Paper deep dive
How to Disclose? Strategic AI Disclosure in Crowdfunding
Ning Wang, Chen Liang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 88%
Last extracted: 7/21/2026, 2:58:50 AM
Summary
This study investigates the impact of mandatory AI disclosure on crowdfunding performance using Kickstarter data and supplementary experiments. It finds that mandatory disclosure significantly reduces funds raised and backer counts. However, this negative effect is moderated by disclosure strategies: high authenticity and explicitness mitigate the decline, while high AI involvement and excessive positive emotional tone exacerbate it. The mechanisms involve perceived creator competence and AI washing concerns.
Entities (10)
Relation Signals (9)
AI Disclosure → negativelyimpacts → Crowdfunding Performance
confidence 95% · mandatory AI disclosure significantly reduces crowdfunding performance: funds raised decline by 39.8% and backer counts by 23.9%
AI Involvement → amplifiesnegativeeffect → AI Disclosure
confidence 90% · Greater AI involvement amplifies the negative effects of AI disclosure
Explicitness → mitigatesnegativeeffect → AI Disclosure
confidence 90% · high explicitness (logos) ... mitigate the negative effect
Authenticity → mitigatesnegativeeffect → AI Disclosure
confidence 90% · high authenticity (ethos) mitigate the negative effect
AI Involvement → reduces → Perceived Creator Competence
confidence 88% · Greater AI involvement decreases pledge intention by reducing perceived creator competence
Emotional Tone → exacerbatesnegativeeffect → AI Disclosure
confidence 85% · excessive positive emotional tone ... backfires and exacerbates negative outcomes
Explicitness → increases → Perceived Creator Competence
confidence 85% · High explicitness increases pledge intention through ... enhancing perceived creator competence
Emotional Tone → increases → AI Washing Concerns
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:As artificial intelligence (AI) increasingly integrates into crowdfunding practices, strategic disclosure of AI involvement has become critical. Yet, empirical insights into how different disclosure strategies influence investor decisions remain limited. Drawing on signaling theory and Aristotle's rhetorical framework, we examine how mandatory AI disclosure affects crowdfunding performance and how substantive signals (degree of AI involvement) and rhetorical signals (logos/explicitness, ethos/authenticity, pathos/emotional tone) moderate these effects. Leveraging Kickstarter's mandatory AI disclosure policy as a natural experiment and four supplementary online experiments, we find that mandatory AI disclosure significantly reduces crowdfunding performance: funds raised decline by 39.8% and backer counts by 23.9% for AI-involved projects. However, this adverse effect is systematically moderated by disclosure strategy. Greater AI involvement amplifies the negative effects of AI disclosure, while high authenticity and high explicitness mitigate them. Interestingly, excessive positive emotional tone (a strategy creators might intuitively adopt to counteract AI skepticism) backfires and exacerbates negative outcomes. Supplementary randomized experiments identify two underlying mechanisms: perceived creator competence and AI washing concerns. Substantive signals primarily affect competence judgments, whereas rhetorical signals operate through varied pathways: either mediator alone or both in sequence. These findings provide theoretical and practical insights for entrepreneurs, platforms, and policymakers strategically managing AI transparency in high-stakes investment contexts.
Tags
Links
- Source: https://arxiv.org/abs/2602.15698v1
- Canonical: https://arxiv.org/abs/2602.15698v1
Trouble viewing inline? Open PDF directly →
Full Text
99,023 characters extracted from source content.
Expand or collapse full text
1 HowtoDisclose?StrategicAIDisclosureinCrowdfunding NingWang SchoolofBusiness UniversityofConnecticut ning.wang@uconn.edu ChenLiang SchoolofBusiness UniversityofConnecticut chenliang@uconn.edu 2 Abstract Asartificialintelligence(AI)increasinglyintegratesintocrowdfundingpractices,strategic disclosureofAIinvolvementhasbecomecritical.Yet,empiricalinsightsintohowdifferent disclosurestrategiesinfluenceinvestordecisionsremainlimited.Drawingonsignalingtheory andAristotle’srhetoricalframework,weexaminehowmandatoryAIdisclosureaffects crowdfundingperformanceandhowsubstantivesignals(degreeofAIinvolvement)and rhetoricalsignals(logos/explicitness,ethos/authenticity,pathos/emotionaltone)moderatethese effects.LeveragingKickstarter’smandatoryAIdisclosurepolicyasanaturalexperimentand foursupplementaryonlineexperiments,wefindthatmandatoryAIdisclosuresignificantly reducescrowdfundingperformance:fundsraiseddeclineby39.8%andbackercountsby23.9% forAI-involvedprojects.However,thisadverseeffectissystematicallymoderatedbydisclosure strategy.GreaterAIinvolvementamplifiesthenegativeeffectsofAIdisclosure,whilehigh authenticityandhighexplicitnessmitigatethem.Interestingly,excessivepositiveemotionaltone (astrategycreatorsmightintuitivelyadopttocounteractAIskepticism)backfiresand exacerbatesnegativeoutcomes.Supplementaryrandomizedexperimentsidentifytwounderlying mechanisms:perceivedcreatorcompetenceandAIwashingconcerns.Substantivesignals primarilyaffectcompetencejudgments,whereasrhetoricalsignalsoperatethroughvaried pathways:eithermediatoraloneorbothinsequence.Thesefindingsprovidetheoreticaland practicalinsightsforentrepreneurs,platforms,andpolicymakersstrategicallymanagingAI transparencyinhigh-stakesinvestmentcontexts. Keywords:ArtificialIntelligence,InformationDisclosure,Crowdfunding,InvestmentDecisions, SignalingTheory 3 1.INTRODUCTION Theincreasingintegrationofartificialintelligence(AI)intobusinessoperationshas intensifiedconcernsregardingitspotentialrisksandunintendedconsequences,prompting regulatorybodiestointroducestricterdisclosurerequirements.Legislativeinitiativessuchas California’sSB1047 1 andtheEUAIAct 2 mandatetransparencyinAIdeployment,aimingto improveoversightandensureresponsibleAIuse.Concurrently,anemergingstreamofresearch hasinvestigatedhowAItransparencyshapesindividualperceptionsanddecisions,highlighting itseffectsontrust(Renierisetal.,2024;Schankeetal.,2024),consumerengagement(Carneyet al.,2024;Luoetal.,2019),andemployeeperformance(Tongetal.,2021). Thesedynamicstakeonparticularimportanceincrowdfunding.Traditionalinvestment contextsofferinformationalanchorssuchasauditedfinancials,third-partyverification,andprior relationshipstoassessquality.Crowdfundingbackerslacksuchanchors,makingcreator disclosure(Bhargavaetal.,2024;Casonetal.,2025;Fuetal.,2025;Kimetal.,2022;Lin& Viswanathan,2016)theprimarychannelforassessingcreatorcompetenceandprojectquality. WhenAIisinvolved,disclosurebecomesespeciallyconsequential:backersarelefttointerpret howAIcontributestotheprojectandwhatitimpliesforhumaninvolvementandcapability.This considerationisparticularlyrelevantintheearly-stagecreativeprojectstypicalofcrowdfunding. Consequently,differentwaysofframingAIusagecanleadtomarkedlydifferentinterpretations amongprospectivebackers,makingcrowdfundinganaturalsettingforexamininghow disclosurestrategiesshapeeconomicbehavior. 1 CaliforniaSenateBill1047isavailableat https://leginfo.legislature.ca.gov/faces/billNavClient.xhtml?bill_id=202320240SB1047(accessedJune5,2025). 2 TheEuropeanParliamentandCouncil'sRegulation(EU)2024/1689,knownastheEUAIAct,isavailableat https://w.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial- intelligence(accessedJune5,2025). 4 DespitegrowingregulatoryandscholarlyattentiontoAItransparency,existingresearch focusesprimarilyonthebinarydecisionofwhetherornottodiscloseAI(e.g.,Baueretal.,2025; Carneyetal.,2024;Luoetal.,2019;Schankeetal.,2024).Muchlessisknownaboutthe potentialroleofAIdisclosurestrategiesinshapingconsequentialstakeholderdecisionsinthe digitaleconomy,particularlyinsettingsthatinvolvehigh-stakeseconomiccommitmentsand relyheavilyoninference,suchascrowdfunding.Understandinghowspecificdisclosure strategiesencourageordeterinvestmentis,therefore,essentialforboththeoryandpracticeinthe digitaleconomy. Giventhatboththecontentandthestyleofdisclosuremayshapebackerinferences,we examinetwocomplementaryformsofsignalinginAIdisclosure:substantivesignals(Connelly etal.,2011)andrhetoricalsignals(Suddaby&Greenwood,2005;Vaaraetal.,2016). Substantivesignalsconveyfact-basedinformationaboutacreator’scapabilitiesortheproject developmentprocess(Connellyetal.,2011;Sahlman,1990),offeringbackerstangiblecues aboutexpectedproductquality(Steigenberger&Wilhelm,2018).AppliedtoAIdisclosure, substantivesignalscapturetheactualdeploymentofAIwithinaproject(“AIinvolvement”), informingbackers’judgmentsaboutprojectqualityandcreatorcompetence. Rhetoricalsignals,incontrast,reflectlanguage-basedfeaturessuchascommunicationstyle, tone,andpresentation(Suddaby&Greenwood,2005;Vaaraetal.,2016)thatshape interpretationindependentoffactualcontent.GuidedbyAristotle’srhetoricaltriangle(Aristotle, 1991),wefocusonthreerhetoricaldimensionsofAIdisclosure:explicitness(logos),referringto thelevelofdetailprovided;authenticity(ethos),referringtotheperceivedcredibilityand trustworthiness;andemotionaltone(pathos),referringtothedegreeofpositivityconveyedinthe disclosure.Whiletheimportanceofsubstantiveandrhetoricalcuesiswell-establishedin 5 contextssuchasfinancialreporting(Feldmanetal.,2010),technologylicensing(Truongetal., 2022),andpublichealthcommunication(Houetal.,2024),itremainsunclearhowthesecues influencepledgedecisionsinthecontextofAIdisclosure.Thisgapisconsequential:unlike traditionaltechnologiesthatprimarilyaugmenthumancapabilities,generativeAIcan complement,substitutefor,orconstrainhumaninput(Songetal.2024;Houetal.2025;Zhang etal.2025),whileexhibitingfargreateropacityandcomplexity.Thesefeaturescomplicate qualityassessmentandcreatestrategicchallengesinbothwhatisdisclosedandhow,challenges largelyabsentfrompriortechnologycontexts.Motivatedbythesegaps,weexamine: 1)HowdoesAIdisclosureaffectcrowdfundingsuccess? 2)HowdoesthisimpactdifferacrossdifferentAIdisclosurestrategies? Toanswerthesequestions,wecollecteddatafromKickstarter,whichintroduceditsAI disclosurepolicyinAugust2023.Thepolicyrequirescreatorstobetransparentandspecific abouthowtheyuseAIintheirprojects.Followingpriorliterature,weuseakeyword-based approachtoidentifyprojectsthatemployAI(Babinaetal.,2024;Lou&Wu,2021;Wuetal., 2025),bothbeforeandafterthepolicy,andtreattheseprojectsasthetreatmentgroup,withall otherprojectsservingasthecontrolgroup.Leveragingthispolicychange,weemploya difference-in-differences(DID)designtoestimatethecausaleffectofAIdisclosureon crowdfundingperformanceandexaminehowtheimpactofAIdisclosurevarieswithdifferent disclosurestrategies. OurfindingsrevealthatAIdisclosuresignificantlydecreasescrowdfundingperformance. Specifically,theintroductionoftheAIdisclosurepolicyleads,onaverage,toa39.8%declinein fundsraisedanda23.9%decreaseinthetotalnumberofbackersforAI-relatedprojects.This negativeeffectissystematicallymoderatedbybothsubstantiveandrhetoricalsignals.Regarding 6 substantivesignals,greaterAIinvolvementamplifiestheadverseimpactonfundingoutcomes. Amongrhetoricalsignals,highexplicitness(logos)andhighauthenticity(ethos)mitigatethe negativeeffect,whilehighpositiveemotionaltone(pathos)exacerbatesit. Touncovertheunderlyingmechanismsdrivingthesemoderatingeffects,weconductfour scenario-basedrandomizedonlineexperimentsonProlific.Eachexperimentmanipulatesone disclosuredimension—AIinvolvement,explicitness,authenticity,oremotionaltone—througha between-subjectsdesignwhereparticipantsevaluateeitherahighorlowversionofthat dimension.ManipulationsareconstructedbyeditingrealKickstarterAIdisclosures:one conditionpresentstheoriginaltext(editedminimallyforreadability);theotherinvertsthefocal dimensionviaLargeLanguageModel(LLM)modification,limitingchangestonon-focal content(Bhattacharjeeetal.,2024;Youssefetal.,2024).Thisdesignallowsustoexaminehow variationsinAIdisclosureshapeinvestors’perceptions,whichinturninfluencetheirwillingness topledge. Foursupplementaryonlineexperimentsconfirmthesemoderatingpatternscausallyand elucidatetheunderlyingmechanisms.GreaterAIinvolvementdecreasespledgeintentionby reducingperceivedcreatorcompetence.Highexplicitnessincreasespledgeintentionthroughtwo pathways:directly,byenhancingperceivedcreatorcompetence,andindirectly,byreducingAI washing(i.e.,deceptiveAIpositioning)concerns,whichinturnbolstercompetenceperceptions. Highauthenticityincreasespledgeintentionpredominantlythroughaserialpathway:itfirst mitigatesAIwashingconcerns,whichsubsequentlyenhancesperceivedcreatorcompetence. Conversely,excessivelypositiveemotionaltonedecreasespledgeintentionthroughthesame serialpathwaybutwithoppositeeffects:itfirstamplifiesAIwashingconcerns,whichthen diminishesperceivedcreatorcompetence,therebyloweringpledgeintention. 7 OurstudycontributestomultiplestreamsofliteratureonAItransparencyandcrowdfunding. First,whereaspriorresearchexaminesAIdisclosureinlow-stakesconsumptioncontextssuchas customerservice(Luoetal.,2019;Xuetal.,2024),contentgeneration(Baueretal.,2025),and newsconsumption(Toff&Simon,2024),wheredecisionsinvolveminimalresource commitmentandareeasilyreversible,weinvestigateAIdisclosureinsettingsrequiring consequentialresourceallocationdecisions.Incrowdfundingcontexts,backersmustevaluate projectqualityundersignificantinformationasymmetrywhilecommittingcapitaltoinherently riskyventures,makingdisclosurestrategiesdirectlyconsequentialforfundingoutcomes.The diversityofprojecttypesandlevelsofAIinvolvementprovidesanidealsettingforexamining howAIdisclosureshapesinvestmentbehaviorforearly-stagecreativeventures. Second,ourstudyextendstheAItransparencyliteraturebyshiftingthefocusfromwhether todiscloseAItohowtodiscloseiteffectively.AsAItransparencybecomesincreasingly mandatedbyregulatorsandplatforms,disclosureisnolongerabinarychoicebutanunavoidable requirement.Thisshiftmakesthedesignofdisclosure,ratherthanitsmerepresence,centralto stakeholderinterpretation,especiallyincrowdfunding,whereinvestorsrelyonnarrativecues andinferenceplaysadominantrole.Existingresearch,however,remainslargelyfocusedonthe effectsofAIdisclosurepresenceversusabsence(Baeketal.,2024;Luoetal.,2019;Tongetal., 2021;Xuetal.,2024),offeringlimitedinsightintohowdifferentdisclosurestrategiesinfluence perceptionsanddecision-makingoncedisclosureisrequired.Weaddressthisgapbyapplyinga substantive–rhetoricalsignalingframeworktotheAItransparencydomain.Bydistinguishing substantivesignals(AIinvolvement)fromrhetoricalsignals(explicitness,authenticity, emotionaltone),weshowthatdisclosurestrategysystematicallyshapesperceivedprojectquality 8 andcreatorcompetence,providingaconceptualfoundationforunderstandingAItransparency interpretationwhendisclosurebecomesmandatory. Finally,ourfindingscontributetothebroaderunderstandingofAIasauniqueresourceby revealingtheunderlyingmechanismsthroughwhichAIdisclosurestrategiesinfluencedecision- making.WhilepriorworkdocumentsthatAIdisclosureaffectsoutcomes,theunderlying processesremainlargelyunexplored,limitingboththeoreticalunderstandingandpractical guidance.GivenAI’suniquecharacteristics(itscapacitytooffloadorsubstitutehumaninput, combinedwithinherentcomplexityandopacity)(Benbyaetal.,2021;LuandZhang,2025),we proposethatAIdisclosureoperatesthroughtwodistinctmechanisms:perceivedcreator competenceandAIwashingconcerns.Critically,wedemonstratethatsubstantiveandrhetorical signalsactivatethesemechanismsdifferently.Ourexperimentalevidenceconfirmsthat substantivesignalsaffectpledgeintentionsprimarilythroughcompetencejudgments,whereas rhetoricalsignalsexhibitmorevariedpathways,operatingthrougheithermechanism independentlyorthroughbothinsequence.Ourinsightsadvancethetheoreticalunderstandingof AItransparencyasauniquesignalingprocessandprovideactionableguidanceforcreators, platforms,andpolicymakersindesigningeffectivedisclosurestrategiesinAI-integrateddigital markets. 2.LITERATUREREVIEW 2.1.AITransparencyintheDigitalEconomy TheexpandingpresenceofAIindigitalserviceshasintensifiedinterestinhow organizationscommunicatealgorithmicinvolvementtousers.AIdisclosure—broadlydefinedas revealingwhenandhowAIcontributestoaproduct,service,ordecision—hasbecomeacentral topicinbothregulatorydebatesandacademicresearch(Luoetal.2019;ElAlietal.2024;Xuet 9 al.2024;Wittenbergetal.2025).Recentpolicyframeworks,includingtheEUAIActand platform-leveltransparencyguidelinesissuedbyfirmssuchasGoogle 3 andTikTok, 4 reflecta broadermovementtowardmakingalgorithmicprocessesmorevisible,understandable,and accountabletothepublic.Correspondingly,researchershavebeguntodocumenthowthe disclosureofanAInatureshapespeople’strust(Yinetal.,2024),evaluationprocesses (Chiarellaetal.,2022),userengagement(Carneyetal.,2024;Xuetal.,2024),andbehavioral responses(Tongetal.,2021)acrossarangeofenvironments.Forexample,inorganizational contexts,disclosingthatperformancefeedbackisgeneratedbyAItendstoloweremployees’ trustinthefeedbackandheightentheirconcernsaboutbeingreplaced,whichsubsequently reducestheirlearningandjobperformance(Tongetal.,2021).Inthecreativeindustries,the disclosureofAIauthorshipreducesaestheticappreciationofartworks(Chiarellaetal.,2022).In thelogisticsindustry,identitydisclosureofAIvoicechatbotsreducesresponselikelihood, whereasanthropomorphiccues(e.g.,fillerwords)positivelyinfluenceuserengagement(Xuet al.,2024). Meanwhile,emergingworksuggeststhatAIdisclosure’sconsequencesarenotuniformly adverse.Onshort-formvideoplatforms,forexample,disclosurecanincreaseusers’willingness toengage,particularlywhentheAIisperceivedascapable(Chenetal.,2025).Voice-based agentsequippedwithdeepfakevoicecloninglikewiseelicithighertrust,andthistrustpersists evenwhenAIinvolvementisdisclosed(Schankeetal.,2024).Thesefindingsindicatethatthe impactofAIdisclosureindigitalcontextsisheterogeneousandcontingentontheplatform,task type,andusers’priorexpectations. 3 Google’sreportonongoingworkonresponsibleAIisavailableathttps://blog.google/technology/ai/responsible-ai- 2024-report-ongoing-work/(accessedNovember1,2025). 4 TikTok’sAI-GeneratedContentSupportGuidelinesareavailablehttps://support.tiktok.com/en/using- tiktok/creating-videos/ai-generated-content(accessedJanuary16,2026). 10 ExistingAIdisclosureresearchconceptualizesdisclosureasabinarytreatment,i.e., disclosedversusnotdisclosed,designedtoisolatetheaverageeffectofan“AIlabel”onusers’ trustorengagement.Thisishighlyinformativeaboutwhetherdisclosurematters,butitleaves underspecifiedthemanagerialquestionthatarisesinthedigitaleconomy:howtodisclose. Moreover,muchofpriorliteraturefocusesonattitudinalreactionsorrelativelylow-stakes engagementmetrics(e.g.,conversationalresponses),whereasmanyconsequentialdecisionsin thedigitaleconomyinvolvehigh-stakeseconomiccommitments.Crowdfundingexemplifies suchsettings:backersmakeirreversibleinvestmentdecisionsunderinformationasymmetry,and transparency,whileaidingevaluation,mayintroduceunintendedcosts(Yangetal.,2022).Tofill thesegaps,westudyhowdifferentAIdisclosurestrategiesshapebackers’actualinvestment decisionsonKickstarter.Movingbeyondthebinary“AIlabel,”weanalyzevariationinthe contentandstyleofdisclosurelanguageandtesthowthesedisclosurechoicestranslateinto downstreamcampaignfundingoutcomes. 2.2.SignalinginCrowdfundingPlatforms Crowdfundingexemplifiesamarketcharacterizedbypronouncedinformationasymmetries: creatorsmustpersuadeadispersedsetofbackerswhopossesslimitedverifiableinformation aboutprojectfeasibilityorcreatorcapability(Ahlersetal.,2015;Yangetal.2016;Courtneyet al.,2017;Sabzehzaretal.2023).Giventheseasymmetries,prospectivebackersrelyheavilyon thecontentpresentedincrowdfundingcampaignstoassessprojectqualityandmakeinvestment decisions.Thisrelianceonsuchobservablesignalsmakescrowdfundingaparticularlywell- suitedcontextforapplyingsignalingtheorytodigitalmarkets,whereverifiabilityislimitedand informationalnoiseishigh(Connellyetal.,2011;Spence,1973). 11 Earlyworkemphasizessubstantivesignals,thatis,relativelycostlyandfact-based indicatorsthatcrediblyreduceinformationasymmetry(Berghetal.,2014;Connellyetal.,2011; Sahlman,1990).Incrowdfunding,prototypes,priorachievements,technicaldetails,andexternal certificationserveasimportantmarkersoffeasibilityandcreator/firmcompetence(Ahlersetal., 2015;Courtneyetal.,2017;Xiaoetal.,2021).Additionalsubstantivesignalsthatspeakto potentialsourcesofuncertainty,suchasdisclosuresofprojectrisk(Kimetal.,2022),unaudited financialstatements(Donovan,2021),additionalbudgetinformation(Fuetal.,2025), prefundingcommunication(Weietal.,2021),timelyupdatesorFAQs(Xiaoetal.,2021),also providebackerswithrelevantinformationaboutthelikelihoodofsuccessfuldelivery.These signalshelpbackersassesswhethercreatorspossesstheabilityandcommitmentneededto deliver. Yetsubsequentresearchshowsthatsubstantivecuesaloneareofteninsufficientinthis high-noiseenvironment,wherebackersmustformimpressionsquicklyandrarelyengageinany formalvettingprocess(Anglinetal.,2025).Insuchsettings,rhetoricalstrategiesalsotendto playanimportantroleinconveyingrelevantinformation(Anglinetal.,2025;Chandleretal., 2024;Steigenberger&Wilhelm,2018;Vaara&Monin,2010).Forinstance,therhetorical framingofsocialresponsibilityexhibitsaninvertedU-shapedrelationshipwithfunding performance—moderateemphasisprovesmoreeffectivethaneitherminimalorexcessive rhetoric(Anglinetal.,2025).Similarly,rhetoricaldisplaysofpassionthroughlinguisticmarkers positivelyaffectsocial-mediaexposureandcrowdfundingsuccess(Lietal.,2017).Attimes, theserhetoricalsignalsmaystrengthenorweakentheinfluenceofsubstantivesignalsonafirm’s fundingperformance(Steigenberger&Wilhelm,2018). 12 Despitesubstantialprogressinunderstandingcrowdfundingsignalingmechanisms,two criticalgapslimitourabilitytoexplainstrategicdisclosureintechnology-intensiveventures. First,priorresearchlargelytreatstechnologyadoptionasanunambiguouslypositivequality signal,implicitlyassumingthatmoreadvancedtechnologiesuniformlyenhanceperceivedeffort andproductquality(Donovan,2021).Thisassumptionisincreasinglyuntenableinthecontextof AI,wheredisclosurecouldsimultaneouslysignalinnovationwhileraisingconcernsabout reducedcreatoreffortandotherpotentialdrawbacks.Moreover,rhetoricalstrategieseffectivefor traditionaltechnologiesmaybackfireinAIcontexts.Unlikeearliertechnologiesthatprimarily augmentedhumancapabilities,AIcansubstitutefororconstrainhumancreativelabor(Zhanget al.2024,2025)whileexhibitinghighlevelsofcomplexityandopacity(Benbyaetal.,2021). Thesecharacteristicsproduceambivalentqualityinferenceswithfewprecedentsinprior technologycontexts.Weaddressthesegapsbyinvestigatinghowsubstantivesignals(e.g.,the extentofAIusage)andrhetoricalsignals(i.e.,howAIusageisdisclosed)jointlyshape crowdfundingperformance.Indoingso,weextendsignalingtheorytotheAIdeployment context,wheresignalsareinherentlyvalue-ladenorcontested,andprovideactionableguidance formanagingAItransparencyinentrepreneurialcontexts. 3.RESEARCHFRAMEWORK BuildingonpriorworkonAItransparencyinthedigitaleconomyandsignalingin crowdfundingplatforms,weproposeaframeworkforunderstandinghowcreatorsstrategically discloseAIusage.Ourframeworkdistinguishesbetweentwocomplementarydimensionsof disclosure:substantivesignals,whichindicatetheextentofAIdeployment,andrhetorical signals,whichcapturehowAIdeploymentiscommunicated. 13 3.1.SubstantiveSignals Substantivesignalsprovideinformationabouttheunderlyingproductionprocess,including theresources,capabilities,andeffortcontributingtoanoutput.Incrowdfunding,suchsignals typicallyincludeprototypes,priorachievements,technicalspecifications,orimplementation plansthathelpreduceinformationasymmetrybyindicatingfeasibilityandcompetence. InAI-enabledprojects,substantivesignalsconcerntheextentofAIinvolvement.These detailsspecifywhetherAIsupportsideation,contentgeneration,refinement,customer interaction,orfulfillment,andwhetheritaugmentsorsubstitutesforhumancreativelabor.These distinctionsmatter:greaterAIintegrationcansimultaneouslyexpandproductioncapabilities (Songetal.,2024)whilepotentiallyraisingconcernsaboutdiminishedhumaneffortor heightenedexecutionrisk(Chiarellaetal.,2022;Tongetal.,2021).Substantivesignals, therefore,mayshapebackers’beliefsaboutcreatorcompetenceandprojectquality,ultimately influencingfundingdecisions. 3.2.RhetoricalSignals Whilesubstantivesignalscommunicatewhatisbeingofferedandhowitisproduced, rhetoricalsignalsshapehowthatinformationisframedandcommunicated.Rhetoricalsignals operatethroughlanguagechoice,framing,narrativestructure,andcommunicationstyle, influencinghowaudiencesevaluateandweighuncertaininformation(Parhankangas&Renko, 2017).AusefulorganizinglensisAristotle’srhetoricaltriangle,namely,logos,ethos,andpathos, whichemphasizesthatpersuasiondependsnotonlyonfactualclaimsbutalsoonrhetorical appealsandaffectiveresonancewithaudiencevalues(Steigenberger&Wilhelm,2018). Consistentwiththisview,policycommunicationresearchshowsthatrhetoricalappealsshape responsestopoliticalandpolicyframing(Gottweis,2017;Stucki&Sager,2018),andcorporate 14 communicationresearchdocumentssimilardynamicsinsustainabilityreporting(Higgins& Walker,2012). Logos-basedsignals.Logos-basedcommunicationemphasizesclear,logicalreasoningand rationalargumentationthroughspecificclaims,technicalexplanations,andfactualdetails (Gottweis,2017).Incrowdfundingsettings,signalsgroundedinclarityandspecificityreduce perceivedexecutionriskandincreaseperceivedfeasibility(Parhankangas&Renko,2017).Prior researchfurthershowsthatnarrativestructureandlanguagechoicesthatenhanceclarityand explicitnessshapehowbackersevaluateprojectfeasibilityand,inturn,influencefunding outcomes(Moradietal.,2024).Consistentwiththisview,riskdisclosuresthatproviderelevant detailsandmaintainabalancedtonearemorepersuasiveforprojectscharacterizedbyhigh uncertainty(Kimetal.,2022). Tooperationalizelogos-basedrhetoricalsignalsinthecontextofAIdisclosure,wefocuson perceivedexplicitness.Inthissetting,perceivedexplicitnesscapturestheextenttowhicha project’sAIdisclosureprovidesaclear,logical,andwell-reasonedexplanationofhowAIisused intheproject.Highperceivedexplicitnessischaracterizedbylanguagethatisspecific, informative,andtechnicallydescriptive,whereaslowperceivedexplicitnessreliesonvague, abstract,orlackingindetail. Ethos-basedsignals.Ethos-basedsignalsemergefromcommunicationthatappearsgenuine, personallyinvested,andtransparentlyhonestratherthanstrategicallymanipulated(Higgins& Walker,2012).Suchsignalsareoftenconveyedthroughauthenticnarrativesthatincorporate personallanguagedescribingcreators’ownjourneys,candiddiscussionsofchallenges,and realisticassessmentsofprojectrisks.Priorresearchshowsthatethos-basedcues,including perceivedfounderauthenticity,enhancebackers’trustandemotionalwarmthtowardaproject 15 (Radoynovska&King,2019).Similarly,brandprominencecanoperateasanethos-basedsignal byconveyingcredibilityandreliability(Moradi&Badrinarayanan,2021). Tocaptureethos-basedrhetoricalsignalinginAIdisclosurecontexts,wefocusonperceived authenticity.Perceivedauthenticityreferstotheextenttowhichthedisclosurereflectsthe creator’scredibilityandtrustworthinessthroughself-revealingandpersonalexpressionrather thanimpersonalanddetachedcommunication. Pathos-basedsignals.Pathos-basedsignalsrefertotheaffectivevalenceembeddedin campaigncommunication,whichcanrangefromrestrained,low-arousalpresentationstohighly enthusiasticandemotionallychargedexpressions(Munyon&Summers,2024).Priorresearch showsthatpositiveemotionaltone,reflectedinhighlyoptimisticlanguageandenthusiastic expressions,cangeneratepsychologicalengagementandidentificationamongaudiences(Liet al.,2017).Incrowdfundingcontexts,pathos-basedsignals,includingemotionalappeals embeddedinnarratives,influencebackerattentionandengagement(Xiangetal.,2019). Relatedly,evidencefromdonationandpersuasionsettingssuggeststhatheightenedemotional cues,particularlywhentheyappearexaggeratedorincongruent,canintensifyaffective processingandshapecontributionbehavior(Yazdanietal.,2025). InAIdisclosurecontexts,however,emotionalframingmayplayamorecomplexrole,as overlyenthusiasticexpressionscanamplifybothexcitementandskepticismtowardthe technology.Tocapturethisdimensionofpathos,wefocusonperceivedexcessivepositive emotion,measuringtheextenttowhichthetoneofAIdisclosureconveysexaggeratedoptimism, inflatedenthusiasm,orforcedconfidenceregardingtheuseofAI. 16 Figure1.ResearchFramework Overall,weproposethatAIdisclosure,whichcontainsbothsubstantiveandrhetorical elements,naturallyfitswithinamultidimensionalsignalingframework.Substantively,disclosure communicatestheextentandnatureofAIinvolvement.Rhetorically,creatorscanframethis involvementthroughvaryinglevelsofexplicitnessassociatedwithlogos-basedappeals, authenticitycuesreflectingethos,andemotionaltonereflectingpathos.Figure1illustratesthe proposedresearchframework. 4.EMPIRICALSETTINGANDMODEL 4.1.ResearchContext OurstudydrawsondatafromKickstarter,oneofthelargestglobalcrowdfundingplatforms, whichhoststhousandsofcampaignsannuallyacrosstechnology,design,andcreativecategories. OnAugust29,2023,Kickstarterintroducedaplatform-widedisclosurepolicyrequiringcreators todeclarewhetherAItoolswereusedinthedevelopment,design,orpromotionoftheir projects. 5 Specifically,duringprojectsubmission,creatorswerepromptedtocompleteanAI disclosuresectiondetailingwhetherandhowgenerativeAIcontributedtoanypartoftheir 5 Kickstarter’sdisclosurepolicyisavailableathttps://updates.kickstarter.com/introducing-our-new-ai-policy/ (accessedJune5,2025). 17 campaign,suchascopywriting,visualdesign,coding,orprototyping.Thesedisclosuresare displayedpubliclyoneachcampaignpageinadedicated“UseofAI”section. Ourobservationwindowspansoneyearbeforeandafterthepolicy,coveringprojectsthat endedbetweenAugust1,2022,andAugust31,2024.Weclassifyprojectsintopre-policyand post-policyperiodsbasedontheircampaignendmonthratherthanlaunchmonth.Becausethe averageprojectdurationexceedsonemonth,projectslaunchedbeforeAugust2023butclosing afterwardmayhavebeenexposedtothepolicy.Classifyingbylaunchmonthwouldmisclassify suchprojectsasuntreated.Classifyingbyclosingmonthensuresaccuratetreatmentassignment andyieldsmoreconservativeestimates. 6 4.2.Keyword-BasedIdentificationofAI-RelatedCampaigns Followingpriorworkthatexploitspolicy-inducedvariationtoidentifycausaleffects(e.g., Chenetal.,2011;Chenetal.,2017;Dewanetal.,2017;Huangetal.,2017),wetreatthe platform’sintroductionofdisclosurerequirementsasanexogenousshockthatappliesonlyto AI-relatedprojects.ThispolicychangegeneratesdifferentialexposurebetweenAI-related (treated)andnon-AI-related(control)projects,therebyenablingdifference-in-differences(DID) estimation. AkeyempiricalchallengeisaccuratelyidentifyingwhichcampaignsareAI-relatedbefore andaftertheimplementationofthedisclosurerequirement.Toaddressthis,wefollowthe existingliterature(Babinaetal.,2024;Lou&Wu,2021;Wuetal.,2025)andidentifyAI-related projectsusingakeyword-basedapproach,whichhasbeenwidelyusedincategorizing technologicalconceptsandsignificantlymitigatessubjectiveclassificationconcerns.Wedevelop acomprehensivekeyworddictionarytoidentifycrowdfundingcampaignsthatpotentially 6 Asarobustnesscheck,weclassifyprojectsintopre-policyandpost-policyperiodsbasedoncampaignlaunchdate (ratherthanenddate),excludingcampaignsthatspanthepolicyimplementationdate.Resultsarehighlyconsistent (seesection7.2). 18 involveAIusage,whicharesubjecttothedisclosurerequirementafterthepolicychangebut wouldnothavebeenrequiredtodosohadtheybeenpostedbeforetheshock. Toensurebothcomprehensivenessandconstructvalidity,wedevelopathree-partkeyword dictionarythatcaptureshowAIisdescribedinbothtechnicalandpracticalterms.Thefirstpart consistsofafoundationalsetofcanonicalAIterms(n=75)fromtheexistingliterature(Babina etal.,2024;Miricetal.,2023),includingwidelyrecognizedtermssuchasmachinelearning, deeplearning,naturallanguageprocessing,neuralnetwork,andcomputervision.These keywordscapturethecoretechnicaldomainsconsistentlyreferencedinAIresearchandpolicy frameworks. ThesecondpartcaptureshowcreatorsdescribeAIinpractice.Weextendedthedictionary withexpressionsextractedfromcampaigndescriptions.Specifically,weusedGPT-4o-minito identifyAI-relatedkeywordsbasedonthespecificquestionsinAIdisclosures(seeOnline AppendixB1 7 fortheprompt).Wethenparsedallcampaignsinoursampleandconducted frequencyanalysisonkeywordsfromcampaignsthatself-identifiedasAI-related.Weapplied threefilters:first,manualscreeningbytheauthorsremovedconceptuallyirrelevantterms; second,weretainedonlykeywordsappearingatleastthreetimes,whichaccountedfor83%of allkeywordoccurrences;third,weexcludedtermsalreadypresentinthecanonicalAIdictionary fromPart1.Thisprocessyielded27additionalkeywords. ThethirdpartreflectsthegrowingprominenceandrapiddevelopmentofgenerativeAI.We supplementedthedictionarywithnamesofmajorgenerativeAImodels,suchasChatGPT, Claude,Gemini,LLaMA,andMistral(n=30).Incorporatingthesetermsallowsustocapture campaignsthatexplicitlyreferencecontemporarygenerative-AItools. 7 Duetospacelimitations,theonlineappendicesarehostedontheOpenScienceFramework(OSF)at https://osf.io/6bcva/overview?view_only=b378f0d57ec9412186292b41c8ea3d8c. 19 AcampaignisclassifiedasAI-relatedifitstitleordescriptioncontainsatleastonekeyword fromthiscomprehensivecodebook.Thisprocedureprovidesatransparentandreproducible methodforAI-relatedcampaignsforourmainDIDanalysis.Importantly,thiskeyword-based classificationcapturesapproximately97%(1,185outof1,220)ofcampaignsthatexplicitly discloseAIusage,indicatingahighdegreeofoverlapbetweenouroperationaldefinitionand self-reportedAIadoption.AllkeywordscuratedduringtheprocessarelistedinAppendixA. 4.3.Variables DependentVariables.Followingtheexistingliterature(Fuetal.,2025;Gevaetal.,2024), weuseLogTotalPledgeandLogTotalBackersasourmaindependentvariables.LogTotalPledge, isthenaturallogarithmofthetotalamountpledgedbybackers(inU.S.dollars),whichcaptures theoverallfundingperformanceofacampaign.Inaddition,weuseLogTotalBackers,definedas thenaturallogarithmofthetotalnumberofbackerswhosupportedtheproject,asanalternative dependentvariabletomeasurethereachofcampaignengagement. FocalVariables.ThekeyexplanatoryvariableisTreatment,adummyvariablethatequals 1ifaprojectisclassifiedasanAI-relatedproject.Toaccountforthetimingoftheplatform policychange,weincludeatimedummyvariableAfter,whichequals1iftheprojectwasclosed followingtheimplementationoftheAIdisclosurepolicy.Additionally,weincludeAIDisclosure, adummyvariablethatequals1ifaprojectdisclosestheuseofAIduringthepost-policyperiod. Moderators.Drawingonthesubstantiveandrhetoricalsignalsframework,weconstruct fourmoderatingvariablestocapturedisclosurevariation:onesubstantivedimension(AI involvement)andthreerhetoricaldimensions(explicitness,authenticity,andemotionaltone). WeemployGPT-4o-minitoclassifythesefeatures. 8 RecentresearchdemonstratesthatLLMs caneffectivelyperformtextualanalysistaskswithaccuracycomparabletoexperthumancoders 8 Wesetthetemperatureparametertozerotoensurefocusedanddeterministicclassifications(deKok,2025). 20 whileofferingsubstantiallygreaterscalabilityandcostefficiency(Bail,2024;deKok,2025; Gilardietal.,2023).LLMsareparticularlywell-suitedforcomplexclassificationtasksrequiring contextualinterpretationandnuancedjudgment(deKok,2025). Weoperationalizethefourdimensionsasfollows:1)HighAIInvolvementisadummy variableequalto1iftheprojectdemonstratesextensiveAIinvolvement,whereAIwascentralto producingtheproject’sprimaryoutputratherthanfunctioninginaperipheralsupportingrole (Eloundouetal.,2024),asclassifiedbyGPT-4o-minibasedonAIdisclosures.2) HighExplicitnessisabinaryvariableequalto1ifthedisclosureexhibitshighexplicitness(above thesamplemedian),wherethedisclosureprovidesaclear,logical,andtechnicallydescriptive explanationofAIusageratherthanvagueorgeneralstatements(Higgins&Walker,2012),as scoredbyGPT-4o-mini.3)HighAuthenticityisabinaryindicatorequalto1ifthedisclosure demonstrateshighauthenticity(abovethesamplemedian),meaningthatthecreatorconveys credibilityandtrustworthinessthroughpersonal,genuine,andhonestwritingratherthangeneric orimpersonallanguage(Bolingeretal.,2024;Higgins&Walker,2012),asscoredbyGPT-4o- mini.4)HighPosEmotionisadummyvariableequalto1ifthedisclosureexhibitsahighly positiveemotionaltone,definedashavingapositivityscoreabovethesamplemedian,indicating excessivelypositiveemotionaltone.ThisemotionmeasureisconstructedusingtheValence AwareDictionaryforSentimentReasoning(VADER)package(Huetal.,2021;Hutto&Gilbert, 2014).Allclassificationsareoperationalizedviatext-basedassessmentofAIdisclosures. DefinitionsandsummarystatisticsofallaforementionedvariablesarereportedinTable1. 21 Table1.DefinitionsandSummaryStatisticsofVariables VariableVariabledefinitionsObsMeanSDMinMax Dependentvariables LogTotalPledgeTotalamount(inU.S.dollars)raisedbytheproject(log- transformed) 35,8327.4382.773015.714 LogTotalBackersTotalnumberofbackerswhosupportedtheproject(log- transformed) 35,8323.6081.726010.515 Focalvariables TreatmentAdummyvariablethatequals1ifaprojectisclassifiedasAI- related 35,8320.1490.35601 AfterAdummyvariablethatequals1iftheAIdisclosurepolicywas alreadyintroducedinagivenmonth 35,8320.5730.49501 AIDisclosureAdummyvariablethatequals1ifaprojectdisclosestheuseof AIduringthepost-policyperiod 20,5140.0590.23701 Moderators HighAIInvolvementAdummyvariableequalto1iftheprojectdemonstrates extensiveAIinvolvement,whereAIwascentraltoproducing mostoftheproject’soutput,asclassifiedbyGPT-4o-minibased onAIdisclosures. 20,5140.0180.13301 HighExplicitnessAdummyvariableequalto1iftheprojectdemonstrateshigh disclosureexplicitness(i.e.,above-medianexplicitness),where thedisclosureprovidesaclear,logical,andtechnically descriptiveexplanationofAIusageratherthanvagueorgeneral statements,asscoredbyGPT-4o-minibasedonAIdisclosures. 20,5140.0320.17601 HighAuthenticityAdummyvariableequalto1iftheprojectdemonstrateshigh disclosureauthenticity(i.e.,above-medianauthenticity),where thedisclosurereflectsthecreator’scredibilityand trustworthinessthroughpersonalandhonestwritingratherthan genericorimpersonallanguage,asscoredbyGPT-4o-minibased onAIdisclosures. 20,5140.0340.18001 HighPosEmotionAdummyvariableequalto1iftheprojectdemonstratesahigh positiveemotionaltone(i.e.,above-medianpositivesentiment score),asmeasuredbyVADERsentimentanalysisbasedonAI disclosures. 20,5140.0300.17101 Notes:Weadd1tovariablesbeforetakingthelogtransformation.ThefocalvariableAIDisclosureandfour disclosure-relatedmoderatorsareonlyobservedinpost-policyperiods.Thefullsamplecomprises35,832projects acrosstheentireobservationperiod.Thepost-treatmentsubsamplecontains20,514projects. 4.4.Difference-in-DifferencesModel OurDIDmodelisspecifiedinEquation(1).Inthismodel,Yrepresentscrowdfunding outcomes,includingLogTotalPledge(totalfundingraised)andLogTotalBackers(totalbackers). Afterisabinaryvariablethatequalsoneifcampaignsendfollowingthepolicyimplementation date. 9 Wealsocontrolforproject-specificandcreator-specificcharacteristics(denotedasX). 9 BecausewedefinetheAfterdummybasedonthecampaignenddate,wecanstillidentifyitsmaineffecteven aftercontrollingforlaunch-monthfixedeffects.Asarobustnesscheck,wealsore-estimatethemodelbasedon 22 Thesecontrolsincludethelengthoftheprojectstoryandtitle,adummyvariableindicating whethertheprojectstoryincludesanyvideo,thefundinggoal(inU.S.dollars),thecampaign duration,campaigncurrencydummiestoaccountforsystematicdifferencesassociatedwithits originalsettlementcurrencies,andthecreator’snumberofpreviouslysuccessfulprojects,along withcategoryfixedeffects(FEs),launchmonthFEs,andday-of-weekFEs.Weclusterstandard errorsattheprojectcategorylevel. 퐀=퐀+퐀 1 퐀ᰀ퐀簀퐀䠀퐀㠀퐀+퐀 2 퐀퀀퐀ᰀ+퐀 3 퐀ᰀ퐀簀퐀䠀퐀㠀퐀×퐀퀀퐀ᰀ+퐀+퐀簀퐀萀ᰀ䠀퐀+ 퐀簀퐀㠀퐀ℎ퐀萀㠀퐀ℎ퐀+퐀簀䠀萀퀀頀퐀+퐀.(1) Apartfromestimatingtheeffectoftheplatform’sdisclosurepolicy,wealsodirectly measuretheeffectofAIdisclosureitselfbycomparingprojectsthatincludeanAI-disclosure statement,capturedbytheAIDisclosuredummy(equals1forprojectswithAIdisclosure),to thosewithoutdisclosureinthepost-policysample.Furthermore,toassesshowthetreatment effectofAIdisclosurevarieswiththesubstantiveorrhetoricalsignalsofaproject(denotedby Signal),weincorporatetheinteractionterm 퐀퀀퐀퐀萀퐀ᰀ퐀×퐀퐀㠀簀퐀 .Toensurethatour estimatedeffectsarenotconfoundedbysystematicdifferencesinthetypesofAIapplications disclosed,wefurtheraccountforpotentialtopic-levelheterogeneityinAIusagebyintroducing thefivemostcommontopicsidentifiedusingGPT-4o-miniascontrolvariables(denotedasZ) (Brynjolfssonetal.,2025). 10 Theremainingcontrolvariablesareidenticaltothoseusedin campaignlaunchdateratherthanenddatetoconstructtheAfterdummybyexcludingcampaignsthatoverlapwith thepolicyimplementationperiod.Theresultsremainconsistentwithourmainfindings(seesection7.2). 10 Herewefollowpriorworkemphasizingtheneedtoaccountforsemanticheterogeneityintext-basedmeasures (e.g.,Brynjolfssonetal.,2025).TodeterminethethematicdimensionofAIinvolvement,weimplementathree-step procedureusingGPT-4o-mini.First,wefeedeachAIdisclosuretothemodelandrequestashorttopicphrase(1–3 words)thatbestsummarizestheprimarypurposeorfunctionofAIintheproject.Second,weaggregateallphrases andpromptthemodeltoclusterthemintofivesemanticallydistinctgroups,eachwithaconciselabel,aone sentencefunctionaldefinition,andrepresentativeexamplephrasesfromthecorpus.Thisprocedureyieldsfive categories:AutomationandOptimization,CreativeGeneration,DataManagementandAnalysis,Supportand Assistance,andUserInteractionandPersonalization.Third,weclassifyeachdisclosureintothefivecategoriesby askingthemodelwhetherthedisclosurebelongstoanyofthem(multilabelassignmentpermitted).Wetheninclude theresultingtopicindicatorsasadditionalcontrolsinourregressionstoisolatedisclosurestyleeffectsfrom 23 Equation(1).ThefullspecificationispresentedinEquation(2). 퐀=퐀+퐀 1 퐀퀀퐀퐀萀퐀ᰀ퐀+퐀 2 퐀퀀퐀퐀萀퐀ᰀ퐀×퐀퐀㠀簀퐀+퐀+퐀+퐀簀퐀萀ᰀ䠀퐀+ 퐀簀퐀㠀퐀ℎ퐀萀㠀퐀ℎ퐀+퐀簀䠀萀퀀頀퐀+퐀.(2) 5.EMPIRICALRESULTS 5.1.ParallelTrendTest AkeyassumptionfortheDIDmodelistheparalleltrendassumption(Abadie,2005),which positsthatintheabsenceoftreatment,thetreatmentandcontrolgroupswouldhaveexhibited similaroutcometrajectories.Wetestthisassumptionbyinteractingthetreatmentgroupdummy withthemonthdummiesinthefollowingmodel: 퐀=퐀+ 퐀=−12 −2 퐀 퐀 ×퐀ᰀ퐀簀퐀䠀퐀㠀퐀×퐀萀㠀퐀ℎ 퐀+퐀 + 퐀=0 12 퐀 퐀 ×퐀ᰀ퐀簀퐀䠀퐀㠀퐀×퐀萀㠀퐀ℎ 퐀+퐀 +퐀 +퐀簀퐀萀ᰀ䠀퐀+퐀簀퐀㠀퐀ℎ퐀萀㠀퐀ℎ퐀+퐀簀䠀萀퀀頀퐀+퐀.(3) Here, 퐀 denotesthecrowdfundingoutcomes,and 퐀 representsthemonthwhentheAI disclosurepolicywasintroducedbytheplatform(i.e.,August2023). 퐀 퐀 indicatesthedifference betweenthetreatmentandcontrolgroupsinmonth퐀+퐀.Thelastpre-treatmentmonth(퐀=−1) issetasthebaseline.Figure2presentstheestimatedcoefficientswith95%confidenceintervals, usingLogTotalPledgeandLogTotalBackersasthedependentvariables,respectively.The estimatedleadcoefficientsarestatisticallyindistinguishablefromzero,indicatingthatthe treatmentandcontrolgroupsfollowedsimilartrendspriortotheplatform’simplementationof theAIdisclosurepolicy.ThisprovidesvalidsupportforourDIDdesign. differencesinthesubstantivenatureofAIapplications.Fullprompts,examples,andcodingrulesareprovidedin OnlineAppendixB3. 24 Figure2.ParallelTrendTest Followingthepolicychange,weobserveasignificantdivergencebetweenthesetwogroups, withAI-relatedprojectsexperiencinganotabledeclineincampaignengagementovertime.This suggeststhatAIdisclosuremayinfluenceprospectivebackers’evaluationsofprojects,thereby affectingcrowdfundingoutcomes.Inthenextsubsection,wewillformallyquantifytheseeffects usingourDIDframework. 5.2.AverageTreatmentEffect Table2showstheaveragetreatmenteffectofAIdisclosureoncrowdfundingsuccess. FollowingtheimplementationofthemandatoryAIdisclosurepolicy,projectsthatdiscloseAI involvementexperienceanapproximate39.8%declineinfundsraised(calculatedasexp(- 0.507)-1),andthetotalnumberofbackersdecreasesby23.9%(calculatedasexp(-0.273)-1). TheseresultssuggestthatmandatoryAIdisclosuresignificantlyreducesbackers’participation andfunding,posingsubstantialeconomicchallengesandnegativelyimpactingentrepreneurship. ThefindingspointtothebroadereconomicconsequencesofAIdisclosuremandatesand highlighttheimportanceofrefiningdisclosurestrategiestoalleviateadversemarketresponses. 25 Table2.ImpactofAIDisclosureonCrowdfundingOutcomes Dependentvariable: LogTotalPledgeLogTotalBackers (1)(2) Treatment0.221 *** 0.079 ** (0.068)(0.033) After-0.484 ** -0.297 *** (0.168)(0.084) Treatment×After-0.507 *** -0.273 *** (0.090)(0.039) ControlVariablesYESYES CategoryFEYESYES LaunchMonthFEYESYES Day-of-weekFEYESYES Observations35,83235,832 R-squared0.2490.305 Notes:a)Wecontrolforprojectandcreatorcharacteristics,includingtitleanddescriptionlength,fundingcurrency, fundinggoal,campaignduration,useofvideo,andthecreator’snumberofpriorsuccessfulprojects.Thedetailed coefficientsforthesecontrolsareomittedfromthetableforbrevity.b)BecausewedefinetheAfterdummybased onthecampaignenddate,wecanstillidentifyitsmaineffectevenaftercontrollingforlaunch-monthfixedeffects. Theresultsremainhighlyconsistentwhenwere-estimatethemodelusinglaunchdate,ratherthanenddate,to constructtheAfterdummy.c)Robuststandarderrorsclusteredatthecategorylevel.*p<0.1,**p<0.05,***p<0.01. 5.3.HeterogeneousTreatmentEffects WenextinvestigatehowtheimpactofAIdisclosureoncrowdfundingperformancevaries acrossprojectsthatadoptdifferentdisclosurestrategies.Tothisend,werestrictouranalysisto observationsfromthepost-policyperiodandcomparecrowdfundingperformanceacrossprojects withdifferingdisclosureapproachesrelativetothosewithoutAIdisclosure.Guidedbyour researchframework,weconsiderfourpotentialmoderatingvariablesthatcapturethesubstantive orrhetoricalsignalsconveyedthroughprojects’AIdisclosurestatements:HighAIInvolvement, HighExplicitness,HighAuthenticity,andHighPosEmotion.Thisapproachallowsustoidentify heterogeneouseffectsofAIdisclosureacrossdifferentdisclosurestrategiesandtouncover whichformsofAIcommunicationaremoreeffectiveinattractingbackersupport. 11 11 Asarobustnesscheck,wecontrolfortextualcharacteristicsofprojectdescriptionstoaddresspotential confounding.Theresultsremainhighlyconsistent(seeSection7.4). 26 Table3.ModeratingEffectsofSubstantiveandRhetoricalSignalsinAIDisclosure Dependentvariable: LogTotalPledge (1)(2)(3)(4)(5) AIDisclosure-1.349 *** -1.864 *** -1.984 *** -1.227 *** -1.716 *** (0.183)(0.225)(0.129)(0.149)(0.121) AIDisclosure×HighAIInvolvement-0.794 *** -0.791 *** (0.255)(0.199) AIDisclosure×HighAuthenticity0.979 *** 0.774 ** (0.206)(0.322) AIDisclosure×HighExplicitness1.201 *** 0.715 ** (0.197)(0.273) AIDisclosure×HighPosEmotion-0.614 *** -0.508 *** (0.140)(0.163) ControlVariablesYESYESYESYESYES AITopicControlsYESYESYESYESYES CategoryFEYESYESYESYESYES LaunchMonthFEYESYESYESYESYES Day-of-weekFEYESYESYESYESYES Observations20,51420,51420,51420,51420,514 R-squared0.2650.2660.2670.2650.269 Table3(continued) Dependentvariable: LogTotalBackers (1)(2)(3)(4)(5) AIDisclosure-0.816 *** -1.106 *** -1.194 *** -0.736 *** -1.027 *** (0.092)(0.109)(0.095)(0.091)(0.080) AIDisclosure×HighAIInvolvement-0.449 *** -0.433 *** (0.150)(0.133) AIDisclosure×HighExplicitness0.548 *** 0.405 ** (0.120)(0.149) AIDisclosure×HighAuthenticity0.723 *** 0.463 *** (0.113)(0.101) AIDisclosure×HighPosEmotion-0.375 *** -0.313 *** (0.071)(0.074) ControlVariablesYESYESYESYESYES AITopicControlsYESYESYESYESYES CategoryFEYESYESYESYESYES LaunchMonthFEYESYESYESYESYES Day-of-weekFEYESYESYESYESYES Observations20,51420,51420,51420,51420,514 R-squared0.3260.3270.3280.3260.329 Notes:a)WeincludethesameprojectandcreatorcontrolsasTable2.Coefficientsareomittedforbrevity.b)In additiontothecontrolvariablesincludedinthepreviousDIDmodel,wealsoaccountfortopic-levelheterogeneity inAIusagebyincorporatingdummyvariablesforthefivemostcommontopicsextractedfromtheAI-disclosure statements.ThesetopicscapturetheprimarypurposeorapplicationofAIdescribedineachdisclosure:Automation andOptimization,CreativeGeneration,DataManagementandAnalysis,SupportandAssistance,andUser InteractionandPersonalization.Includingthesetopicindicatorsensuresthatourestimatesreflecttheeffectsof disclosureitselfratherthandifferencesintheunderlyingAIapplications.Detailsontheextractionprocedureare providedinSection4.4andOnlineAppendixB3.c)Robuststandarderrorsclusteredatthecategorylevel.*p<0.1, **p<0.05,***p<0.01. 27 Table3summarizesthemoderatingeffectsofAIdisclosurestrategiesoncrowdfunding outcomes.Forthesubstantivesignal,thecoefficientoftheinteractiontermAIDisclosure× HighAIInvolvementissignificantlynegativeinColumns1and5,indicatingthatgreaterAI involvementamplifiesthenegativeeffectofAIdisclosureonfundingperformance.Thisfinding suggeststhatbackersrespondmorefavorablywhenAIismainlyusedtoassist,ratherthan replace,humaneffort. Turningtorhetoricalsignals,weobservethreekeyfindings.First,higherlevelsof explicitnessinAIdisclosuresareassociatedwithimprovedcrowdfundingperformance (Columns2and5).WhencreatorsclearlyarticulatehowAIisintegratedintotheproject,they reduceuncertaintyandenablepotentialbackerstobetterassessprojectfeasibilityandcreator commitment.Second,authenticityindisclosurecontributespositivelytofundingoutcomes (Columns3and5).Thissuggeststhat,languagethatconveyssincerityandinternalconsistency tendstoincreasebackerwillingnesstosupportthecampaign.Third,andsomewhatunexpectedly, greaterpositiveemotioninAIdisclosureisassociatedwithlowercrowdfundingperformance (Columns4and5).Althoughpositiveemotionalframingistypicallylinkedtofavorable outcomesinotherdisclosurecontexts,suchasfinancialreporting(Feldmanetal.,2010), prosocialbehaviors(Moran&Bagchi,2019),andlivestreamingshopping(Linetal.,2021),our findingdeviatesfromthispattern.Onepossibleexplanationisthat,giventhehighcomplexity andopacityofAItechnologies,overlyenthusiasticdescriptionsofAIusageandpotentialmay leadbackerstosuspect“AIwashing”, 12 whereexaggeratedclaimsaboutAIcapabilitiescreate doubtsaboutthecredibilityandprospectsofAI-involvedprojects. 12 FordiscussionsofAIwashing,seemoreonhttps://w.techtarget.com/whatis/feature/AI-washing-explained- Everything-you-need-to-know(accessedJune5,2025). 28 Onthewhole,whileAIdisclosurenegativelyimpactscampaignsuccessandtotalfunding raised,strategicframingcanmitigatetheseeffects,highlightingtheroleofsubstantiveand rhetoricalsignalsinshapingbackers’perceptions.Together,ourfindingsextendsignalingtheory byshowinghowAI-relatedtransparencyshapesbackers’responsesandhowstrategicframing canhelpcounterthedownsidesofdisclosure. 6.MECHANISMEXPLORATIONTHROUGHSCENARIO-BASEDEXPERIMENTS WhilepreviousanalysesshowthatAIdisclosurereducescrowdfundingperformance,the mechanismsunderlyingthiseffectremainunclear.Drawingonpriorresearchonsignalingin crowdfunding(Blanchardetal.,2023;Kimetal.,2022;Steigenberger&Wilhelm,2018)and workonAI-relatedsocialperception(Reifetal.,2025),weproposethatAIdisclosuremay changehowbackersevaluateboththeprojectanditscreator.Accordingly,wefocusontwo specificmechanisms:perceivedAIwashingandperceivedcreatorcompetence.First,backers mayinterpretAIdisclosureasanattempttoartificiallyinflatethetechnologicalsophisticationof acampaign,aphenomenonwerefertoasAIwashing.Suchperceptionscanraiseconcernsabout authenticityorstrategicobfuscation,therebyunderminingbackers’projectevaluation.These concernsareparticularlyimportantincrowdfundingsettings,wherebackersrelyheavilyon limited,informal,andoftenambiguoussignalstoguidetheirdecisions(Blanchardetal.,2023; Kimetal.,2022;Steigenberger&Wilhelm,2018).Second,becauseAIreducestheeffort requiredtocompletetasks,prospectivebackersmayinferthatcreatorswhorelyonAIareless capable,lessskillful,orpossessfewerrelevantresources.Thisinterpretationalignswithrecent evidenceshowingthatAIadoptersareoftenjudgedmoreharshlyorperceivedaslesscompetent thanthoseusingtraditionalmethods(Reifetal.,2025). 29 6.1.ExperimentDesignandDataCollection Toempiricallyexaminetheseproposedmechanisms,weconductaseriesofpre- registered, 13 scenario-basedonlineexperimentsonProlific(https://w.prolific.com/).These experimentsaredesignedtocausallyidentifyhowAIdisclosureshapesbackers’perceptionsof boththeprojectanditscreatorundervaryingmoderatorconditions.Specifically,weassess whetherAIdisclosureelevatesconcernsregardingAIwashinganddecreasesperceptionsof creatorcompetence,twoprocessesthatmayaccountforthenegativeperformanceeffects observedinourmainanalyses. Weimplementfourseparateexperiments,eachfocusedononemoderator:AIinvolvement, explicitness,authenticity,orpositiveemotion.Eachexperimentemploysabetween-subjects designinwhichparticipantsarerandomlyassignedtoeitherahighorlowlevelofthefocal moderator.AllexperimentalstimuliareadaptedfromactualKickstarterprojectsthatfeatureAI disclosures.Toensureprecisemanipulationofthefocaldimension,werefinedthedisclosure languageusingChatGPTtogenerateacounterfactualversion:whentheoriginalprojecttext reflectedahighlevelofthemoderator,werephrasedittoproducealow-levelcounterpart(and viceversa),whileminimizingchangestoothertextualfeatures. Eachexperimentfollowsaconsistentthree-stepprocedure.Step1:Participantsreviewbrief studyinstructions.Step2:TheyreadacrowdfundingprojectdescriptionthatincludesanAI disclosuretailoredtotheirassignedcondition.Thedisclosurewordingissystematicallyvariedto reflecttheintendedlevelofinvolvement,explicitness,authenticity,oremotionaltone.We administeramanipulationchecktoverifythattheperceivedlevelsofthemoderatoralignwith theassignedcondition.Step3:Participantsreporttheirpledgeintention,whichservesasour primarydependentvariable,alongwiththeirperceptionsoftheprojectanditscreator,including 13 Allhypotheses,materials,andanalysisplanswerepre-registered. 30 perceivedAIwashing(Haridasanetal.2015)andperceivedcreatorcompetence(Kuan&Chau, 2001),allmeasuredon7-pointLikertscales.Thisstructured,three-stepdesignallowsusto isolatehowspecificfeaturesofAIdisclosureinfluencebackers’psychologicalresponses(the proposedmediators)andtheirsupportintentions.Additionalproceduraldetailsandfull experimentalmaterialsareprovidedinOnlineAppendixE. 6.2.ManipulationCheck Toensuredataqualityandreducelinguisticorculturalconfounds,weappliedstringent prescreeningcriteriaonProlific.Participantswererequiredtohaveapriorapprovalrateof95– 100%,over50previoussubmissions,Englishastheirfirstlanguage,andbothcountryofbirth andnationalityrestrictedtotheUnitedStates.Thesecriteriahelpensurethatrespondentscan reliablyinterpretdisclosurelanguageandthatanyobservedeffectsarenotattributableto differencesincomprehensionorculturalbackground.Eachparticipantwaspermittedtotakepart inonlyoneofthefourexperimentstoavoidpotentialcross-exposureorcontaminationacross studies. Weconductedfourexperiments,eachmanipulatingonedisclosuresignalattwolevels(high vs.low).Intotal,weobtainedcompleteresponsesfrom318participantsacrossthefour experiments,withapproximately40participantsassignedtoeachcondition.Arandomization check(OnlineAppendixF)confirmsthatdemographicandpriorcrowdfundingexperience variablesarebalancedacrossconditions,supportingthevalidityofourrandomizationprocess. Amongtheserespondents,52failedthemanipulationcheckandwereexcludedfromfurther analysis.Oursubsequentregressionresultsarebasedontheremaining266participants. 14 14 Asarobustnesscheck,weconductedregressionanalysesusingthefullsample,includingparticipantswhofailed theattentioncheck.Resultsremainhighlyconsistent. 31 Table4presentstheresultsofourmanipulationchecks.Participantsevaluatedtheconstruct associatedwiththeexperimenttowhichtheywereassigned,indicatingwhetherthefocal attributeappearedatahighorlowlevelbasedontheirperceptions.Responsestothe manipulationcheckquestionwerecodedas1forhigh-levelperceivedvaluesand0forlow-level perceivedvaluesoftheconstruct.Acrossallfourexperiments,themanipulationcheckswere successful. Table4.ManipulationCheck Construct SamplesizeMean(SD)t-statistic HighconditionLowconditionHighconditionLowcondition AIInvolvement41380.88(0.33)0.13(0.34)9.84*** Explicitness42370.83(0.38)0.14(0.35)8.53*** Authenticity41390.88(0.33)0.23(0.43)7.60*** PositiveEmotion38420.87(0.34)0.26(0.45)6.78*** Note:*p<0.1,**p<0.05,***p<0.01 ForAIinvolvement,participantsinthehighconditionweresignificantlymorelikelyto reportahighlevelofAIinvolvement(M=0.88,SD=0.33)thanthoseinthelowcondition(M =0.13,SD=0.34;t=9.84,p<0.01),with87.8percentand86.8percentcorrectlyidentifying theintendedlevels,respectively.Forexplicitness,thehighconditionwasagainviewedasmore explicit(M=0.83,SD=0.38)thanthelowcondition(M=0.14,SD=0.35;t=8.53,p<0.01), and83.3percentand86.5percentofparticipantscorrectlyidentifiedtheintendedlevels.For authenticity,participantsinthehighconditionsimilarlyrateddisclosuresassignificantlymore authentic(M=0.88,SD=0.33)thanthoseinthelowcondition(M=0.23,SD=0.43;t=7.60,p <0.01),withcorrespondingcorrectidentificationratesof87.8percentand76.9percent.Finally, foremotionaltone,thehighconditionelicitedsignificantlygreaterperceivedexcessivepositive emotionality(M=0.87,SD=0.34)thanthelowcondition(M=0.26,SD=0.45;t=6.78,p< 0.01),withcorrectidentificationratesof86.8percentand73.8percent,respectively.Overall, 32 theseresultsconfirmthatparticipantsconsistentlydetectedtheintendedhighversuslow manipulationsacrossallconstructs. 6.3.ExperimentalResults Thefollowingresultsarebasedonthe266participantswhopassedthemanipulationcheck, including69intheAIinvolvementmodule,67intheexplicitnessmodule,66intheauthenticity module,and64intheemotionmodule. 6.3.1.ResultsonAIInvolvement AsillustratedinFigure3,projectsinvolvinghigherlevelsofAIelicitednoticeablydifferent reactionsfromparticipants.Thoseevaluatinghigh-AI-involvementprojectsarefarlesswillingto pledgethanthosereviewingprojectswithlimitedAIinvolvement(M=1.9vs.4.6,p<0.01). Theyalsojudgethecreatorassignificantlylesscompetent(M=2.6vs.5.8,p<0.01),whilethe increaseinperceivedAIwashing,thoughdirectionallyconsistent,isrelativelymodestand insignificant(M=3.0vs.2.5,n.s.). Figure3.EffectsofLowvs.HighAIInvolvement Wefurthertestthepotentialunderlyingmechanismwithasequentialmediationanalysis usingAIinvolvementastheindependentvariable,pledgeintentionasthedependentvariable, andperceivedAIwashingandperceivedcreatorcompetenceasserialmediators(Model6; Hayes,2017).AsshowninFigure4,highAIinvolvementproducesonlyaslightriseinAI 33 washing(b=0.57,n.s.),andthisshifthaslittlebearingonperceivedcreatorcompetence(b= −0.06,n.s.).Bycontrast,highAIinvolvementhasastrongandnegativedirecteffecton competence(b=−3.17,p<0.01),whichinturnisakeydriverofdecreasedpledgeintention(b =0.71,p<0.01).NeitherthedirectpathwayfromhighAIinvolvementtopledgeintention(b= −0.36,n.s.)northelinkfromAIwashingtopledgeintention(b=−0.08,n.s.)contributes meaningfully.BasedontheeffectsizessummarizedinTableG1ofOnlineAppendixG,the indirecteffectofhighAIinvolvementonpledgeintentionthroughperceivedcreatorcompetence is−2.263(bootstrapped95%CI:[-3.402,-1.314]),whilethetotaleffectofhighAIinvolvement onpledgeintentionis−2.689(p<0.01),indicatingthatdiminishedperceptionsofcreator competenceaccountforthemajorityofthenegativeimpactofhighAIinvolvementonbackers’ support. Figure4.TreatmentEffectofAIInvolvementonPledgeIntention,MediatedbyAI WashingandCreatorCompetence 6.3.2.ResultsonExplicitness Figure5showsthathighexplicitnessinAIdisclosureproducesnoticeablymorefavorable evaluations.Participantsevaluatingprojectswithhigh-explicitnessAIdisclosuresaremore willingtopledgethanthoseexposedtolow-explicitnessstatements(M=4.8vs.3.2,p<0.01). Theseparticipantsalsoformmorepositiveimpressionsofthecreator,reportinglower perceptionsofAIwashing(M=2.1vs.2.9,p<0.05)andhigherperceivedcreatorcompetence (M=5.8vs.3.8,p<0.01). 34 Figure5.EffectsofLowvs.HighPerceivedExplicitness EvidencefromFigure6furtherclarifieshowAIdisclosureexplicitnessshapesthese outcomes.Higherexplicitnessdirectlyincreasesperceivedcreatorcompetence(b=1.61,p< 0.01),whichinturnelevateswillingnesstopledge(b=0.76,p<0.01).Additionally,higher explicitnessalsoreducesperceptionsofAIwashing(b=−0.76,p<0.05),andlowerAIwashing istiedtohigherperceivedcompetence(b=−0.49,p<0.01). Figure6.TreatmentEffectofPerceivedExplicitnessonPledgeIntention,Mediatedby PerceivedAIWashingandPerceivedCreatorCompetence Bycontrast,thedirectpathwayfromexplicitnesstopledgeintention(b=0.05,n.s.)andthe linkfromAIwashingtopledgeintention(b=−0.07,n.s.)arenotsignificant.Thispattern indicatesthatexplicitnessstrengthenssupportprimarilybyenhancingperceivedcompetence. TableG2ofOnlineAppendixGfurtherclarifieshowAIdisclosureexplicitnessshapes theseoutcomes.HighexplicitnessintheAIdisclosureprimarilyincreasespledgeintentionby enhancingbackers’perceptionsofcreatorcompetence(indirecteffect=1.226,bootstrapped95% 35 CI:[0.584,1.991]),andthetotaleffectis1.612(p<0.01).Meanwhile,highexplicitnessalso exertsanadditionalpositiveeffectonpledgeintentionthroughasequentialmediationpathway (PerceivedHighExplicitness→PerceivedAIWashing→PerceivedCreatorCompetence→ PledgeIntention;indirecteffect=0.286,bootstrapped95%CI:[0.042,0.590]). 6.3.3.ResultsonAuthenticity Figure7presentstheeffectsofperceivedauthenticityonpledgeintentionandperceived creatorcompetence.Participantsexposedtohigh-authenticitydisclosuresexpresssignificantly greaterwillingnesstopledgethanthoseexposedtolow-authenticitydisclosures(M=4.3vs.2.7, p<0.01).TheyarealsolesslikelytoperceiveAIwashing(M=2.4vs.3.8,p<0.01)andjudge thecreatortobemorecompetent(M=5.1vs.3.7,p<0.01). Figure7.EffectsofLowvs.HighPerceivedAuthenticity ThemediationanalysisresultspresentedinFigure8furtherillustratehowauthenticity shapestheseoutcomes.HighauthenticityintheAIdisclosurestatementstronglyreduces perceptionsofAIwashing(b=−1.32,p<0.01),andlowerperceivedAIwashingisassociated withhigherperceivedcompetence(b=−0.67,p<0.01).Highauthenticityalsodirectlyimproves perceivedcreatorcompetence(b=0.86,p<0.01). 36 Figure8.TreatmentEffectofPerceivedAuthenticityonPledgeIntention,Mediatedby PerceivedAIWashingandPerceivedCreatorCompetence AssummarizedinTableG3ofOnlineAppendixG,neitherthedirecteffectofhigh authenticityonpledgeintention(b=0.469,n.s.)norsimplemediationthroughperceived competence(indirecteffect=0.371,bootstrapped95%CI:[-0.244,0.953])orAIwashing (indirecteffect=-0.051,bootstrapped95%CI:[-0.479,0.339])aloneprovessignificant.Instead, highauthenticityoperatesthroughsequentialmediation:reducingAIwashingperceptions,which enhancescompetenceperceptionsandincreasespledgeintention(PerceivedHighAuthenticity →PerceivedAIWashing→PerceivedCreatorCompetence→PledgeIntention;indirecteffect =0.761,bootstrapped95%CI:[0.335,1.351]). 6.3.4.ResultsonEmotionalTone Figure9showsthatexcessivelypositiveemotionaltoneleadstomarkedlylessfavorable reactionsfromparticipants.IndividualsexposedtoAIdisclosurestatementscontaininghighly positiveemotionaresignificantlylesswillingtopledgethanthoseencounteringmore moderatelywordedstatements(M=3.2vs.4.6,p<0.01).Theyalsoformmorenegative impressionsoftheprojectandcreator,reportinghigherperceptionsofAIwashing(M=5.5vs. 2.3,p<0.01)andlowerperceivedcreatorcompetence(M=3.6vs.5.4,p<0.01). 37 Figure9.EffectsofLowvs.HighPositiveEmotionalTone Figure10.TreatmentEffectofPerceivedHighlyPositiveEmotiononPledgeIntentionas MediatedbyPerceivedAIWashingandPerceivedCreatorCompetence Figure10andTableG4furtherillustratewhyexcessivepositiveemotionintheAI disclosurestatementcanproduceunintendednegativeeffectsonbackers’pledgeintention.In particular,excessivelypositiveemotiondoesnotdirectlyaffectpledgeintentionorperceived creatorcompetence;instead,itexertsasignificantindirecteffectthroughincreasedAIwashing perceptionsandtheresultingdecreaseinperceivedcreatorcompetence(PerceivedHighly PositiveEmotion→PerceivedAIWashing→PerceivedCreatorCompetence→Pledge Intention).Thissequentialindirecteffectissizable(b=−0.572,bootstrapped95%CI:[-1.558,- 0.054]).Thus,ratherthanattractingbackerengagement,overlypositiveemotionalframingcan backfirebytriggeringAIwashingconcernsthatunderminecompetenceperceptions. 38 7.ROBUSTNESSCHECKS Toverifytherobustnessofourresults,weconductseveraladditionalchecks,summarized inTable5.Specifically,were-estimatethemodelsusingatwo-stagepropensityscorematching (PSM;Rosenbaum&Rubin,1983)approach,reruntheanalysisusingcampaignlaunchdate insteadofenddatetodefinetreatmentstatus,andexaminewhetherthefindingsholdwith alternativedependentvariablesandadditionalcontrols.Wefurtherassesstherobustnessofthe emotionaltoneresultsusingtwoalternativeapproaches:(1)applyinganalternativesentiment analysismodel(SieBERT)and(2)re-performingthetextclassificationsusingClaude-Sonnet-4, anadvancedLLMdevelopedbyAnthropic.Theresultsremainhighlyconsistentacrossall specifications,lendingstrongsupporttoourmainfindings. Table5.SummaryofRobustnessChecks AnalysisObjectiveLocation AlternativeDIDsetupEmployingatwo-stagePSMapproachtoconstructanalternative sample 7.1 AlternativetimingdefinitionUsingthecampaignlaunchdatetoclassifythetreatmentperiod7.2 AlternativedependentvariablesRobustnessofresultstoalternativedependentvariable7.3 AdditionalcontrolvariablesRobustnessofresultstoadditionalcontrolvariables7.4 Alternativemethodforsentiment detection UsingSieBERTtoassesstheemotionaltoneofAIdisclosure statements 7.5 AlternativeLLMforLabelingRobustnesscheckusinglabelsgeneratedbyClaude-Sonnet-47.6 7.1.AlternativeDIDSetup Inourmainanalysis,weadoptatraditionalkeyword-basedapproachtoclassifyprojects intotreatmentandcontrolgroups.Thismethodallowsforastraightforwardandtransparentway toidentifyAI-relatedprojects.However,itidentifieswhetherprojectsmentionAIwithout capturingthedeploymentorsubstantiveuseofAI.Asarobustnesscheck,weemployatwo- stagePSMproceduretoidentifyprojectssimilartothosethatdiscloseAIusageandusethemas counterfactuals. 39 Figure11illustratesthematchingprocessandsampleconstruction.Inthefirststage,weuse one-to-onePSMtomatchprojectspostedbeforethedisclosurepolicyimplementationwiththose thatincludeAIdisclosurestatementsinthepost-policyperiod;thesepre-policymatchesserveas thepre-policycounterpartsfortheAI-disclosedprojects.Inthesecondstage,weconstructthe controlgroupbyselectingprojectsthatdonotdiscloseAIusageandapplyingone-to-onePSM againtoidentifytheircorrespondingpre-policycounterparts(Liangetal.,2025).Thistwo-stage procedureensuresthatboththetreatmentandcontrolgroupsarepairedwithcomparablepre- policyprojects,therebyenablingacrediblecounterfactualcomparison. Figure11.Two-stagePSM Itshouldbenotedthatourapproachreliesontheassumptionthatcreatorstruthfully disclosetheiruseofAIinthepost-policyperiod.ThisassumptionissupportedbyKickstarter’s enforcementpractices:theplatformmaysuspendprojectsthatfailtodiscloseAIusage,and creatorswhomisrepresentorattempttobypassthedisclosurerequirementmaybeprohibited fromsubmittingfutureprojects. 15 Kickstarteralsoemployshumanreviewerswhocanrequest additionalevidencewhenaproject’sAIusageisunclear,ratherthanrelyingsolelyoncreators’ self-reportedstatements.Importantly,anyremainingnon-compliance(e.g.,creatorswhouseAI 15 AsstatedpubliclybyKickstarter,“Ifcreatorsdon'tproperlydisclosetheiruseofAIduringthesubmissionprocess, Kickstartermaysuspendtheproject.ThosewhotrytobypassKickstarter'spoliciesorpurposefullymisrepresent theirprojectwon'tbeallowedtosubmit.”Seemoreonhttps://w.engadget.com/kickstarter-projects-will-soon- have-to-disclose-any-ai-use-145100394.htmlandhttps://w.neowin.net/news/kickstarter-to-require-ai- transparency-from-creatives/(accessedNovember17,2025). 40 butdonotdiscloseit)wouldbiasourestimatestowardzero,makingourtreatmenteffectmore conservative. Ourfinalsamplecomprises4,880projects,including2,440treatmentprojectsand2,440 matchedcontrolprojects.Figure12presentstheparalleltrendstestusingthetwo-stagePSM sample. 16 Ourresultsshowthatthetreatmentandcontrolgroupsfollowedparalleltrendspriorto thepolicyimplementationandexhibitedcomparablecrowdfundingperformanceduringthepre- policyperiod.Followingthepolicymandate,however,projectswithAIdisclosureexperienced significantlymorenegativecrowdfundingoutcomesthantheirmatchedcounterparts,consistent withourmainfindings. Figure12.ParallelTrendTestBasedonaTwo-StagePSMSample Table6reportstheaveragetreatmenteffectofAIdisclosureoncrowdfundingsuccessusing LLM-classifiedAIinvolvement.Table7reportsheterogeneoustreatmenteffects,whichare highlyconsistentwithourmainfindings.Theseresultsconfirmthatourfindingsarerobustto alternativemeasurementapproaches. Table6.MainEffectBasedonaTwo-StagePSMSample Dependentvariable: 16 NotreatedprojectendsinAugust2023becausethepolicybecameeffectiveonAugust29,2023,leavingonlya fewdaysinthatmonthforprojectstocloseunderthenewdisclosurerequirement.Sincecrowdfundingcampaigns typicallyspanseveralweeks,almostallprojectsaffectedbythemandateclosedinSeptember2023orlater.This timingexplainstheabsenceoftreatedprojectswithanenddateinAugust2023andisconsistentwiththe implementationwindowofthepolicyshock. 41 LogTotalPledgeLogTotalBackers (1)(2) Treatment-0.064-0.042 (0.105)(0.054) After-0.349-0.157 (0.341)(0.184) Treatment×After-1.003 *** -0.643 *** (0.200)(0.123) ControlVariablesYESYES CategoryFEYESYES LaunchMonthFEYESYES Day-of-weekFEYESYES Observations4,8804,880 R-squared0.2990.371 Notes:a)WeincludethesameprojectandcreatorcontrolsasTable2.Coefficientsareomittedforbrevity.b)See Table2fortheAfterdummydefinition.c)Robuststandarderrorsclusteredatthecategorylevel.*p<0.1,**p<0.05, ***p<0.01. Table7.HeterogeneousTreatmentEffectBasedonaTwo-StagePSMSample (withLogTotalPledgeasDependentVariable) Dependentvariable: LogTotalPledge (1)(2)(3)(4)(5) AIDisclosure-1.213 *** -1.625 *** -1.759 *** -1.084 *** -1.564 *** (0.164)(0.192)(0.156)(0.161)(0.152) AIDisclosure×HighAIInvolvement-0.592 ** -0.660 *** (0.244)(0.198) AIDisclosure×HighExplicitness0.846 *** 0.716 ** (0.209)(0.300) AIDisclosure×HighAuthenticity0.987 *** 0.579 * (0.219)(0.272) AIDisclosure×HighPosEmotion-0.505 *** -0.446 *** (0.125)(0.149) ControlVariablesYESYESYESYESYES AITopicControlsYESYESYESYESYES CategoryFEYESYESYESYESYES LaunchMonthFEYESYESYESYESYES Day-of-weekFEYESYESYESYESYES Observations2,5802,5802,5802,5802,580 R-squared0.3270.3320.3310.3260.341 Notes:a)WeincludethesameprojectandcreatorcontrolsasTable2.Coefficientsareomittedforbrevity.b)We includethesameAItopiccontrolsasTable3.Relatedvariabledefinitionsandcoefficientsareomittedforbrevity.c) Forbrevity,wereportonlytheresultswithLogTotalPledgeasthedependentvariable;resultswithLogTotalBackers asthedependentvariablearehighlyconsistentandreportedinOnlineAppendixH.d)Robuststandarderrors clusteredatthecategorylevel.*p<0.1,**p<0.05,***p<0.01. 7.2.AlternativeTimingDefinition Inourmainanalysis,werelyoneachproject’scampaignenddatetodeterminewhethera projectbelongstothepre-orpost-policyperiod,ensuringaclearmodelsetupandprecise 42 identificationofpolicyexposure.Tofurtherassessrobustness,weusethecampaignlaunchtime asthetemporalreferencepoint.Specifically,weincludeallprojectslaunchedbetweenAugust1, 2022,andAugust31,2024,asouralternativesample,excludingthoselaunchedbeforebut closingafterthepolicyintroduction.Thisprocedureyieldsafinalsampleof33,583projects, comprising5,138treatmentprojectsand28,445controlprojects.AsshowninTables8and9,the resultsbasedonthisalternativespecificationremainhighlyconsistentwithourmainfindings. Table8.MainEffectBasedonAlternativeTimeVariable Dependentvariable: LogTotalPledgeLogTotalBackers (1)(2) Treatment0.210 *** 0.069 * (0.067)(0.033) Treatment×After-0.500 *** -0.276 *** (0.094)(0.041) ControlVariablesYESYES CategoryFEYESYES LaunchMonthFEYESYES Day-of-weekFEYESYES Observations33,58333,583 R-squared0.2490.305 Notes:a)WeincludethesameprojectandcreatorcontrolsasTable2.Coefficientsareomittedforbrevity.b)See Table2fortheAfterdummydefinition.c)Robuststandarderrorsclusteredatthecategorylevel.*p<0.1,**p<0.05, ***p<0.01. Table9.HeterogeneousTreatmentEffectBasedonanAlternativeTimeVariable (withLogTotalPledgeasDependentVariable) Dependentvariable: LogTotalPledge (1)(2)(3)(4)(5) AIDisclosure-1.374 *** -1.883 *** -2.015 *** -1.254 *** -1.763 *** (0.201)(0.237)(0.120)(0.167)(0.117) AIDisclosure×HighAIInvolvement-0.746 ** -0.745 *** (0.261)(0.197) AIDisclosure×HighExplicitness0.992 *** 0.761 ** (0.194)(0.303) AIDisclosure×HighAuthenticity1.244 *** 0.775 *** (0.195)(0.261) AIDisclosure×HighPosEmotion-0.589 *** -0.489 *** (0.132)(0.150) ControlVariablesYESYESYESYESYES AITopicControlsYESYESYESYESYES CategoryFEYESYESYESYESYES LaunchMonthFEYESYESYESYESYES Day-of-weekFEYESYESYESYESYES Observations19,36019,36019,36019,36019,360 R-squared0.2690.2710.2700.2690.273 43 Notes:a)WeincludethesameprojectandcreatorcontrolsasTable2.Coefficientsareomittedforbrevity.b)We includethesameAItopiccontrolsasTable3.Relatedvariabledefinitionsandcoefficientsareomittedforbrevity.c) Forbrevity,wereportonlytheresultswithLogTotalPledgeasthedependentvariable;resultswithLogTotalBackers asthedependentvariablearehighlyconsistentandreportedinOnlineAppendixH.d)Robuststandarderrors clusteredatthecategorylevel.*p<0.1,**p<0.05,***p<0.01. 7.3.AlternativeDependentVariables Inthemainanalysis,weusethetotalamountraisedandthetotalnumberofbackersas primarydependentvariables.GivenKickstarter’sall-or-nothingfundingmodel,where entrepreneursreceivefundsonlyiftheyreachtheirpredeterminedfundinggoalwithinthe campaignduration(Calicetal.,2023),weexaminetwoadditionaloutcomevariables: CampaignSuccessandLogpledgedAdj.CampaignSuccessisabinaryindicatorofwhethera projectultimatelymetitsfundinggoal,andLogpledgedAdjcapturestherealizedfundingamount undertheall-or-nothingpolicy(Soublière&Gehman,2020).AsshowninTables10and11,the resultsbasedonthisalternativeoutcomemeasureremainconsistentwithourmainfindings. Table10.MainEffectBasedonAlternativeDependentVariables Dependentvariable: CampaignSuccessLogPledgedAdj Treatment0.0110.216 ** (0.010)(0.096) After-0.057 * -0.529 * (0.030)(0.263) Treatment×After-0.057 *** -0.634 *** (0.007)(0.080) ControlVariablesYESYES CategoryFEYESYES LaunchMonthFEYESYES Day-of-weekFEYESYES Observations35,83235,832 R-squared0.3090.264 Notes:a)WeincludethesameprojectandcreatorcontrolsasTable2.Coefficientsareomittedforbrevity.b)See Table2fortheAfterdummydefinition.c)Robuststandarderrorsclusteredatthecategorylevel.*p<0.1,**p<0.05, ***p<0.01. Table11.HeterogeneousTreatmentEffectsBasedonAlternativeDependentVariables Dependentvariable: CampaignSuccess (1)(2)(3)(4)(5) AIDisclosure-0.174 *** -0.221 *** -0.243 *** -0.162 *** -0.209 *** (0.019)(0.017)(0.024)(0.017)(0.021) AIDisclosure×HighAIInvolvement-0.093 ** -0.086 *** 44 (0.034)(0.029) AIDisclosure×HighExplicitness0.081 *** 0.055 ** (0.016)(0.020) AIDisclosure×HighAuthenticity0.128 *** 0.088 *** (0.022)(0.024) AIDisclosure×HighPosEmotion-0.065 *** -0.053 ** (0.020)(0.021) ControlVariablesYESYESYESYESYES AITopicControlsYESYESYESYESYES CategoryFEYESYESYESYESYES LaunchMonthFEYESYESYESYESYES Day-of-weekFEYESYESYESYESYES Observations20,51420,51420,51420,51420,514 R-squared0.3290.3290.3300.3290.330 Table11.(continued) Dependentvariable: LogPledgedAdj (1)(2)(3)(4)(5) AIDisclosure-1.879 *** -2.431 *** -2.647 *** -1.745 *** -2.258 *** (0.171)(0.182)(0.188)(0.130)(0.154) AIDisclosure×HighAIInvolvement-1.056 *** -1.001 *** (0.354)(0.310) AIDisclosure×HighExplicitness0.948 *** 0.677 ** (0.210)(0.242) AIDisclosure×HighAuthenticity1.409 *** 0.934 *** (0.229)(0.203) AIDisclosure×HighPosEmotion-0.754 *** -0.625 ** (0.199)(0.217) ControlVariablesYESYESYESYESYES AITopicControlsYESYESYESYESYES CategoryFEYESYESYESYESYES LaunchMonthFEYESYESYESYESYES Day-of-weekFEYESYESYESYESYES Observations20,51420,51420,51420,51420,514 R-squared0.2850.2850.2860.2850.287 Notes:a)WeincludethesameprojectandcreatorcontrolsasTable2.Coefficientsareomittedforbrevity.b)We includethesameAItopiccontrolsasTable3.Relatedvariabledefinitionsandcoefficientsareomittedforbrevity.c) Robuststandarderrorsclusteredatthecategorylevel.*p<0.1,**p<0.05,***p<0.01. 7.4.AdditionalControlVariables ToaccountforrhetoricalvariationinprojectdescriptionsbeyondtheAIdisclosuresection, weapplythesamecodingframeworktoextractstyle-relatedtextualfeaturesfromthefullproject narrative. 17 ThesemeasuresareincludedasadditionalcontrolstoisolatetheeffectsofAI 17 FollowingthesameapproachusedtoextracttextfeaturesintheAIdisclosure,weapplythesameGPT-4o-mini prompttomeasurethedescription’sexplicitnessandauthenticity,anduseVADERtoquantifyitsemotionaltone. 45 disclosurefromgeneralstylisticdifferencesacrosscampaigns.AsshowninTables12and13, afteraccountingforthesefactors,ourresultsremainqualitativelyunchanged. Table12.MainEffectwithAdditionalControlVariables Dependentvariable: LogTotalPledgeLogTotalBackers (1)(2) Treatment0.195 *** 0.062 (0.053)(0.035) After-0.443 ** -0.272 *** (0.161)(0.081) Treatment×After-0.472 *** -0.252 *** (0.088)(0.040) ControlVariablesYESYES AdditionalProjectDescriptionControlsYESYES CategoryFEYESYES LaunchMonthFEYESYES Day-of-weekFEYESYES Observations35,83235,832 R-squared0.2650.320 Notes:a)WeincludethesameprojectandcreatorcontrolsasTable2.Coefficientsareomittedforbrevity.b)See Table2fortheAfterdummydefinition.c)Toensurethatourestimatesarenotdrivenbystylisticvariationinhow projectcreatorswritetheirdescriptions,wealsoextractlinguisticfeaturesdirectlyfromtheprojectdescriptions. UsingthesameGPT-4o-minipromptsappliedtotheAI-disclosurestatements,weobtainmeasuresofexplicitness andauthenticity,anduseVADERtocomputesentimentscores.Thesevariablesareincludedasadditionalcontrols toseparatetheeffectofdisclosurestylefromgeneralwritingstyle.d)Robuststandarderrorsclusteredatthe categorylevel.*p<0.1,**p<0.05,***p<0.01. Table13.HeterogeneousTreatmentEffectswithAdditionalControlVariables (withLogTotalPledgeasDependentVariable) Dependentvariable: LogTotalPledge (1)(2)(3)(4)(5) AIDisclosure-1.288 *** -1.795 *** -1.930 *** -1.181 *** -1.660 *** (0.178)(0.221)(0.129)(0.149)(0.119) AIDisclosure×HighAIInvolvement-0.822 *** -0.810 *** (0.254)(0.197) AIDisclosure×HighExplicitness0.941 *** 0.731 ** (0.207)(0.321) AIDisclosure×HighAuthenticity1.203 *** 0.737 ** (0.190)(0.261) AIDisclosure×HighPosEmotion-0.592 *** -0.485 *** (0.144)(0.162) ControlVariablesYESYESYESYESYES AITopicControlsYESYESYESYESYES ProjectDescriptionControlsYESYESYESYESYES CategoryFEYESYESYESYESYES LaunchMonthFEYESYESYESYESYES Day-of-weekFEYESYESYESYESYES Observations20,51420,51420,51420,51420,514 R-squared0.2810.2820.2830.2810.284 46 Notes :a)WeincludethesameprojectandcreatorcontrolsasTable2.Coefficientsareomittedforbrevity.b)We includethesameAItopiccontrolsasTable3.Relatedvariabledefinitionsandcoefficientsareomittedforbrevity.c) WeincludethesamesetofadditionalprojectdescriptioncontrolsasdescribedinTable12. d )Forbrevity,wereport onlytheresultswithLogTotalPledgeasthedependentvariable;resultswithLogTotalBackersasthedependent variablearereportedinAppendixH.e)Robuststandarderrorsclusteredatthecategorylevel.*p<0.1,** p<0.05,***p<0.01. 7.5.AlternativeMethodforSentimentDetection Inthemainanalysis,weuseVADER,awidelyadoptedtoolintheliterature,todetect whetheraprojectexhibitsexcessivelypositivesentimentinitsAIdisclosurestatementbasedon theintensityofpositiveemotion.Inthissection,weemploySieBERT,anLLMfine-tunedfor sentimentanalysis(Brynjolfssonetal.,2025;Hartmannetal.,2023),toconstructabinary indicatorofwhetherthesentimentispositive.AsshowninTable14,wefindthatwhenusing thisalternativemeasure,theresultsremainconsistentwithourmainfindings. Table14.HeterogeneousTreatmentEffectBasedonanAlternativeSentimentTool (withLogTotalPledgeasDependentVariable) Dependentvariable: LogTotalPledge (1)(2)(3)(4)(5) AIDisclosure-1.349 *** -1.864 *** -1.984 *** -1.073 *** -1.494 *** (0.183)(0.225)(0.129)(0.303)(0.228) AIDisclosure×HighAIInvolvement-0.794 *** -0.795 *** (0.255)(0.208) AIDisclosure×HighExplicitness0.979 *** 0.844 *** (0.206)(0.284) AIDisclosure×HighAuthenticity1.201 *** 0.715 ** (0.197)(0.261) AIDisclosure×PosEmotion-0.491 ** -0.548 ** (0.222)(0.196) ControlVariablesYESYESYESYESYES AITopicControlsYESYESYESYESYES CategoryFEYESYESYESYESYES LaunchMonthFEYESYESYESYESYES Day-of-weekFEYESYESYESYESYES Observations20,51420,51420,51420,51420,514 R-squared0.2650.2660.2670.2640.268 Notes:a)ProjectandcreatorcontrolsasdescribedinTable2.b)AItopiccontrolsasdescribedinTable3.c)For brevity,wereportonlytheresultswithLogTotalPledgeasthedependentvariable;resultswithLogTotalBackersas thedependentvariablearehighlyconsistentandreportedinAppendixH.d)Robuststandarderrorsclusteredatthe categorylevel.*p<0.1,**p<0.05,***p<0.01. 47 7.6.AlternativeLLMforLabeling Inthemainanalysis,weuseGPT-4o-minitolabelthelevelsofexplicitness,authenticity, andAIinvolvementbasedontheAIdisclosurestatements.Toexaminewhetherourresultsare robusttothechoiceofLLM,wereplicatetheclassificationusingClaude-Sonnet-4,astate-of- the-artmodeldevelopedbyAnthropicthatemphasizesreasoningaccuracyandreliabilityintext understanding 18 .AssuggestedbyTable15,theresultsbasedonlabelsgeneratedbyClaude- Sonnet-4closelyalignwiththoseobtainedusingGPT-4o-mini,indicatingthatourfindingsare stableacrossdifferentLLMarchitecturesandproviders. Table15.HeterogeneousTreatmentEffectBasedonanAlternativeLLMforLabeling (withLogTotalPledgeasDependentVariable) Dependentvariable: LogTotalPledge (1)(2)(3)(4)(5) AIDisclosure-1.360 *** -2.248 *** -2.024 *** -1.227 *** -2.025 *** (0.192)(0.138)(0.164)(0.149)(0.172) AIDisclosure×HighAIInvolvement-1.037 *** -0.910 *** (0.272)(0.180) AIDisclosure×HighExplicitness1.474 *** 1.300 *** (0.172)(0.140) AIDisclosure×HighAuthenticity1.057 *** 0.307 * (0.239)(0.168) AIDisclosure×HighPosEmotion-0.614 *** -0.398 ** (0.140)(0.160) ControlVariablesYESYESYESYESYES AITopicControlsYESYESYESYESYES CategoryFEYESYESYESYESYES LaunchMonthFEYESYESYESYESYES Day-of-weekFEYESYESYESYESYES Observations20,51420,51420,51420,51420,514 R-squared0.2660.2670.2660.2650.269 Notes:a)WeincludethesameprojectandcreatorcontrolsasTable2.Coefficientsareomittedforbrevity.b)We includethesameAItopiccontrolsasTable3.Relatedvariabledefinitionsandcoefficientsareomittedforbrevity.c) Forbrevity,wereportonlytheresultswithLogTotalPledgeasthedependentvariable;resultswithLogTotalBackers asthedependentvariablearehighlyconsistentandreportedinAppendixH.d)Robuststandarderrorsclusteredat thecategorylevel.*p<0.1,**p<0.05,***p<0.01. 18 AnthropicNewsisavailableathttps://w.anthropic.com/news/claude-4(accessedDecember1,2025). 48 8.DISCUSSIONANDCONCLUSION 8.1.SummaryandDiscussion Thisstudyexamineshowplatform-mandatedAIdisclosureaffectscrowdfundingoutcomes andhowitseffectsvaryacrossdifferenttypesofsubstantiveandrhetoricalsignals.Usingthe policyshockasanexogenoussourceofvariation,weidentifythecausalimpactofAIdisclosure andanalyzethemechanismsthroughwhichitshapesbackers’evaluations. OurresultsshowthatdisclosingAIuselowersthelikelihoodofprojectsuccess,particularly whencreatorsrevealhighAIinvolvement.Thisfindingsuggeststhattransparencyabout extensiveAIusemayraisedoubtsaboutcreatorcompetence.Incontrast,disclosingpartialAI useappearstobalancetheappealofinnovationwiththereassuranceofhumancreationand oversight,resultinginweakernegativeeffects. WealsofindthathowcreatorsframeAIdisclosuremattersgreatly.Clearandspecific statements(logos)andanauthentic,credibletone(ethos)reduceinvestorskepticismand improvefundingperformance.However,disclosuresthatrelyonhighlypositiveoremotional language(pathos)oftenbackfire,astheymaybeinterpretedasexaggerationorhype.These resultshighlightthatrestrained,specific,factualcommunicationismoreeffectivethanoverly promotionalframingwhenaddressingbackers’concernsaboutcreatorcompetenceandpotential deceptiveAIpositioning. Ourexperimentalmechanismanalysisilluminateswhythesedifferentialeffectsemerge. Substantivesignals(AIinvolvement)operateprimarilythroughcompetenceperceptions:greater AIintegrationreducesperceivedcreatorexpertise.Thethreerhetoricalsignals,however,exhibit morecomplexcausalpathways.(1)HighexplicitnessinAIdisclosuremainlybolsterspledge intentionbydirectlyenhancingperceivedcreatorcompetence,whilealsohavinganindirect 49 sequentialmediationwhereinexplicitdisclosurefirstreducesAIwashingconcerns,which subsequentlyelevatescompetenceperceptions.(2)Highauthenticityinfluencespledgeintentions predominantlyviaanindirectserialpathway:authenticcommunicationalleviatesAIwashing concerns,whichthenelevatescompetencejudgments.(3)Excessivelypositiveemotionaltone triggersthereverseserialprocess:itfirstheightensAIwashingsuspicions,whichthenerode competenceperceptions.Thisrevealsacriticalasymmetry:promotionalrhetoricintendedto enhanceperceptionsactuallybackfiresbyactivatingsuspicionthatunderminesthevery judgmentsitaimstostrengthen.Overall,ourfindingsdemonstratethateffectiveAIdisclosure requiresnotmerelytransparencybutstrategiccalibrationofhowinformationisframed.AI disclosurefunctionsasastrategiccommunicationtoolthatcanbuildcredibilitywhenexecuted thoughtfullyorundermineitwhenhandledcarelessly. 8.2.ManagerialandPolicyImplications Ourfindingsofferseveralimplicationsforcreators,platforms,andpolicymakersnavigating thegrowingimportanceofAItransparencyindigitalmarkets.Forcreators,theresultshighlight thattransparencyshouldbemanagedstrategicallyratherthanmechanically.Thekeyisto communicateAIinvolvementasameansofenhancingefficiencyorcreativitywithout underminingthehumancontribution.DisclosinglimitedorassistiveAIuseallowscreatorsto signalbothinnovationandauthenticity.Creatorsshouldavoidambiguousoremotionally exaggeratedclaims,whichmayleadtoperceptionsofoverstatementoropportunism.Instead, theyshouldprovideconcrete,factualdisclosurespecifyingAIfunctions,explainingtheir benefits,andemphasizinghumanoversightinensuringqualityandoriginality.Platformscould helpbyofferingguidanceortemplatesforeffectivedisclosurestatements,makingiteasierfor creatorstocomplywithpolicywhilepreservingtrust. 50 Forplatforms,ourfindingsemphasizetheimportanceofbalancingregulationwithuser experience.PlatformscanplayanactiveroleinshapinghowAIdisclosureisperceivedby introducingstandardizedformats,explanatorycues,orcredibilityindicatorsthatclarifythe meaningofdisclosedAIuse.Forexample,avisualtagsystemcoulddistinguishbetweenAI- assistedandAI-generatedcontent,helpingbackersinterpretdisclosuresaccurately.Platforms couldalsouseautomatedtextscreeningtoflagvagueormisleadingdisclosuresandprompt revisionsbeforepublication.Moreover,platformsmayconsiderintegratingcreatoreducation modulesonAIcommunicationethicsorauthenticity,whichcouldhelpsustainbothcreator successandoverallmarketplaceintegrity.Beyondcompliance,platformsshouldrecognize disclosureaspartoftheinstitutionalarchitecturegoverningcreator-backerrelationshipsand sustainedlong-termplatformengagement. Forpolicymakers,ourevidencesuggeststhatmandatoryAIdisclosurepoliciescanhave mixedeffects.Whilesuchrulesimprovetransparencyandaccountability,theymay unintentionallydiscourageinnovationifaudiencesinterpretAIinvolvementasanegativequality signal.Policymakersshouldthereforecomplementdisclosurerequirementswithinitiativesthat improvepublicliteracyaboutAI,clarifyingthatpartialorassistiveusedoesnotnecessarily diminishcreativityorreliability.Atieredorcontextualizeddisclosureframeworkcouldalsobe useful,distinguishingdifferenttypesanddegreesofAIinvolvementratherthanimposinga single,uniformrequirement.Thisapproachwouldhelpminimizeoverreactiontobenignusesof AIandencouragetruthful,nuancedreporting.Regulatorsmightalsocollaboratewithplatforms tostandardizedefinitionsandbestpracticesacrossindustries,ensuringconsistencyandfairness (Hoetal.2024)whilemaintainingflexibilityforinnovation. 51 Takentogether,theseimplicationsunderscorethatAIdisclosureisnotonlyacompliance obligationbutalsoacommunicationchallenge.Creatorsandplatformsshouldtreatdisclosureas aformofbrandmanagementandtrust-building,whereclarity,honesty,andproportionalityare central.Policymakers,inturn,shoulddesignandenforcedisclosureframeworksthatencourage transparencywithoutcreatingunnecessaryfearorstigmatowardAIadoption.Effective implementationacrossallthreestakeholdergroupscanpromotebothaccountabilityand innovationinAI-mediateddigitalecosystems. 8.3.LimitationsandFutureResearch Ourstudyhasseverallimitations.First,ouranalysisisbasedonareward-based crowdfundingplatformforcreativeprojects,wherebackersaremotivatedprimarilybyobtaining productrewardsandsupportingventurestheyfindcompelling.Thedynamicsweobservemay thereforedifferinothercontexts,suchasdonation-basedplatformsorsocialmediacontent platforms.ThesalienceofcreatorcompetenceandAIwashingconcernsmayvaryacrossthese contexts.FutureresearchcouldcomparethesesettingstoassesswhethertheeffectsofAI disclosuregeneralizebeyondcreativecampaignsandhowcontextualfactorssuchasregulatory environmentsorculturalnormsshapeaudienceresponses. Moreover,ouranalysisfocusesonshort-termfundingoutcomes.Futureworkcould examinelonger-termconsequences,suchascreatorreputation,repeatfunding,orsustained communityengagement.AsAIdisclosurebecomesmorecommon,researcherscouldalso explorehowcreatorscalibratetheirsignalingstrategiesinresponsetoevolvingbacker expectations,andhowtheyintegrateAIcapabilitiesintoprojectdevelopment. [Duetospacelimitations,theonlineappendicesarehostedontheOpenScienceFramework(OSF)at https://osf.io/6bcva/overview?view_only=b378f0d57ec9412186292b41c8ea3d8c.] 52 REFERENCES Abadie,A.(2005).Semiparametricdifference-in-differencesestimators.ReviewofEconomic Studies,72(1).https://doi.org/10.1111/0034-6527.00321 Ahlers,G.K.C.,Cumming,D.,Günther,C.,&Schweizer,D.(2015).Signalinginequity crowdfunding.EntrepreneurshipTheoryandPractice,39(4),955–980. Anglin,A.H.,Short,J.C.,Wolfe,M.T.,Mckenny,A.F.,&Reid,S.W.(2025).Whoserhetoric matters?Socialresponsibilityrhetoricandthemoderatinginfluenceofsocialrolesin crowdfundingpitches.JournalofManagementStudies.https://doi.org/10.1111/joms.13262 Aristotle.(1991).OnRhetoric.KennedyGA,trans.(OxfordUniversityPress,NewYork). Babina,T.,Fedyk,A.,He,A.,&Hodson,J.(2024).Artificialintelligence,firmgrowth,and productinnovation.JournalofFinancialEconomics,151,103745. Baek,T.H.,Kim,J.,&Kim,J.H.(2024).EffectofdisclosingAI-generatedcontentonprosocial advertisingevaluation.InternationalJournalofAdvertising,1–22. Bail,C.A.(2024).CanGenerativeAIimprovesocialscience?ProceedingsoftheNational AcademyofSciences,121(21),e2314021121. Bauer,K.,Jussupow,E.,Heigl,R.,Vogt,B.,&Hinz,O.(2025).HowdisclosinggenerativeAI useimpactshumancreativeinvolvementinhuman-genaicollaboration.AvailableatSSRN. Benbya,H.,Pachidi,S.,&Jarvenpaa,S.(2021).Specialissueeditorial:Artificialintelligencein organizations:Implicationsforinformationsystemsresearch.JournaloftheAssociationfor InformationSystems,22(2),10. Bergh,D.D.,Connelly,B.L.,KetchenJr,D.J.,&Shannon,L.M.(2014).Signallingtheoryand equilibriuminstrategicmanagementresearch:Anassessmentandaresearchagenda. JournalofManagementStudies,51(8),1334–1360. Bhargava,H.K.,Qu,X.,Son,Y.,&Sun,D.(2024).Capitalizingonthemoment:Thestrategic roleofinformationdisclosureinonlinecrowdfunding.ProductionandOperations Management,10591478241231916. Bhattacharjee,A.,Moraffah,R.,Garland,J.,&Liu,H.(2024).Zero-shotLLM-guided counterfactualgeneration:AcasestudyonNLPmodelevaluation.2024IEEEInternational ConferenceonBigData(BigData),1243–1248. Blanchard,S.J.,Noseworthy,T.J.,Pancer,E.,&Poole,M.(2023).Extractionofvisual informationtopredictcrowdfundingsuccess.ProductionandOperationsManagement, 32(12),4172–4189. Bolinger,M.T.,Brownell,K.M.,&Covin,J.G.(2024).Keepingitreal:Howentrepreneurs effectivelydiscloserisk.StrategicEntrepreneurshipJournal,19(2),283-312. Brynjolfsson,E.,Li,D.,&Raymond,L.(2025).GenerativeAIatwork.TheQuarterlyJournal ofEconomics,140(2),889-942. Calic,G.,Arseneault,R.,&Ghasemaghaei,M.(2023).ThedarksideofMachiavellianrhetoric: Signalinginreward-basedcrowdfundingperformance.JournalofBusinessEthics,182(3), 875–896. Carney,S.,Riveros,I.,&Tully,S.(2024).MadewithAI:Consumerengagementwithmedia containingAIdisclosures.AvailableatSSRN. Cason,T.N.,Tabarrok,A.,&Zubrickas,R.(2025).Signalingquality:Howrefundbonusescan overcomeinformationasymmetriesincrowdfunding.ManagementScience,71(7),5933– 5947. 53 Chandler,J.A.,Anglin,A.H.,Kanwal,F.,&Short,J.C.(2024).Nopoliticsinfundingpitches: Anexpectancyviolationstheoryperspectiveofentrepreneurs’politicalexpressionsin crowdfunding.JournalofBusinessVenturing,39(1),106365. Chen,H.,Wang,P.,&Hao,S.(2025).AIinthespotlight:Theimpactofartificialintelligence disclosureonuserengagementinshort-formvideos.ComputersinHumanBehavior,162, 108448. Chen,P.-Y.,Hong,Y.,&Liu,Y.(2017).Thevalueofmultidimensionalratingsystems: Evidencefromanaturalexperimentandrandomizedexperiments.ManagementScience, 64(10),4629–4647. Chen,Y.,Wang,Q.,&Xie,J.(2011).Onlinesocialinteractions:Anaturalexperimentonword ofmouthversusobservationallearning.JournalofMarketingResearch,48(2),238–254. Chiarella,S.G.,Torromino,G.,Gagliardi,D.M.,Rossi,D.,Babiloni,F.,&Cartocci,G.(2022). Investigatingthenegativebiastowardsartificialintelligence:Effectsofpriorassignmentof AI-authorshipontheaestheticappreciationofabstractpaintings.ComputersinHuman Behavior,137,107406. Connelly,B.L.,Certo,S.T.,Ireland,R.D.,&Reutzel,C.R.(2011).Signalingtheory:Areview andassessment.JournalofManagement,37(1),39–67. Courtney,C.,Dutta,S.,&Li,Y.(2017).Resolvinginformationasymmetry:Signaling, endorsement,andcrowdfundingsuccess.EntrepreneurshipTheoryandPractice,41(2), 265–290. deKok,T.(2025).ChatGPTfortextualanalysis?HowtousegenerativeLLMsinaccounting research.ManagementScience,71(9),7888-7906. Dewan,S.,Ho,Y.-J.,&Ramaprasad,J.(2017).Popularityorproximity:Characterizingthe natureofsocialinfluenceinanonlinemusiccommunity.InformationSystemsResearch, 28(1),117–136. Donovan,J.(2021).Financialreportingandentrepreneurialfinance:Evidencefromequity crowdfunding.ManagementScience,67(11),7214–7237. ElAli,A.,Venkatraj,K.P.,Morosoli,S.,Naudts,L.,Helberger,N.,&Cesar,P.(2024). TransparentAIdisclosureobligations:Who,what,when,where,why,how.Extended AbstractsoftheCHIConferenceonHumanFactorsinComputingSystems,1–11. Eloundou,T.,Manning,S.,Mishkin,P.,&Rock,D.(2024).GPTsareGPTs:Labormarket impactpotentialofLLMs.Science,384(6702),1306–1308. Feldman,R.,Govindaraj,S.,Livnat,J.,&Segal,B.(2010).Management’stonechange,post earningsannouncementdriftandaccruals.ReviewofAccountingStudies,15,915–953. Fu,W.,Shang,G.,&Tong,X.(2025).Budgetdisclosureincrowdfunding:Information asymmetryandcosttransparency.Manufacturing&ServiceOperationsManagement,27(2), 659–678. Geva,H.,Barzilay,O.,Goldstein,A.,&Oestreicher-Singer,G.(2024).Equalopportunityforall? thelongtailofcrowdfunding:EvidencefromKickstarter.MISQuarterly,48(3),1223–1238. Gilardi,F.,Alizadeh,M.,&Kubli,M.(2023).ChatGPToutperformscrowdworkersfortext- annotationtasks.ProceedingsoftheNationalAcademyofSciences,120(30),e2305016120. Gottweis,H.(2017).Rhetoricinpolicymaking:Betweenlogos,ethos,andpathos.InHandbook ofPublicPolicyAnalysis.Routledge. Haridasan,A.C.,Xavier,N.,Krishnan,M.A.N.,&Glorindal,G.(2025).AIWashingin Marketing:Feasibility,Viability,andEthicalConstraints.InStrategicBlueprintsforAI- DrivenMarketingintheDigitalEra(p.495-528).IGIGlobalScientificPublishing. 54 Hartmann,J.,Heitmann,M.,Siebert,C.,&Schamp,C.(2023).Morethanafeeling:Accuracy andapplicationofsentimentanalysis.InternationalJournalofResearchinMarketing, 40(1),75–87. Hayes,A.F.(2017).Introductiontomediation,moderation,andconditionalprocessanalysis:A regression-basedapproach.Guilfordpublications. Higgins,C.,&Walker,R.(2012).Ethos,logos,pathos:Strategiesofpersuasionin social/environmentalreports.AccountingForum,36(3),194–208. Ho,Y.J.I.,Jabr,W.,&Zhang,Y.(2024).AIenforcement:ExaminingtheimpactofAIon judicialfairnessandpublicsafety.WaelandZhang,Yifan,AIEnforcement:Examiningthe ImpactofAIonJudicialFairnessandPublicSafety(November25,2024). Hou,J.,Liang,C.,&Chen,P.(2024).HowSociallyPerceivedThreatShapesPreventive BehaviorintheContextofCOVID-19.ProductionandOperationsManagement, 10591478241231864. Hou,J.,Wang,L.,Wang,G.,Wang,H.J.,&Yang,S.(2025).Thedouble-edgedrolesof generativeaiinthecreativeprocess:experimentsondesignwork.InformationSystems Research. Hu,T.,Wang,S.,Luo,W.,Zhang,M.,Huang,X.,Yan,Y.,Liu,R.,Ly,K.,Kacker,V.,&She,B. (2021).RevealingpublicopiniontowardsCOVID-19vaccineswithTwitterdatainthe UnitedStates:spatiotemporalperspective.JournalofMedicalInternetResearch,23(9), e30854. Huang,N.,Hong,Y.,&Burtch,G.(2017).Socialnetworkintegrationandusercontent generation:evidencefromnaturalexperiments.MISQuarterly,41(4),1035–1058. Hutto,C.,&Gilbert,E.(2014).Vader:Aparsimoniousrule-basedmodelforsentimentanalysis ofsocialmediatext.ProceedingsoftheInternationalAAAIConferenceonWebandSocial Media,8(1),216–225. Kim,K.,Park,J.,Pan,Y.,Zhang,K.,&Zhang,X.(2022).Riskdisclosureincrowdfunding. InformationSystemsResearch,33(3),1023–1041. Kuan,K.K.,&Chau,P.Y.(2001).Aperception-basedmodelforEDIadoptioninsmall businessesusingatechnology–organization–environmentframework.Information& Management,38(8),507-521. Li,J.J.,Chen,X.-P.,Kotha,S.,&Fisher,G.(2017).Catchingfireandspreadingit:Aglimpse intodisplayedentrepreneurialpassionincrowdfundingcampaigns.JournalofApplied Psychology,102(7),1075. Liang,C.,Hong,Y.,&Gu,B.(2025).Monitoringandthecoldstartproblemindigitalplatforms: Theoryandevidencefromonlinelabormarkets.InformationSystemsResearch,36(1),600– 620. Lin,M.,&Viswanathan,S.(2016).Homebiasinonlineinvestments:Anempiricalstudyofan onlinecrowdfundingmarket.ManagementScience,62(5),1393–1414. Lin,Y.,Yao,D.,&Chen,X.(2021).Happinessbegetsmoney:Emotionandengagementinlive streaming.JournalofMarketingResearch,58(3),417–438. Lou,B.,&Wu,L.(2021).AIondrugs:Canartificialintelligenceacceleratedrugdevelopment? Evidencefromalarge-scaleexaminationofbio-pharmafirms.MISQuarterly,45(3),1451– 1482. Lu,T.,&Zhang,Y.(2025).1+1>2?Information,humans,andmachines.InformationSystems Research,36(1),394-418. 55 Luo,X.,Tong,S.,Fang,Z.,&Qu,Z.(2019).Frontiers:Machinesvs.humans:Theimpactof artificialintelligencechatbotdisclosureoncustomerpurchases.MarketingScience,38(6), 937–947. Miric,M.,Jia,N.,&Huang,K.G.(2023).Usingsupervisedmachinelearningforlarge‐scale classificationinmanagementresearch:Thecaseforidentifyingartificialintelligencepatents. StrategicManagementJournal,44(2),491–519. Moradi,M.,&Badrinarayanan,V.(2021).Theeffectsofbrandprominenceandnarrative featuresoncrowdfundingsuccessforentrepreneurialaftermarketenterprises.Journalof BusinessResearch,124,286–298. Moradi,M.,Dass,M.,Arnett,D.,&Badrinarayanan,V.(2024).Thetime-varyingeffectsof rhetoricalsignalsincrowdfundingcampaigns.JournaloftheAcademyofMarketing Science,52(2),370–398. Moran,N.,&Bagchi,R.(2019).Thepowerofemotionalbenefits:Examiningtheroleofbenefit focusondonationbehavior.JournalofAdvertising,48(3),284–291. Munyon,T.P.,&Summers,J.K.(2024).Emotion,persuasion,andteamadaptation:Advancing theorythroughcinema.AcademyofManagementReview,49(1),182–196. Parhankangas,A.,&Renko,M.(2017).Linguisticstyleandcrowdfundingsuccessamongsocial andcommercialentrepreneurs.JournalofBusinessVenturing,32(2),215–236. Radoynovska,N.,&King,B.G.(2019).Towhomareyoutrue?Audienceperceptionsof authenticityinnascentcrowdfundingventures.OrganizationScience,30(4),781–802. Reif,J.A.,Larrick,R.P.,&Soll,J.B.(2025).Evidenceofasocialevaluationpenaltyforusing AI.ProceedingsoftheNationalAcademyofSciences,122(19),e2426766122. Renieris,E.M.,Kiron,D.,&Mills,S.(2024).ArtificialIntelligenceDisclosuresAreKeyto CustomerTrust.MITSloanManagementReview(Online),1–4. Rosenbaum,P.R.,&Rubin,D.B.(1983).Thecentralroleofthepropensityscorein observationalstudiesforcausaleffects.Biometrika,70(1),41–55. Sabzehzar,A.,Burtch,G.,Hong,Y.,&Raghu,T.S.(2023).Puttingreligiousbiasincontext: Howofflineandonlinecontextsshapereligiousbiasinonlineprosociallending.MIS Quarterly,47(1),33-62. Sahlman,W.A.(1990).Thestructureandgovernanceofventure-capitalorganizations.Journal ofFinancialEconomics,27(2),473–521. Schanke,S.,Burtch,G.,&Ray,G.(2024).DigitalLyrebirds:ExperimentalEvidenceThat Voice-BasedDeepFakesInfluenceTrust.ManagementScience.72(1),386-405. Song,F.,Agarwal,A.,&Wen,W.(2024).TheimpactofgenerativeAIoncollaborativeopen- sourcesoftwaredevelopment:EvidencefromGitHubCopilot.AvailableatSSRN. Soublière,J.-F.,&Gehman,J.(2020).Thelegitimacythresholdrevisited:Howpriorsuccesses andfailuresspillovertootherendeavorsonKickstarter.AcademyofManagementJournal, 63(2),472–502. SpenceM(1973)Jobmarketsignaling.TheQuarterlyJournalofEconomics.87(3),355–374. Steigenberger,N.,&Wilhelm,H.(2018).Extendingsignalingtheorytorhetoricalsignals: Evidencefromcrowdfunding.OrganizationScience,29(3),529–546. Stucki,I.,&Sager,F.(2018).Aristotelianframing:logos,ethos,pathosandtheuseofevidence inpolicyframes.PolicySciences,51(3),373–385. Suddaby,R.,&Greenwood,R.(2005).Rhetoricalstrategiesoflegitimacy.Administrative ScienceQuarterly,50(1),35–67. 56 Toff,B.,&Simon,F.M.(2024).“OrTheyCouldJustNotUseIt?”:TheDilemmaofAI DisclosureforAudienceTrustinNews.TheInternationalJournalofPress/Politics,30(4), 881-903. Tong,S.,Jia,N.,Luo,X.,&Fang,Z.(2021).TheJanusfaceofartificialintelligencefeedback: Deploymentversusdisclosureeffectsonemployeeperformance.StrategicManagement Journal,42(9),1600–1631. Truong,T.J.,Ludwig,S.,Mooi,E.,&Bove,L.(2022).Themarketvalueofrhetoricalsignalsin technologylicensingcontracts.IndustrialMarketingManagement,105,489–501. Vaara,E.,&Monin,P.(2010).Arecursiveperspectiveondiscursivelegitimationand organizationalactioninmergersandacquisitions.OrganizationScience,21(1),3–22. Vaara,E.,Sonenshein,S.,&Boje,D.(2016).Narrativesassourcesofstabilityandchangein organizations:Approachesanddirectionsforfutureresearch.AcademyofManagement Annals,10(1),495–560. Wei,X.,Fan,M.,You,W.,&Tan,Y.(2021).Anempiricalstudyofthedynamicanddifferential effectsofprefunding.ProductionandOperationsManagement,30(5),1331–1349. Wittenberg,C.,Epstein,Z.,Péloquin-Skulski,G.,Berinsky,A.J.,&Rand,D.G.(2025). LabelingAI-generatedmediaonline.PNASNexus,4(6),pgaf170. Wu,L.,Lou,B.,&Hitt,L.M.(2025).InnovationstrategyafterIPO:HowAIanalyticsspurs innovationafterIPO.ManagementScience,71(3),2360–2389. Xiang,D.,Zhang,L.,Tao,Q.,Wang,Y.,&Ma,S.(2019).Informationaloremotionalappealsin crowdfundingmessagestrategy:Anempiricalinvestigationofbackers’supportdecisions. JournaloftheAcademyofMarketingScience,47(6),1046–1063. Xiao,S.,Ho,Y.,&Che,H.(2021).Buildingthemomentum:Informationdisclosureandherding inonlinecrowdfunding.ProductionandOperationsManagement,30(9),3213–3230. Xu,Y.,Dai,H.,&Yan,W.(2024).Identitydisclosureandanthropomorphisminvoicechatbot design:afieldexperiment.ManagementScience,72(1),223-241. Yang,M.,Zheng,Z.,&Mookerjee,V.(2022).Howmuchisfinancialadviceworth?the transparency-revenuetensioninsocialtrading.ManagementScience,68(7),5252-5268. Yang,Y.,Wang,H.J.,&Wang,G.(2016).Understandingcrowdfundingprocesses:adynamic evaluationandsimulationapproach.JournalofElectronicCommerceResearch,17(1),47. Yazdani,E.,Chakravarty,A.,&Inman,J.(2025).(Mis)alignmentbetweenfacialandtextual emotionsanditseffectsondonorsengagementbehaviorinonlinecrowdsourcingplatforms. JournaloftheAcademyofMarketingScience,53(4),968–988. Yin,Y.,Jia,N.,&Wakslak,C.J.(2024).AIcanhelppeoplefeelheard,butanAIlabel diminishesthisimpact.ProceedingsoftheNationalAcademyofSciences,121(14), e2319112121. Youssef,P.,Seifert,C.,&Schlötterer,J.(2024).LLMsforgeneratingandevaluating counterfactuals:Acomprehensivestudy.FindingsoftheAssociationforComputational Linguistics:EMNLP2024,14809–14824. Zhang,X.andZhou,M.andLee,GM,HowDoesAI-GeneratedVoiceAffectOnlineVideo Creation?EvidencefromTikTok(2024).Availableathttps://ssrn.com/abstract=4676705. Zhang,X.,Lee,M.,Zhou,M.,&Lee,G.M.(2025).LargeLanguageModelsintheInstitutional Press:InvestigatingtheEffectsonNewsProductionandConsumption.AvailableatSSRN 5357471.Availableathttps://papers.ssrn.com/sol3/papers.cfm?abstract_id=5357471.