-Wulffetal.,
2023;
Elkhatatetal.,
2023
),ourfindingssuggestthattransparencyregardingcontentoriginplaysacriticalroleinshapingevaluativejudgments,par-ticularlyamongsubgroupswithspecifictechnicalproficienciesanddemographicprofiles.
Theimplicationsoftheseresultsaresignificantforthedeploymentoflargelanguagemodelsinvarioussectors.AsAI-generatedcontentbecomesincreasinglyprevalentinbusinesscommunications,publicmessaging,andeducationalsettings,ensur-ingthatendusersunderstandtheoriginsofthecontentmayenhanceitsperceivedcredibilityandeffectiveness.Futureresearchshouldfurtherex-aminethelong-termeffectsofsourcetransparency,extendtheinquirytoencompassnon-textualmedia,andexplorethedynamicsofaugmentedhuman-AIcollaboration.Suchinvestigationswillbeessentialinrefiningourunderstandingofhowbesttointe-grateAItechnologiesintocontextswheretrustandauthenticityremainparamount.
6Limitations
Itisimportanttonotethatourresearchislimitedtotextualcontentanddoesnotencompassnon-textualformssuchasgraphicsoraudio,whicharewidelyutilizedincommunication.Additionally,maintain-inghumanoversightremainscrucialtoensurethatgenerativeAI-producedcontentisappropriateforsensitivetopicsandtopreventthedisseminationofunsuitablematerial.Furthermore,ourstudydoes
notexploreperceptionsofaugmentedhumanoraugmentedAI-generatedcontent,anemergingandpromisingareaofresearch(
Vaccaroetal.,
2024;
ZhangandGosline,
2023
).
Thisstudyfacesseverallimitationsthatwar-rantcarefulconsideration.First,ouranalysisisconfinedtotextualcontentanddoesnotencom-passnon-textualmediasuchasgraphics,audio,orvideo,whichareincreasinglyintegraltomoderncommunication.Thisfocusontextmaylimitthegeneralizabilityofourfindingsincontextswheremultimodalcontentplaysacriticalrole.
Anotherlimitationrelatestoourdatacollectionmethods.AlthoughweutilizedplatformssuchasQuoraandStackOverflowtoensureadiversesetofquestionsandrespondentbackgrounds,theon-linenatureofthesesourcesmayintroduceselectionbiases.Thesample,whilesubstantial,mightnotfullycapturethebroaderpopulation’scultural,de-mographic,ortechnologicaldiversity,whichcouldinfluencetheobservedpreferencesandperceptions.
Moreover,ourinvestigationintoevaluativejudg-mentsofcontentoriginreliesonself-reportedre-sponsesandobservablechoiceswithinanexperi-mentalframework.Despiteeffortstomitigatepo-tentialbiases,suchastheHawthorneeffect,thereremainsariskofdemandeffectsorsocialdesir-abilityinfluencingparticipants’selections.Thismethodologicalconstraintsuggeststhatfurtherstudiesemployingalternativedesignsoradditionalqualitativemeasuresmaybeneededtovalidateourconclusions.
Additionally,whileourworkcontributestounderstandingthedistinctionsbetweenhuman-generatedandAI-generatedtext,itdoesnotex-ploretheemergingdomainofaugmentedhumanoraugmentedAI-generatedcontent(
Vaccaroetal.,
2024;
ZhangandGosline,
2023
).Thedynamicsofcollaborativecontentcreation,wherehumancre-ativityisintertwinedwithAIassistance,presentapromisingavenueforfutureresearch.Similarly,theinfluenceofnon-textualelementsandthein-tegrationofmultimodalcommunicationonuserperceptionsremainopenquestions.
Finally,itisimportanttoacknowledgethattheplatformsandcontextsinwhichthedatawerecol-lectedmayimposetheirownnormsandbiasesonbothhumanandAI-generatedresponses.Thesecontextualfactorscouldaffectthestyle,substance,andevaluativejudgmentsofthecontent.Futureresearchshouldaimtoreplicateandextendthese
findingsacrossdifferentmediaandculturalsettingstoenhanceourunderstandingofhowsourceaware-nessinfluencescontentperceptioninabroaderspectrumofcommunicationenvironments.
7EthicsStatement
ThisworkadherestotheACLCodeofEthicsandcomplieswiththeethicalguidelinesestablishedforACL2023.Inconductingthisresearch,weensuredthatalldatacollectionprocesseswereperformedinaccordancewithethicalstandards,includingob-taininginformedconsentfromallparticipantsandanonymizingthecollecteddatatoprotectprivacy.Werecognizethatadvancesinlargelanguagemod-elsandtheincreasingprevalenceofAI-generatedcontenthavesignificantsocietalimplications.Ac-cordingly,ourstudyhasbeendesignedwithacom-mitmenttotransparencyandaccountability.Wehavecarefullyconsideredpotentialethicalrisks,includingthepropagationofbiasesandthemis-useofAI-generatedcontent,andhaveincorporatedmeasurestomitigatetheseconcerns.Ouraimistocontributetoabetterunderstandingofhowsourceawarenessinfluencescontentperception,whileun-derscoringtheimportanceofhumanoversightinthedeploymentofAItechnologies.
7.1PotentialRisks
Potentialrisksassociatedwiththisresearchpri-marilyarisefromthemisinterpretationandmis-useofourempiricalfindings.Inparticular,theobservedshiftinuserpreferences
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 企业电子买卖合同样本
- 农村集体场地租赁合同样本
- 农村房屋赠送合同标准文本
- 供热 bot 合同样本
- 团队创意思维的激发计划
- 公司办公维修合同样本
- 关于学校保洁合同范例
- 公司购买设备合同样本
- 中介公司用人合同样本
- 中央空调工程合同样本
- 基于单片机的停车场计数系统设计
- 氩弧焊培训课件-氩弧焊焊接技术培训
- 公司法务管理手册全部程序葵花宝典
- 弱电工程施工进度表(甘特图)
- 大航海时代4威力加强版各种宝物遗迹
- 幼儿园故事课件:《胸有成竹》
- GB/T 43200-2023机器人一体化关节性能及试验方法
- 五年级下册综合实践活动说课稿-合理消费 全国通用
- SIMTRADE外贸模拟实训报告
- 2022公务员录用体检操作手册(试行)
- 卖石斛怎么给顾客说:石斛卖的方法
评论
0/150
提交评论