Anda di halaman 1dari 39

CounterspeechonTwitter:AFieldStudy

AreportforPublicSafetyCanadaundertheKanishkaProjectby

SusanBenesch1,4,DerekRuths2,KellyPDillon3,HajiMohammadSaleem2,andLucasWright4

BerkmanKleinCenterforInternet&Society,HarvardUniversity,Massachusetts
SchoolofComputerScience,McGillUniversity,Montreal
3
SchoolofCommunication,TheOhioStateUniversity,Ohio
4
DangerousSpeechProject,WashingtonD.C.

Contactauthors:
SusanBeneschsbenesch@cyber.law.harvard.edu

DerekRuthsderek.ruths@mcgill.ca

TableofContents
Introduction

ReviewofExistingResearch

HateSpeechResearch

CounterspeechResearch

Findings

Methods

Coding

11

MethodologicalChallenges

11

Definitions

13

Vectors

14

OnetoOne

14

OnetoMany

16

ManytoOne

16

ManytoMany

17

Taxonomyofcounterspeech

18

Presentingfactstocorrectmisstatementsormisperceptions

18

Pointingouthypocrisyorcontradictions

20

Warningofofflineoronlineconsequences

20

Affiliation

22

Denouncinghatefulordangerousspeech

24

Visualcommunication

25

Humor

27

Tone

30

TypesofResponses

31

Apologyorrecanting

31

Deletion

32

Morehatefulspeech

32

Civilconversation

32

Morecounterspeech

33

ConclusionsandIdeasforFutureResearch

33

WorksCited

35

Introduction
Ashatefulandextremistcontentproliferatesonline,counterspeechisgainingcurrencyasa
meansofdiminishingit.1Nowonder:counterspeechdoesntimpingeonfreedomof
expressionandcanbepracticedbyalmostanyone,requiringneitherlawnorinstitutions.The
ideathatmorespeechisaremedyforharmfulspeechhasbeenfamiliarinliberal
democraticthoughtatleastsinceU.S.SupremeCourtJusticeLouisBrandeisdeclareditin
1927.2Wearestillwithoutevidence,however,thatcounterspeechactuallydiminishes
harmfulspeechoritseffects.Thiswouldbeveryhardtomeasureofflinebutisabiteasier
online,wherespeechandresponsestoitarerecorded.Inthispaperwemakeamodeststart.
Specificallyweask:inwhatformsandcircumstancesdoescounterspeechwhichwedefine
asadirectresponsetohatefulordangerousspeechfavorablyinfluencediscourseand
perhapsevenbehavior?
Toourknowledge,thisisthefirststudyofInternetusers(notagovernmentororganization)
counterspeakingspontaneouslyonapublicplatformlikeTwitter.Ourfindingsarequalitative
andanecdotal,sincereliablequantitativedetectionofhatefulspeechorcounterspeechisa
problemyettobefullysolvedduetothewidevariationsinlanguageemployed,althoughwe
madeprogress,asreportedinanearlierpaperthatwaspartofthisproject(Saleem,Dillon,
Benesch,&Ruths,2016).
Wehaveidentifiedfourcategoriesorvectorsineachofwhichcounterspeechfunctions
quitedifferently,ashatefulspeechalsodoes:onetooneexchanges,manytoone,
onetomany,andmanytomany.Wealsopresentasetofcounterspeechstrategies
extrapolatedfromourdata,withexamplesoftweetsthatillustratethosestrategiesatwork,
andsuggestionsforwhichonesmaybesuccessful.

ReviewofExistingResearch
Herewereviewpreviousstudiesofonlinecounterspeechandalsoofthespeechitisintended
tocounter.Bothfieldsarestillyoung,soliteratureislimited.Hatespeechandextremist
speechonlinehavebothbeenstudiedbymultipleauthors.3Asweandothershavenoted
elsewhere,hatespeechisawidelyusedtermbutthereisnoconsensusonitsdefinition,in
laworthesocialsciences(Benesch,2014Mendel,2012).Ingeneralitmeansdenigratinga
personorpeoplebasedonhis/her/theirmembershipinagroup,butstudiesofhatespeech,
likelaws,haveemployedavarietyofdefinitions.Inourownstudy,weavoidedtheterm
1

ForexampleFacebooksCOOSherylSandbergsaidinaJanuary2016speechattheWorldEconomicForumin
Davos,Switzerland,Thebestantidotetobadspeechisgoodspeech.Thebestantidotetohateistolerance.
Amplifyingcounterspeechtothespeechthatsperpetratinghateis,wethink,byfarthebestanswer.
http://www.recode.net/2016/1/21/11588986/wanttocombathatespeechonfacebooktryalikeattacksaysco
o
2
JusticeBrandeisassertedinhisconcurringopinioninWhitneyvCaliforniathattoexposefalsehoodand
fallaciesandtoaverttheevil,theremedyismorespeech,notenforcedsilence(WhitneyvCalifornia,
1927,U.S.SupremeCourt,p.377)
3
Cyberbullyingalsohasanextensiveliterature,butisoutsidethescopeofthisproject.

becauseofitsambiguityandthedifficultyincodingconsistentlyforit,insteadusinghateful
speechanddangerousspeech,asexplainedbelow.
HateSpeechResearch
Thereissubstantialliteraturedescribingandanalyzinghatespeech,onlineandoffline.
Withinit,wehavedrawnonrecentanalyticalstudiessuchasTheContentandContextof
HateSpeech(Herz&Molnar,2012)andDanielleCitronsHateCrimesinCyberspace
(2014),e.g.forideasonhowtodrawdefinitionalboundaries.Otherworksusefullydescribe
thebreadthandvarietyofhatespeechonline,includingViralHate(Foxman&Wolf,2013),
ClickHeretoEndHate(2014)bythehumanrightsorganizationMuslimAdvocates,and
IntermediariesandHateSpeech:FosteringDigitalCitizenshipforOurInformationAge
(Citron&Norton,2011).
Anotherlineofresearchonhatespeechconsistsofeffortstodetectitautomatically,online.
Thisismuchmoredifficultthanitmightatfirstseem,notonlybecauseofthedefinitional
problem,butbecauseofthewidevarietyofwaysinwhichpeopleexpresshatredand
contempt:theydonotalwaysuseslurs,anddonotalwaysuseslurshatefully.Onestudy
attemptedtocodeforracismusingaNaiveBayesclassifier(Kwok&Wang,2013).This
workconfirmedthedefinitionalchallengeofhatespeechbyshowingannotatorscouldagree
only33%ofthetimeonwhethertextscontainedhatespeech.Anotherconsideredthe
problemofdetectingantiSemiticcommentsinYahoonewsgroupsusingsupportvector
machines(Warner&Hirschberg,2012).Notably,thetrainingdataforthisclassifierwere
handcoded.Manuallyannotatingtrainingdataisthemostfamiliarwaytoproduceadataset
fortraining,butitadmitsthepotentialforhardtotracebiasinthespeechultimatelydetected.
Athirdstudyusedalinguisticrulebasedapproachontweetsthathadbeencollectedusing
offensivekeywords(Xiangetal.,2012).Likemanuallyannotateddata,keywordbaseddata
hassignificantbiasingeffectsaswell.
Dinakar,Reichart,andLieberman(2011)exploredthedetectionofharassing(notnecessarily
hatespeech)commentsonYouTube.Theyidentifiedsexuality,race&culture,and
intelligenceasthethreemajorthemesinuserharassment.Fromtheperformanceof
classificationsystemsthatincludenaiveBayesandSVM,theyconcludedthatlabelspecific
classifiersaremoreeffectivethanmulticlassclassifiersatdetectingharassment.Theyfurther
observedthatblatantharassmentiseasiertomodelthanexpressionsinvolvingsarcasmand
euphemismwhichoftenappearinhatespeechaswell.Theauthorshadtoperformmanual
codingtogeneratesmalltrainingandtestingdatasets.
Inanothereffortatautomatedcodingofhatespeech,theFloatingSheepscholarscollective,
whichspecializesinmappinggeolocateddata,publishedamapin2013,identifyingthe
ostensiblymosthatefulplacesintheUnitedStateswithgeolocatedtweetsthatusedoffensive
words(Stevens,2013b).Theblogwascriticizedforextrapolatingbroadlyfromasmalland
arguablyskewedsample(tweetsarenotallgeolocated,notallhatredisexpressedwithslurs,
andsomeformsofhate,suchasmisogyny,wereomitted).Somereaderswereoffended,also,

thattheblogpublishedthehatefultermsforwhichtheysearched.Theresearchersresponded
byclarifyingtheirgoalsandmethods(Stevens,2013a).InanotheronlineeffortaCanadian
NGO,theSentinelProject,launchedasitein2013calledHateBase,whichinvitesInternet
userstoaddtoalistofslursandinsultingwordsinmanylanguages.HateBasedoesnot,as
farasweknow,attemptcodingordetection.
Althoughwedidnotfocusonextremistorterroristspeechinthisproject,wedrewon
relevantliteraturesuchasWhoMattersOnline:MeasuringInfluence,EvaluatingContent
andCounteringViolentExtremisminOnlineSocialNetworks(Berger&Strathearn,2013),
TheISISTwitterCensus:DefiningandDescribingthePopulationofISISSupporterson
Twitter(Berger&Morgan,2015),andTheIslamicState(Barrett,2014).
CounterspeechResearch
Thereisasmallbodyofexistingresearchoncounterspeechwhichissummarizedinsome
detailhereasitshouldbeausefulcontributionnootherreviewseemstobeavailable.Nearly
alloftheworksofarfocusesoncounteringeitherhatefulspeechorextremistspeech.Many
authorsobserve,aswedo,thatcounterspeechvariesgreatly,intoneandincommunicative
strategies,andseveralpapersoffercategoriesofcounterspeech,providingusefulframeworks
forobservationandfurtherstudy.Itisimportanttonote,though,thatsomeauthorsusethe
termcounterspeechbroadly,torefertoanycontentthatcountersorcontradictshatefulor
extremistcontentgenerallynotnecessarilyinresponsetoanyparticularstatementorspeech
act.Weusecounterspeechtoreferonlytoaresponse.Forexpressionthatcountersanother
narrativeorviewingeneralamuchbroadercategorythatcouldincludeformsofeducation,
propaganda,andpublicinformationweusethetermcounternarrative.
Theliteraturethusfarcontainsverylimitedevidencethatcounterspeech(oranyparticular
varietyofit)iseffectiveinchangingthought,onlinespeech,orofflinebehavior.Nonetheless,
someauthorsoffersuggestionsonwhichformsofcounterspeechmaybemostusefulfor
counteringhatredand/orextremism.
Amongfeaturesortechniquesofcounterspeechthatotherresearchershaveidentified,the
U.S.basedAntiDefamationLeagueidentifiesthreeinitsBestPracticesforRespondingto
Cyberhate(2014).Theyare:respondingtotheoriginalspeaker(theuserwhoproducedthe
hatefulordangerousspeech),usingcomedyorsatire,andcorrectingfalsehoods.Inastudyof
counterspeechonFacebook,theBritishthinktankDemosdistinguishesamongconstructive
counterspeechandnonconstructivecounterhate(respondinghatefullytohatred),fact
checking,constructivediscussion,satiricaloppositionalpages,andseriousoppositionalpages
(Bartlett&KrasodomskiJones,2015).TheDemosreportalsooffersthreetypesofmetrics
formeasuringthesuccessofcounterspeech:quantitativemetrics,suchasuserengagement
andthespreadofcounterspeechcontentmetrics,suchassentimentanalysisandrealworld
metricsthatmeasurethelongtermimpactsofonlinecounterspeechonofflinebehavior.

TheDemosreportalsousefullysoughttounderstandwhethercounterspeechreachespeople
whodonotalreadyagreewithit,sinceifitdoesnot,itcannotbeeffectiveatchangingminds
ordiscoursenorms.DemosfoundthatpopulistrightwingFacebookpagesgotmorelikesand
commentsfrompeoplewhodidnotalreadylikethosepages,thanpagesthatcountered
hatefulpopulistrightwingviews.(Bartlett&KrasodomskiJones,2015).Thissuggests
populistrightwingcontentreachedabroaderaudience,andhighlightstheimportanceof
disseminatingcounterspeechcontenttothosewhodonotalreadyagreewithit.
Researchintotheusefulnessofcounterspeechagainsthatefulspeechhasalsofoundthatthe
identityofthemessengerandthecontentofthemessagearecrucial.Theconflictprevention
activistRachelBrown,inherDefusingHate:aStrategicCommunicationGuideto
CounteractDangerousSpeech,highlightstheimportanceofthemediumorplatformfor
determininghowandwhenacounterspeechmessagewillreachtheaudience.Brownalso
notesthataparticularspeakercanmakeinformationseemmorereliable,andthecontentofa
messagecaninfluencethebehavioroftheaudience(Brown,2016).Sherecommendsa
centrallyplannedstrategywithacleargoalinwhichmedium,speaker,andcontentare
carefullytailoredtothetargetaudience.
Researchonusingspeechandnarrativetocounterviolentextremism(CVE)makesmany
similarpoints.AreportongovernmentantiextremistprogramspublishedbytheInstitutefor
StrategicDialogue(ISD),aLondonbasedthinktankthatstudiesviolentextremismand
methodstodiminishit,identifiesfivepossiblemethodsforcounterspeakers:toerodethe
originalspeakersintellectualframework,tomockorridicule,tohighlightthenegative
impactofextremistspeech,todemonstrateinconsistenciesintheextremistarguments,andto
questiontheeffectivenessofextremistsatachievingtheirgoals(Briggs&Feve,2013).The
reportmakestheimportantpointthatthegoalforallofthesepossibilitiesisnottowinthe
argumentbuttodelegitimizeextremistnarratives.Inawhitepaperonpreventingextremism,
theQuilliamFoundation,aBritishcounterextremismthinktank,distinguishescounterspeech
thattriestodiscreditextremistcontent,fromcounterspeechthatoffersalternative,more
positivenarratives,andcounterspeechintendedtoinformandprovidetransparencyarounda
topic(Saltman&Russell,2014).
AnothercommoncontentioninCVEresearchisthatextremismcanbemoreeffectively
underminedbyaddingfavorable,peacefulmessages(suchascounterspeech)thanbydeleting
extremistcontent(Saltman&Russell,2014Hussain&Saltman,2014Bipartisan,2012).
Forexample,suchresearchassertsthatattemptstoremove,monitor,orbanextremistcontent
arestymiedbythechallengesofdefiningsuchcontent.Also,contentislikelytoreappear
elsewhereontheInternetafterithasbeendeleted(Saltman&Russell,2014).Counterspeech
andcounternarrativesseemtobemoreeffectivesincetheycanbetterreachaudiencesthatare
vulnerabletoextremism.Onlinecounterspeechalsoallowsfortrackingviewership,and
coordinatingwithofflinecounternarrativecampaigns(Hussain&Saltman,2014).Finally,

counterspeechcanbeconductedbycivilsocietyorplatforms,whichareusuallymore
credible,asspeakers,thangovernments.
ThisleadstoanothercommonfeatureofCVEresearchthatisalsofoundintheantihate
speechliteraturetheimportanceofthemessenger,message,andmessaging(tone)in
constructingeffectivecounterspeech(Saltman&Russell,2014).TheISDreporton
governmentantiextremistcounternarrativesassertsthattheidentityofthemessengeris
criticalwhencounteringonlineextremistspeech(Briggs&Feve,2013).Accordingtothe
authors,tobeaneffectivemessengeronemusthavebothcredibilityintheeyesofthe
audienceandtheexpertisenecessarytomarketastrongnarrative(Briggs&Feve,2013).
BergerandStrathearnfoundinastudyononlineextremistnarrativesthatwhitenationalist
extremistsweremorelikelytoengagewithnewpeopleonlinewhoseemtobeconservative,
thanwithliberals.Theauthorsthereforerecommendthatinterventionstoreachwhite
nationalistextremistsbedirectedbyconservatives(Berger&Strathearn,2013).Similarly,a
studyofonetooneinterventionswithindividualsconsideredsusceptibletoviolentextremist
speechfoundthatthetoneandcontentoftheinterventionmessagessenttosuchindividuals
greatlyinfluencedtheresponserate(Frenett&Dow,2015).Responseratestointerventions
onlinewerehighestwhenthetoneofthemessagewascasualratherthanantagonisticandthe
contentincludedoffersofassistanceandpersonalstoriesratherthanpersonalquestionsor
attemptstohighlightthepossibleconsequencesofextremism.
Thesepreviousstudiesoncounteringhatredorextremismonlinefocusontwoquitedifferent
kindsofefforts:thoseintendedtoinfluenceindividuals,andthosethatseektoinfluence(or
shiftthenormsof)groups.Inthefirstcategory,thereareseveraldocumentedprojectsto
persuadeindividualstochangetheirviews,especiallyincaseswherethetargetisconsidered
vulnerabletoextremistspeechandlikelytoengageinviolence.AsFrenettandDow(2015)
demonstrate,suchcounterspeechisusuallydeliveredbyoneperson(evenwhenactingaspart
ofalargereffort),directlyaddressinganotherindividual.Counternarrativecampaignsare
alsoorganizedandplanned,buttheyattempttochangethenormsorbeliefsofalarger
audiencethroughchangesinnarratives.Suchcampaignscanbeorchestratedbygovernments,
civilsociety,privatecorporations,orindividuals.Theyusuallytakeplaceattheonetomany
ormanytomanylevel.
Nopreviouspublishedstudyseemstohavedonewhatweattemptinthepresentwork:to
collectandexamineorganic,spontaneouscounterspeech,whichappearswhenInternetusers
choose,foravarietyofreasonsandinavarietyofways,toreplytootheruserswhoexpress
hatredorextremistviews.Spontaneouseffortsarevariedbutmuchmorenumerous,and
thereforecanhavealarger,moresustainedimpactthanprogrammedinterventionsand
counternarrativecampaigns.

Findings
ThisworkwasinspiredbyseveralexamplesofspontaneouscounterspeechonTwitterthat
seemtohavebeensuccessful,inthesensethattheywerefollowedbyapologiesorother

signsoffavorableimpactontheoriginalaccount(theaccounttowhichthecounterspeech
responded).Weweresurprisedtocomeacrossthis,andresolvedtodiscoverandstudymore
examples,hopingtolearnsomethingabouthowandwhytheyworked.Ineachoftheoriginal
cases(observedamongTwitterusersinKenya,France,andtheUnitedStates)4,therewasa
threestepprocesswhichweeventuallynicknamedagoldenconversation,sincewecameto
seesuchaTwitterexchangeasavaluableprize.
Thethreestepsarethese:oneaccountproducedatweetthatwashateful,racist,orthat
includedathreattocommitviolence.Oneormoreaccountsapparentlyunknowntothe
originaluserrepliedbyrebukingthatuser,whothenrecantedorapologized.Sucha
responsewassurprising,andbeliedthefamiliarmaximdontfeedthetrolls.Iffeedingcan
sometimesbesuccessful,whatwouldbethebestrhetoricalfood?
CounterspeechsurgesonTwitter,wefound,whenpeoplewhodontknowoneanotherand
whodisagreedeeplymeetandargueonline,oftenbecausetheyareinterestedinthesame
offlineevent.Usuallypeopleofverydifferentconvictionsdonotexchangethemsince,like
mostInternetusers,theyspendmostoftheirtimeinlikemindedsilos,readingcontentthey
agreewithandremainingwiththeirowntribe(Conoveretal,2011,Lewandowskyetal.,
2012,p.111,Zuckerman,2013,Anderson&Raine,2010,p.18).
Wefoundbothextensivehatefulspeechandcounterspeechinresponsetocertaineventsthat
provokedwidespreademotionalresponseanddebate,suchastheBaltimoreprotestsofApril
2015,theJune2015U.S.SupremeCourtdecisiononsamesexmarriage,ortheNovember
2015Parisnightclubmassacre.Ineachofthese,newhashtagswereformedandpeoplewith
opposingviewsmetonthosehashtags.Insomecases,differenthashtagsexpressed
opposingviews,suchas#lovewinsand#1m1wor#stopislamand#loveislam.Somepeople
tweetedonthehashtagofanopposingopinion,apparentlyinordertoreachandengagewith
thosewithopposingopinions,andinsomecases,peopletweetedusingtwoopposing
hashtags.
Someconversationsbetweenpeopleofopposingviewsarefarfromgolden,unfortunately.
Notonlydotheyfailatimprovingdiscourseonline,butcounterspeechissometimesas
vitriolicandhostileasthespeechtowhichitresponds.Stillworse,somerespondtohatefulor
simplyoffensivespeechwiththreats,dogpiling(manypeoplesendinghostileresponsestoan
individualinashortperiodoftime),andharassment.Thesearenottobeconfusedwith
counterspeech.

WefirstobservedsuccessfulcounterspeechonTwitterinKenya2013,duringaprojecttostudyhatefuland
dangerousspeechonlineduringthemonthsleadingtoapresidentialelection.SeeiHubResearch,UmatiFinal
Report,Sept.2012May2013availableathttp://research.ihub.co.ke/uploads/2013/june/1372415606__936.pdf.
Subsequently,weworkedwithTwitterstafftofindotherexamplesofsuccessfulcounterspeech,includingin
responsetotheselectionofNinaDavuluriasMissAmerica2014,andinresponsetohomophobiaonTwitterin
France.

Methods
Alldatainthestudyaretweets,whenpossiblecapturedinconversations,ortweetsin
whichuserstaggedotherusers,respondingtotheirtweets.Virtuallyallofourdataarein
English,sincewerealizedthatcodingandanalysismustbedonebynativespeakerswhoare
familiarwithconnotationsandculturalcontext.Thereareexceptionsamongourdata,suchas
tweetswhosemeaningisconveyedalmostentirelywithimages.

Figure1.Datacollectionworkflow
Twitterisadynamicplatformwhereusersoftenreacttoofflineevents.Twitterallows
researchersaccesstoaportionofpublictweetsthroughthesearchAPIandthestreaming
API,andwedevelopeduserfriendlycommandlinetoolsforthosesearches.
Bothtoolsrequireahashtagorkeywordtoinitiateasessionandthesearchtoolreturns
relevanttweetsfromthepastweek,whilethestreamingtoolstartstherealtimecollectionof
relevanttweetsfromthatpointforward.Theoutputofbothtoolsisacollectionoftweetsas
JSONobjectsinatextfile.
Wealsohavedevelopedadditionaltoolstoaidtheanalysisofthecollecteddata.They
includeaparsingtoolthatconvertsthelistofJSONsinthetextfiletoanExcelfriendly
tabularformatandahashtagtoolthatliststhemostfrequenthashtagsinthecollectionof
tweets.Together,thiscanopyoftoolsallowedustoeffectivelycollecttweetsaroundevents
astheywerehappeningandpresentthedatainaformatthatisconducivetoouranalysis.
Weoftenfocusedonofflineeventsthatledtoconversationwithracist,sexist,orbigoted
overtones,suchasterroristattacksinEurope(e.g.Paris,Brussels),protestsagainstU.S.
police(e.g.Ferguson,Baltimore,McKinney,Waco),SouthCarolinasdisplayofthe
Confederateflag,andsamesexmarriage.Relevanthashtagslike#baltimoreor#mizzou
appearedonTwitter.Multiplehashtagscanappearinasingletweet,soeachtimewebegan
collectingtweetswithahashtag,wealsosearchedforthetop10hashtagsinthatsetof
tweets,oftenleadingustotherelevantsubconversations.
ToillustratehowsubhashtagsgaveshapetoconversationsonTwitteronevents,aword
cloudgeneratedfromtophashtagsin#stopislamdatasetappearsinFigure2.Tweets
containing#stopislamwerecollectedforthreedayssinceitappearedasatophashtagin
tweetscontaining#Brussels,aftertheterroristattacksthereinMarch2016.Userswere
taggingtheirtweetswithhatefulhashtagslike#banislamor#muzzlemuzzies,andalsowith
counterspeechhashtagslike#loveislamor#muslimsforpeace.Generally,tweetscontaining
competinghashtagswereeitherhateful,#StopIslamdg#LoveIslamMAKEPEOPLEGET

10

RIDOFTHEMUSLIMFAITHETERNALHELL,ortheywerecounterspeechLet'smake
#loveislamtrendingthen#stopislam.terrorismhasnoreligion!

Figure2.Wordcloudillustratingtophashtagsin#stopislamdataset,March2016
Datasetswerecategorizedbasedontopic,inanefforttoincludeavarietyofhateful,
dangerous,andcounterspeechexamples(seeTable1).Withinthedatasets,weusedsortand
filterfunctionstofindcounterspeech.SearchingTwitterswebsitebyusernames,phrases,or
hashtagsalsoyieldedconversations.
Oursearchtoolscapturedataembeddedintweetssuchasthenumberoffollowersauserhas,
whetheratweetisaretweet,andhowmanytimesatweethasbeenseenatthetimeof
collection.Sortingbyretweetstatus(yes/no)didnotyieldusefulinformationasitonly
indicatedwhetheratweethadbeenretweeted.Retweetsdonotnecessarilymeanagreement
withtheoriginalsentiment,astheRTfunctionisusedtoinform,commenton,validate,agree
ordisagreewith,orreferothersto,theoriginaltweet(boyd,Golder,&Lotan,2010).Since
retweetersintentwasoftenhardtodetermine,thisvariablewasnotparticularlyuseful.
Ifatweetappearedtobecounterspeech,theoriginalconversationwassearchedforonthe
usersTwitterfeed.Sortingbytweetdatewashelpfultoidentifytrendsonhashtags,or
multipleretweets,orevenreclaimingofhashtags,asinthecaseof#KillAllMuslimsinthe
wakeoftheJanuary2015CharlieHebdoshooting.Thelargestnumberoftweetsonthat
hashtaginourdatawasonthedayofthemassacre,January7.Afterthatthenumberof
tweetsdiminished,andthenroseagain,thistimedominatedbycounterspeech.

Table1.Datasummary
Topic

No.ofdatasets

No.oftweets

Counterphrases(e.g.Imsorry)

68,036

LGBTQ(e.g.ObergefellvHodges,U.S.Supreme
Courtrulingonsamesexmarriage)

18

55,234

Islamophobia(e.g.BrusselsandParisattacks)

23

91,847

Race(e.g.Baltimoreprotests)

37

305,273

Refugee/immigration(e.g.Syrianrefugees,anti
PEGIDAprotests)

22,577

11

Transgender(e.g.NC,Targetbathrooms)

13,341

Misogyny(e.g.#getbackinthekitchen)

4345

Trump(i.e.U.S./Mexicowall,Pope,DavidDuke)

23

498,932

Coding
Apreliminarycodebookwasdevelopedtobetteridentifyhateful,dangerous,and
counterspeech.Sixteenvariableswereincludedinthepreliminarycodebook:relevance,
trend,direction,tone,hatespeech,dangerousspeech,counterspeech,defensivespeech,
conciliatoryspeech,endorsementorreferenceofstereotypes,endorsementorreferenceof
generalizations,counterspeecheducation,counterspeechramifications,hypocrisy,media,
mediatype,anddeletedpost.
Somevariableswereforreferenceororganizationalpurposes,likerelevance,trend,direction,
media,mediatype(ifany),andwhetherthepostwasdeleted.Thedirectionalityofatweet
(retweet,originaltext,andresponsive)helpeddeterminewhereconversationsmightbe
hiding.Codingtweetsashateful,dangerous,defensive,orconciliatoryhelpedorganizethe
dataquicklyforfurtherprobing.Theremainingvariablesinthecodebookassistedinshaping
thecontextofthetweetsandhelpedframethediscussionofthestrategiesobserved.
MethodologicalChallenges
Usingtweetsastheunitofanalysis,itwasdifficulttodeterminewhichpartofaconversation
atweetformed,ifany.Twitterexchangesareasynchronous,anduserscancomeuponatweet
hours,days,ormonthsafterithasbeenposted.Partsofaconversationmayhavebeen
deleted,oraccountssuspended,orrepliescouldappearoutoforderassideconversations
developed.Counterspeechtweetsmaynotappearsequentiallylikeascriptedornatural
conversation,whichcanmakeitdifficulttoreassembleconversations.Also,sincewehad
accesstoastreamofonlyalimitedsamplingoftweets(usually10percent),wewereoften
unabletofindallpartsofaconversation.However,someTwitterconversationsonhateful
trendinghashtagswerethesubjectofnewsreportsonline(suchas#FuckPhyllisand
#KillAllMuslims),whichmadefindingconversationsonthesehashtagsmucheasier.
Lackofcontextoftenmadeitchallengingtoidentifybothhatefulandcounterspeechtweets
bytheircontentalone.Forexample,thetweet@realDonaldTrumpweneedtostart
catapultingillegalsbacktoMexico,Theyllstopcomingmighthavebeenpostedbya
DonaldTrumpsupporteroranopponentsarcasticallyimitatingasupporter.Inthiscase,we
checkedtheaccountandrealizedthatthetweetwasnotsarcastic.
Itwasalsodifficulttofindhatefulanddangerousspeechtransmittedasweblinksorimages,
sinceourtools,whicharetextbased,couldnotsearchthatcontent.Forexample,onetweet
collectedinresponsetotheBrusselsattacks(March2016)containedonlyamentionof

12

anotheruserandalink.Thelinkledtoanothertweet,reproducingathirdone.Theimageis
atFigure3.

Figure3.Imageembeddedinatweet,illustratingdifficultyoftextonlyanalysis
Interraterreliability(IRR),ortheextenttowhichcodersagreewhenclassifyingthesame
content,isnotoriouslydifficulttoachieve.InsocialsciencesIRRisgenerallyacceptableif
thealphaisabove.80,orthecodersagreeatleast80%ofthetime.ToimproveourIRR,we
codedsetsof10tweetstogether.Then,totestapreliminarycodebook,100randomtweets
fromthe#KillAllMuslimsdatasetwerecodedbyoneoftheauthorsandaresearchassistant,
resultinginCohensalphasbetween.42and.85(eachvariablehasitsownreliabilityscore).
Another33tweetscodedusingthesamecodebookproducedaCohensKappaof.37and.82.
Itwasespeciallydifficulttoreachagreementonwhethertweetsinvolvedgeneralizationsor
stereotypes,andwhethertheoveralltonewastoxic,hateful,ordangerous.Thedatacouldbe
analyzedatvariouslevels:theconversations(affordingcontext),theindividualtweets(with
givendata),orthecomponentsofthetweet(ifincludingvisuals,retweets,mentions,or
combinations).Sincethecodebookwaspreliminary,broadcodingandresearchassistant
trainingdidnotgofurther.Futureprojectsshoulddetermine(a)levelofanalysisand(b)a
solidcodebookorqualitativememoingsystem(Glaser&Strauss,1967).
Definitions
Counterspeechisdefinedinthisstudyasaresponsethattakesissuewithhateful,harmful,or
extremistcontent.Counterspeechisconsideredsuccessfulwhenitisfollowedbyafavorable
responsefromtheInternetuseroruserstowhomthecounterspeechwasdirected.
Threeclarifyingremarksareinorderhere.First,itisimpossibletobesure,inan
observationalstudylikethisone,thatonlinecounterspeechwasinfacttheinspirationfora
favorableresponsesuchasanapology.Second,afavorableresponsemayhavean
unfavorableeffect,ultimately,whereusersfeeloverwhelmedbyangryorabundant

13

counterspeechandapologizeordeletetheiraccounts,simplytogetittostop.Third,
counterspeechmayhaveapositiveimpactonInternetusersotherthantheoriginalspeaker,
andwherethataudienceislargeasitoftenisonTwitter,theimpactofcounterspeechonthe
audiencecanbemuchgreaterandmoreinfluential,thanononeoriginalaccount.This
suggestsaseconddefinitionofsuccess:counterspeechthatfavorablyshiftstheexpression
and/orbeliefsofInternetusers.Pastresearchhasfocusedoncounternarrativecampaignsasa
necessaryvehicleforinfluencingalargeraudience,butwebelievethatnormscanbeshifted
whenanaudiencewitnessespublic,direct,andorganiccounterspeechthattheywouldnot
havebeenexposedtoifithadbeenpublishedaspartofacounternarrativecampaignor
privatelyinapersonalintervention.Sincewedidnotattempttomeasuresuchaneffect,we
definesuccessintermsofafavorableresponsefromtheoriginalaccount.Thatmayormay
notindicateadurablechangeinspeechorbehaviorthisisdiscussedfurtherintheTypesof
Responsessectionbelow.
Asnotedabove,weavoidedthetermhatespeech.Itsdefinitionsspecifydifferenttypesof
groupmarkers,including(ornot),religion,ethnicity,nationality,sexualorientation,
disability,bodytype,age,orgender,amongothers.Wheretodrawthelinebetweenhate
speechandspeechthatismerelyoffensivedependssomuchonprevailingsocialnorms,
context,andindividualandcollectiveinterpretationthatitisverydifficulttocode
consistently.Arecentstudydemonstratedamere33%agreementbetweencodersfrom
differentraces,whenaskedtoidentifyracisttweets(Kwok&Wang,2013).Anadditional
ambiguityinthetermhatespeechisinthewordhateitselfwhichmightrefertothe
speaker/authorshatred,orhis/herdesiretomakethetargetsofthespeechfeelhated,or
desiretomakeothershatethetarget(s),ortheapparentcapacityofthespeechtoincrease
hatred.Mostoften,itimpliesthatthespeakerfeelsorintendshatred.
Sinceitisdifficultorimpossibletoknowaspeakersintent,especiallyfromatweet,we
proposethetermhatefulspeechtofocusontheexpressionofhate.Thisisanuancedbut
usefuldistinctionsinceexpressioniseasiertodetectthanintent,andmorelikelytobelinked
tolanguagescapacitytocauseharm.Hatefulspeech,then,isspeechwhichcontainsan
expressionofhatredonthepartofthespeaker/author,againstapersonorpeople,basedon
theirgroupidentity.Wesearchedforcounterspeechinresponsetothiskindofcontent,and
alsoinresponsetotweetsthatconstitutedangerousspeech,whichwehavedefinedin
previousworkasspeechthatcaninspireorcatalyzeintergroupviolence(Benesch,2014).
Intenselyhatefulandevenviolentspeechisnothardtofindonline,includingonTwitter,but
mostofoursuccessfulcounterspeechresponded,notsurprisingly,totweetsthatwerenot
amongthemostvitriolic.
Vectors
Weobservedimportantdistinctionsincounterspeechconversations,accordingtothenumber
ofparticipantsineachstageorsideofanexchange.Aconversationbetweentwopeopleis
qualitativelydifferent,forexample,fromacaseinwhichmanypeoplerespondtoonetweet.

14

Notsurprisingly,theeffectivenessofcounterspeechstrategiesvarieswiththesedistinctions
aswell.
Weidentifiedfourtypesofcounterspeechexchanges,orvectors,apparentlyforthefirsttime
intheliterature:onetoone,onetomany,manytoone,andmanytomany.Literatureand
policydiscussionsonhatespeechusuallyconflatethesecategoriesalso,althoughthe
distinctionsareequallyimportantinthatcontext:harassmentofanindividualbyagroupof
people,forexample,isverydifferentinnatureandlikelyconsequences,fromhatreddirected
byonepersonagainstanentireracialorethnicgroup.Below,wediscusseachofthefour
vectorsinthecontextofcounterspeech.
OnetoOne
Someofthemoststrikingexamplesofgoldenconversations,orcasesinwhich
counterspeechseemstoconvinceapersontostopspeakinghatefully,areconversations
betweenonlytwopeople.Wefoundextendedexchangesbetweenanoriginalspeakeranda
counterspeakerwhoengagesrepeatedly,eveninthefaceofapparentlyimplacableresistance
andastreamofhatefultweets.
Wheresomeoneseemsfirmlycommittednotonlytohatefulideologybuttodeclaringit
publicly,wewouldnotexpectcounterspeechtoswaythatperson.Yetinsomecases,it
apparentlyhasandhasevenhelpedtobringaboutlastingchangeinbeliefs,notonlyin
speech.Inthesecases,weobservecounterspeechstrategiesincluding:anempathicand/or
kindtone,theuseofimages,andtheuseofhumor.Thissuccessfulcounterspeechusually
doesnotlabeltheoriginalspeakerashatefulorracist,butdoesidentifythespeechassuch.
Asuccessfulcounterspeechconversation,inwhichnearlyallofthesestrategieswereused,
tookplaceonJanuary19,2015,MartinLutherKingDayintheUnitedStates.Itbeganwith
thistweet:
InhonorofMLKdaytoday,I'mtakingavowtousetheword"nigger"asmanytimesas
possibleandinthemostinappropriatetimes
Awriterandactivist5discoveredthetweetandrespondedwithanantihatredquotefrom
King:
Letnomanpullyousolowastohatehim.MartinLutherKingJr.
Theoriginalaccountrespondedohsoyou'reoneofthoseniggerloverstoo?towhichthe
counterspeakerreplied,againcitingKing,hatedestroysthehater.
Theoriginalspeakerrespondedwithegregiousracistattacksonthecounterspeaker(theonly
fuckingfoolisyourinferiorniggerass)whorepliedwithquoteafterantihatredquotefrom
5

Weveerredonthesideofnotrevealingtheidentitiesofpeopleinthecaseswedescribeinorderto
protectthemandtopreservetheirprivacy.Wevemadeexceptions,however,forpublicfigureslike
DonaldTrumpand/orthosewhohavealreadychosentodiscussthecasepublicly.

15

MartinLutherKing,untilshewrote,Iwishyoupeaceandloveandfreedomfromthehatred
thathurtsyourheart.Herinterlocutorreplied,whosthataquotefrom.
[t]hatsme,thecounterspeakertweeted.Sendingloveandhopetoyou.Fromthere,the
counterspeakerandherinterlocutorexchangedmorethanadozentweets,inwhichshe
describedlynchingsandotherviolenceagainstAfricanAmericans,withphotographsasking,
youmockthis?Theoriginalspeakerrepliedthathewas14yearsold.Thecounterspeaker
askeddoesyourmomknowyouspendyourtimeontheInternettryingtohurtpeople
pointingouttheharmfulconsequencesthathisTweetscouldinflictonmembersofthegroup
hewasdenigrating.Intryingtomakehimfeelaccountabletohismother,thecounterspeaker
seemstohaveunexpectedlymadeanemotionalconnectionwiththestrangerwhohadbeen
continuallyattackingher.Hereplied,Idoubtit.Shesbeendeadforayearandahalfso
Imsorryforyourloss,thecounterspeakerwrote.AndIhopeyoufindabetterwayto
honorher.Afterseveralmoreexchanges,hewrotetothewomanhehadbeenattacking
viciously,youresoniceandImsosorry.Thiscasewaspublicizedonlineandthe
counterspeakerwaspraisedforherforbearance(Payne,2015).
AnotherstrikingexampleofonetoonecounterspeechisthecaseofMeganPhelpsRoper,
whowasfullyconvincedoftheextremehomophobictenetsoftheWestboroBaptistChurch,
whichhergrandfatherFredPhelpsfoundedandinwhichshewasraiseduntilshestarteda
Twitteraccounttospreadtheviewsofthechurch.OnTwittersheencounteredpeoplewho
challengedherviewsandengagedherinotherways,includinghumorandsuggestionsfor
musicshemightenjoy.Extendedconversationswithtwoofthem(DavidAbitbol,arabbi
whofoundedtheblogJewlicious,andalawyerwithwhomPhelpsRopergraduallyfellin
love)completelychangedPhelpsRopersviews,byherownaccount.Sheendedupleaving
thechurch.ThiscaseisrecountedbyAdrianChen(2015)inalongarticlethatPhelpsRoper
haspraisedforcapturinghertransformationverywell.
Akeydistinctionbetweentheseonetoonecasesandtheotherexchangeswestudiedisthat
thesedialoguesarelong,goingfarbeyondthethreeessentialpartsofwhatweaffectionately
callagoldenconversation.Itisnosurprisethatdeepand/orlastingchangeindiscourseand
beliefsdifficulttoachievebyanymeans,onlineorofflinecantakemanytweets.Another
distinguishingfeatureofonetooneconversationsisthat,evenonTwitter,theyarenot
alwayspublic.OnecansendadirectmessageonTwitterwhich,likeanSMSortext
message,isvisibleonlytoitsdesignatedrecipient.InMeganPhelpsRoperscase,sheand
hernewinterlocutorsalsousedonetoonemessagingappsothersthanTwitter,suchas
WordsWithFriends.Inalesspubliconlinecontext,peoplemayfeellessguardedand
thereforemoreopentodissentingviews.Ontheotherhand,iftheirconversationsare
invisibletothelargeraudience,theaudiencecanneitherjoininnorbefavorablyinfluenced
bytheconversation,exceptinrarecaseswhenitisdescribedelsewhere,asinChensarticle
(2015).

16

OnetoMany
SomeTwitterusershavetakenituponthemselvestotrytochangewayinwhichothers
expressthemselvespubliclyonTwitter,bysearchingfortheuseofcertaintermsorphrases
andrebukingthosewhousethem.Thissortofactivisteffortisonetomanycounterspeech.
Inoneexample,DawudWalid,anAfricanAmericanMuslim,searchedforvariationsofthe
wordabeedwhichmeansslaveinArabic,whereitwasusedintweetstorefertoblack
people.Hesentanopedhehadwritten,entitledFellowhumansarenotabeedtomany
Twitteruserswhohadtweetedtheterm.Hereceivedavarietyofresponses,fromapologies
andpromisesnottousethewordagain,toatweetthatrepeatedthewordasmanytimesas
possiblein140characters(Walid,2013).Othersimilareffortsare@YesYoureRacistand
@YesYoureSexist.Ineachofthesecases,counterspeechismetwitharangeofresponses,
fromapologiestoangryargument.Inanotherexampleofonetomanycounterspeech,some
usersdeliberatelytweetonahashtagwithwhichtheydisagree,suchas#stopislam,toreach
peoplewhoagreewithit.
ManytoOne
Insomecases,newsofanobjectionabletweet(orhashtag)goesviral,andmanysometimes
thousandsofTwitterusersjoinincounterspeech.Thiscanbesalutarywhereitcatches
enoughoftheattentionoftheoriginalspeakertobesuccessfulbutnotharrassing,asinthe
caseofauserwhotweetedhisoutragethatNinaDavuluri(whoheerroneouslyidentifiedas
anArab)hadbeenchosenasMissAmerica2014.Afterreceivingtweetsthatvariously
correctedhiserrorandcalledhimaracist,hefirstrespondedIdidntrealizeitwould
explodelikethat#unrealandthentweetedatDavuluri,apologizing.Thefurordieddown
quickly,andtheuserisstillonTwitter,wishinghappybirthdaystofamilymembersand
tweetingphotosofhisfishingtrips.
Inothercases,however,hugenumbersofangryTwitterusershaveoverwhelmedothers,
risingtothelevelofharassment.Originalspeakershastilydeletetweetsoreventheir
accounts,buteventhatcanbeaninsufficientrefugeinthefaceof,forexample,
counterspeakerswhocontacttheiremployers,demandingthattheybefiredforhavingposted
perceivedhatefulorracistcontent.ThiswasthecaseforJustineSacco,whoinadvertently
createdanumberoneworldwidetrendonTwitterinDecember2013bytweeting
(tongueincheek,sheinsistedlater),GoingtoAfrica.HopeIdontgetAIDS.JustKidding.
Imwhite!assheboardedaflighttoSouthAfrica.Tensofthousandsoffurioustweetscame
withinhours,andSaccowassoonfiredfromherjobinpublicrelations(Ronson,2015).
TheblogRacistsGettingFiredmadeapracticeofpunishingpeoplewhomaderacistposts,
bycontactingtheiremployersand,similarly,demandingthattheybefired(McDonald,2014).
Suchresponsesarenodoubtsuccessfulatchangingtheonlinespeechoftheirtargets,but
mayonlyhardenthehatefulconvictionsofthosetargets,andconstituteonlinemobjustice.

17

ManytoMany
Conversationsamonglargenumbersofpeopleonlineareofinterest,notleastbecauseofthe
impressivescaleonwhichtheyoftentakeplace.Inourexperience,theyareoftencatalyzed
byofflineeventsthatareofstronginteresttoalargenumberandlargevarietyofpeople.On
Twitter,suchconversationsgenerallyformaroundhashtags.
Hashtagscanconstitutehatefulanddangerousspeech,orcounterspeech,andtheyoften
gatherorinspiremanytomanyconversations.Theuseofahashtagcanbeseenasan
explicitattempttoaddressanimaginedcommunityofusersaseachuserparticipatingina
hashtagconversationactspotentiallyasabridgebetweenthehashtagcommunityand
membersoftheirownfollowernetwork(Bruns&Burgess,2012,p.804).Often,one
hashtagrepresentsonegeneralviewornormativegroup,suchas#BlackLivesMatter,and
othersrepresentopposingordissentingviews,suchas#BlueLivesMatter(whichreferstothe
policefortheirblueuniforms),or#AllLivesMatter.
Oneofthemostvitriolichashtagswefound,#KillAllMuslims,trendedintheimmediate
aftermathoftheCharlieHebdomassacreofJanuary2015andthenwasquicklytakenover
bycounterspeakersexpressingtheirdismaythatitexisted.Onecounterspeechtweetthatuses
thehashtagwasretweetedmore10,000times:Notmuslimbutneverthoughtaboutthisb4
#CharlieHebdo#KillAllMuslims#Muslimspic.twitter.com/LL1pkPk6uk. Thelinkwastoan
imageofvisualsimilaritiesamongreligioustraditions,e.g.aCatholicnuninahabitanda
Muslimwomaninhijab.

Notably,hashtagscanbemorewidelyandquicklydisseminated,whentheytrend,thanany
tweet.When#KillAllMuslimstrended,forexample,thousandsofpeopleonTwittercould
nothelpbutnoticetwothings:thehashtagcalledformassmurderorgenocide,andthousands
ofpeoplehadtypeditandsentit,aspartoftheirtweets.Thefactthatahashtagistrending
canalsohaveamajorimpactonhowTwitterusersperceivenormsontheplatform.Itis
dismayingwhenhatefulhashtagstrend,andreassuringwhencounterspeechdoes.The
hashtag#YouAintNoMuslimBruv,forexample,trendedafterabystanderyelledthesame
phraseinDecember2015,atawouldbeattackerinLondon.Aworthytopicforfurtherstudy
wouldbethenorminfluencingcapacityofhashtagsaroundpubliceventsandcontroversies.
Taxonomyofcounterspeech
Wehaveobservednumerouscommunicativestrategiesatworkincounterspeechtweets,
whichwepresentanddiscusshere.Somearewelldescribedinpersuasionand
communicationliterature.Wedistinguisheightstrategies:1)presentationoffactstocorrect
misstatementsormisperceptions,2)pointingouthypocrisyorcontradictions,3)warningof
possibleofflineandonlineconsequencesofspeech,4)identificationwithoriginalspeakeror
targetgroup,5)denouncingspeechashatefulordangerous,6)useofvisualmedia,7)useof
humor,and8)useofaparticulartone,e.g.anempatheticone.Thesearenotexclusive

18

counterspeakersoftenemploymultiplestrategiesinasingletweet.Norisourlist
comprehensiveitisintendedasastart,basedondatacollectedforthepresentproject.
Presentingfactstocorrectmisstatementsormisperceptions
Likepreviousauthors(Bartlett&KrasodomskiJones,2015AntiDefamationLeague,2014
Briggs&Feve,2013),weobservedthatcounterspeakersoftentrytopersuadebycorrecting
misstatements.Unfortunatelythereisabundantresearchindicatingthatthisrarelysucceeds,
especiallyinthefirstinstance.6Evenwhenpresentedwithfactsthatcorrectanegregious
mistakeorfalsehood,mostpeopleareunlikelytochangetheirbeliefsaboutthe
misperceptionasfalsebeliefsbasedonmisinformation,areoftenheldstronglyandwith
(perhapsinfectious)conviction(Lewandowskyetal.,2012,p.108).Paradoxically,
individualswithminimalknowledgeoffactsandexperienceofthesubjectareleastlikelyto
beswayedbycorrections(Kuklinskietal.,2000).Someresearchershaveevenfounda
backfireeffectwherecorrectionattemptsresultinfirmerbeliefsinthemisperceptionor
misinformation(Nyhan&Reifler,2010Nyhan&Reifler,2015).Whereasthecorrections
maynotbesuccessfulinchangingthespeechorattitudesoftheoriginalspeaker,whomay
haveatightlyheldbeliefonthetopic,theycouldpersuadecyberbystandersi.e.membersof
thegeneralaudiencewhohavenotyetalliedthemselvesfirmlywitheitherside.
Perhapsawareofhowdifficultitistochangemindsevenwithfacts,counterspeakers
sometimessupporttheircontentionswithevidenceoreveninviteotherstolookupthe
evidenceforthemselves.Forexample,oneuserwhowasagainstTargetsnewpolicyof
genderneutralbathrooms,saidover400,000peoplehadsignedapetitiontoboycottTarget.
Anotherusertriedtocorrectthenumber(<5,000),andinvitedthefirstusertocheck.The
originalspeakerinsultinglydeclined(seeFigure4).

Figure4.Counterspeakerusingfactstocorrectmisinformationandmisperception
Counterspeakerssometimesgotostrikinglengthstopersuadestrangersthattheirfactsor
understandingarewrong.Forexample,inresponsetothesuggestionthattheimageofthe
abolitionistHarrietTubmanbeplacedonanEBT(welfare)cardinsteadoftheUS$20bill,a
6

Psychologistsandotherscholarsusethetermmotivatedreasoningtodescribepeoplesextraordinary(and
routine)effortstoreachtheconclusionstheyseek,inspiteoffactsthatdonotsupportthoseconclusions.See,
e.g.DanKahan,NeutralPrinciples,MotivatedCongnition,andsomeProblemsforConstitutionalLaw.Harvard
LawReview,2011.

19

counterspeakerengagedtheoriginalspeakerinadebateabouttheproportionsofwhiteand
blackAmericansonwelfare(seeFigure5).Thecounterspeakerfounddatafromaseparate,
morereputablesource(theU.S.DepartmentofAgriculture),andthenpostedagraphicto
illustratethedata.Theeffort,likemostoftheothersweobserved,wasunsuccessful.

Figure5.Counterspeakermarshallingstatistics
InadiscussionofIslamicprayerroomsinAmericanpublicschools,onespeakertriedtouse
historicaleventstocorrecttheoriginalspeakersassertions,PointofhistoricalfactSpanish
soldiersmurderedPuebloIndianreligiousleadersbecausetheywerecompetition.Oneof

20

thosetaggedrespondedtothiscounterspeech,andapoliteexchangecontinuedbutthefact
didnotapparentlychangeanymindsonthetopic.
Pointingouthypocrisyorcontradictions
Inastrategysimilartocorrectingmistakesormisstatements,counterspeakersoftenpointed
outhypocrisyorcontradictions.This,too,maycauseanoriginalspeakertodiginhisorher
heels,sinceitisdisturbingtobecalledahypocrite.Inordertodiscredittheaccusation,the
individualmayexplainawayandrationalizetheirpreviousbehavior,oriftheyare
persuadable,resolvetoavoidthedissonantbehaviorinthefuture(Beauvois,Joule,&
Brunetti,1993).Buttomembersoftheaudience,whomwecancallcyberbystanders,
pointingouthypocrisymaybepersuasive.
Inoneexample,ausertweeted#KillAllMuslims!!!!Washitlerabadperson?Yeshow
darehekillsomeonejustcosoftheirreligionUmmm...Ok.Anotherexampleofcallingout
hypocrisyorcontradictionsistopainttheoriginalspeakerwiththesamebrushheorsheis
usingtopaintagroup.Forexample,onecounterspeakertweetedIfamuslimtwtd
#KillAllChristianshemustba"terrorist"btachristiantwtng#KillAllMuslimsissimply
exprsnghisfreedomwhileanothersimplystated,#KillAllMuslims"..butyettheycallus
terrorists.
Warningofofflineoronlineconsequences
Warningusersofthepossibleconsequencesoftheirhatefulordangerousspeechisacommon
strategyamongcounterspeakers.Likethepreviousstrategy,thismayleadanoriginalspeaker
toholdmoretightlytoopinions.However,inourdata,thiscounterspeechstrategyappeared
tobesuccessful,leadingtorecantingordeletionofoffendingtweets.Thismaybebecause
peopleforget,untilreminded,thattheirspeechonlinecanhaveseriousconsequencesoffline.
Online,theycanbecomedisinhibited,actingasiftheyliveinanmakebelievedimension,
separateandapartfromthedemandsandresponsibilitiesoftherealworld(Suler,2004,p.
323).Itmayalsobethatwarningsorthreatsareeffectivebecausetheyhavebeencarriedout
inwellpublicizedcases,suchassuccessfuldemandsthatpeoplebefiredfromtheirjobsfor
contentthattheypostedonline.
Inonesuchcase,aTexasschoolteacher7postedaracistdiatribeonFacebookinthewakeofa
scandaloverpolicebrutalityinMcKinney,Texas(Michael&Young,2015).Screenshotsof
herdiatribemadetheirwayontoTwitter,andwerecirculatedwidely.Inresponse,many
userstaggedheremployer,FrenshipIndependentSchoolDistrict,onTwittercallingforher
ouster,"@FrenshipISDShehasgottoGO!!GetridofherNOW.#[teachersname],
"@FrenshipISDIf#[teachersname]hasanyblackkidsinher4thgradeclassPLEASEget
themoutofthere!#Mckinney,"@FrenshipISDistheracist#[teachersname]firedyetoris

Weavoidrepeatingthenameoftheteachersinceshehasnotchosentodiscussthecasepublicly,nor
issheapublicfigure.

21

thistypeofideologyencouragedinyourdistrict?Theteacherwaseventuallyfiredfor
failingtoabidebythestaffcodeofconductonsocialmedia.
Inmanyothercases,counterspeechfeatureswarningsoftheriskoffutureunemployment.
RespondingtomisogynistandracisttweetsregardingUniversityofIllinoisat
UrbanaChampaignChancellorPhyllisWisesdecisionnottocancelclassesonaverycold
dayinJanuary2014,counterspeakerswerequicktowarnofrealworldconsequences
(Simeone,2014).Arecurringthemewastheimplicationsonfutureemployment.Some
tweeted,"Allofyoustudentswhotweetedracistcomments,GOODLUCKGETTINGAJOB
DUMMIES#fuckphyllisand"Ppldefending#FuckPhyllisfocusingonwrongthing.You
shouldbeworryingabhowtoundoyourtweetscomeyoureventualemploymentsearch.
Anotherremindedstudentsthattheirprofessorsandfutureemployersmightseetheirtweets
(seeFigure6).

Figure6.Imageoftweetwarningusersofofflineconsequences
Oneuserdidreportattemptingtopunishstudentsbycontactingtheiremployers,tweeting"11
studentsreportedtoWise.1reportedtoemployer.seeifthey'llbelaughingabt#fuckphyllis
Monday#diversityMadnessandlater"lmfao...takingscreencapsandnamesoftheseracist
kids.Foundtheemployerforone....stillworking.Anotherusertweeted,"Asanemployer
whohasrecruitedmanyfreshoutofcollegestudents,Idohavesecondthoughttohire
anyoneoutofUIUCseeing#fuckphyllis."Somealumnievencomplainedthatthehateful
tweetingwouldputaburdenonthem:"Policeyourselves,UofIstudentsburdenof
#fuckphyllisalsofallsontheshouldersofusalumniwhotrytogetuhiredwhenugraduate.
Onlineconsequencesofhatefulspeechareblockingandostracism.Manycounterspeakers
documentedtheireffortstolimitinstancesofhatefulspeechintheirpersonalfeedsas
warningstothegeneralaudience:"Thesituationin#Baltimorebroughtoutthe#racists.My
friend'slistisdownadozensofarblockedandunfriended,"Callingthe#protesters
#animalsanddefending#Baltimoreforkilling#FreddieGrayyou'regonnagetblocked!
#AmericaIsBetterThanThat,Anyoneusingthehashtag#KillAllMuslimsonmyTLget
blocked,WhoEverCameupwiththis#KillAllMuslimshe/Sheshouldbeblockedfromall
socialnetwork.
Counterspeakerssometimesalsotrytoprovokeempathybypointingouttotheoriginal
speakerthathatefulordangerousspeechmayhavepainfulconsequencesforothersonline.
Forexample,thewriterwhorepeatedlyrepliedtoanotherTwitterusersracisttweetson
MartinLutherKingDay2015,asdescribedabove,tweetedathim,rememberthatthereisa
goodchancethatsomeonecomingacrossyourhatefultweetsissufferingfromdepression
too.Inothercases,counterspeakerswarnofharmtothegroupsdenigratedinhatefultweets,

22

IfUofIllinoisstudentscansay#FuckPhyllisandberacist/sexisttotheirChancellorthenI
fearforotherAsiansand/orwomenoncampus.However,warningoriginalspeakersofthe
offlineconsequencesthattheythemselvesmightfaceappearstobethemosteffectivetypeof
thisstrategy.
Unfortunately,warningscaneasilyswellintothreatsandharassment,sometimesbylarge
numbersofpeople.Counterspeakerscanalsotargetthewrongindividuals,suchaspeople
whohavewrittensomethingoffensiveorinsensitive,butnothatefulordangerousoreven
peoplewhosimplyhappentosharethenameoftheintendedrecipient.Nomatterwhat
contenttheyarecountering,wouldbecounterspeakersshouldbeawareofthelinebetween
counterspeech,ontheonehand,andtrolling,harassment,dogpiling,oreventhreatsof
physicalharmontheother.Itisbothwrongandineffectivetoattackinthoseways,andit
mustnotbeconfusedwithcounterspeech.
Affiliation
Affiliationis...establishing,maintaining,orrestoringapositiveaffectiverelationshipwith
anotherpersonorgroupofpersons(Atkinson,Heyns,&Veroff,1954,p.407).Associal
creatures,weunderstandweareamemberofnumeroussocialgroupsandthemembership
contributes,positivelyornegatively,totheimageonehasofhimorherself(Tajfel,1974,p.
69).Peoplemayalsomorelikelytocreditthecounterspeechofthosewithwhomthey
affiliate,sincetheytendtoevaluateingroupmembersasmoretrustworthy,honest,loyal,
cooperative,andvaluabletothegroupthanoutgroupmembers.(Kane,Argote,&Levine,
2005,p.58).
Ifacounterspeakerpersonallyaffiliateswithahatefulspeaker,itcanreducetheperceived
distancebetweenthetwo(Tanis&Postmes,2003).Counterspeakers,usingimpliedsocial
identitieseitherinthetextofhatefulordangeroustweetsorfromthehatefulspeakersprofile
information,canusetheinformationgleanedtoaffiliatewiththeoriginalspeakersassumed
orperceivedsocialidentities.Insomecases,counterspeakersuseasharedidentitytoclaim
thatcertainspeechisunacceptableforamemberofthatgroup.Inthetweetbelow,the
speakeruseshisidentityasarightwingertobothassociatewith,anddistancehimself
from,theoriginalspeakers(Figure7).Asnotedabove,BergerandStrathearnsresearch
indeedfoundthatconservativesarebettersuitedtoreachingouttowhitenationalistsonline
thanliberals,becauseofaperceivedaffiliation(2013).

Figure7.Illustrationofcounterspeakeraffiliatingwithoriginalspeaker.
Inthe#FuckPhyllisdata,alumniassertedtheiraffliationwiththestudentswhostartedthe
trendtolendcredibilitytotheircounterspeech,andtodeclarethestudentsbehavior

23

outofboundsfortheirsharedidentity.Forexample:AsaUofIalumna,Iamcompletely
embarrassedbythis#FuckPhyllishashtag.Realitycheckkids:professors&employerscan
seeeverything.Thisaffiliationtacticwasalsosuccessfulinrecruitingothercounterspeakers.
Itcanfacilitatehelpingorinterveningbehaviors,mostlikelyduetogroupnorms,suchas
expectationstohelponeanotherorbehavelikeoneanother(Rutkowski,Gruder,&Romer,
1983).
Inanotheruseofaffiliation,usersfromagroupvilifiedintweetssometimesdisclosetheir
affiliationineffortstocounterthosetweets.Forexample,inresponseto#KillAllMuslims,
onetweetedAsaMuslimI'mdisgustedandoffendedthatthisisevenahashtag.
#KillAllMuslims.IdentifyingasMusliminthefaceofonlineIslamophobiawasfoundin
otherevents.Oneuser,inresponsetotheGarland,TXshootingtweeted,AsaMuslimliving
inTexasthetweetsunderthe#garlandshootinghashtagarebothalarminganddisturbing.
Prayingforpeace&tolerance.Othershumanizedadangerousproposalinthehashtag
#KillAllMuslims,#KillAllMuslimsi'mamuslimandidon'tthinkishouldbekilledbc
someonedidsomething,Can'tbelievethatthisistrending,I'maMuslimmyselfsoyou
wantmedead?#KillAllMuslims.Theeffectsofracistspeechonatargetgroupwerespelled
outbyselfdisclosingtargetgroupmembers,AsanAsianAmericanwoman,trendslike
#FuckPhyllismakeitclearthatnoamountofsuccesscaneraseracistandsexistattacks.
Sometakeituponthemselvestoapologizeonbehalfofidentitiestheypresumedtheyhadin
commonwithhatefulspeakers.Forexample,oneusertweeted,Asachristian,I'msorryfor
thehashtag#KillAllMuslims.It'sashamethathumanityisgonefromthemodernworld.
Others,havingidentitiesincommonwithbothatargetgroupandtheoriginalspeakersofthe
hatefulspeech,speakoutagainstthehatefulspeechgenerally(seeFigure8).

Figure8.Imageoftweetsaffiliatingwithtargetandhatefulspeakinggroup.
Denouncinghatefulordangerousspeech
Counterspeakersoftenidentifyspeechashateful,especiallyracistormisogynist.Asoneuser
tweeted,Iknowyoucantalwayschangearacistsmindandtbhidontwanttobother
wastingmytimeonthembutillbefuckedifidontatleasthumiliateandshamethem.

24

Denouncingthecontentandnotthespeakercanbeuseful,andeveneducationalifthe
speakerisunawaretheyareengaginginhatefulordangerousspeech(seeFigure9).

Figure9.Counterspeakerdistinguishingspeechfromspeaker
Somecounterspeakerssimplycallouttweetsasracist(seeFigure10).Otherselaborateon
whyspeechishatefulordangerous:YES.ITDOES.Theentiretagismisogynisticand
encouragesrape.#fuckphyllis.

Figure10.Counterspeakersdenunciation,quotinganoriginalspeakerstweet.
Wheredenunciationisdirectedataperson,notatcontent,itcanbeindistinguishablefrom
hostilenamecalling,asintheexchangeinFigure11.OneTwitterusermakesacritical
commentaboutatheists(andatheism)andanotherrespondswithaninsult:pissoff,bigot.
Aseconduserinsteadexplainswhythefirsttweetmightbeunderstoodasbigoted.Afterthat,
theoriginalspeakersayshertweetwassnottyandclarifiesherpoint.

Figure11.Twocounterspeakersdifferentstrategies

25

Counterspeakersoftendenouncehashtags.Forexample,oneusertweetedaskingrhetorically,
whyis#FeminismIsAwfulathing?Itsupportsequality.Whyarepeopleagainstthat?Many
tooktoTwittertodenouncethehashtag#KillAllMuslims,untiltweetscondemningit
outnumberedthosesupportingit.Examplesare#KillAllMuslimsisaviciousandcrueltrend
anditjustneedstostopand#KillAllMuslimsisliterallythemostdisgracefulthingI'veseen
onTwitterandoneofthemostoftenretweetedpostsinthewakeoftheCharlieHebdo
shootingsinJanuary2015wasTothosewhotrended#KillAllMuslimsYou'renodifferent
thanAlQaeda,Osama&ISISYouarehumanityatitsworst!
Visualcommunication
Twitterallowsuserstoembedimages(i.e.memes,graphics,photographs,animatedgifs,
videos)intweets,andcounterspeakersoftenusethem,sinceimagesaremorepersuasivethan
textalone.Theysendpeoplealongemotivepathwayswheretextual/verbalmaterialleaves
theminamorerational,logicalandlinearpathwayofthought(Joffe,2008,p.84).Online
specifically,visualscanaugmenttextualinformationviaparatasis,thatis,bybeingplaced
nexttosuchinformationasacoordinate,supportivestructure(LaGrandeur,2003,p.124).In
oneexample,aheavilyretweetedcounterspeechpostcontainedthetextthefactthat
#KillAllMuslimsis/wastrendingmakesmesicksoheresafriendlyreminderwithanimage
illustratingtheproportionofterroristsamongallMuslims(seeFigure12).

Figure12.Visualcommunicationexample.
Otherstrytoputcurrenthatefulspeechintohistoricalcontext(seeFigure13).

26

Figure13.VisualcomparisonofIslamophobia,Nazis,andKKK.
Visualcaricaturescanmaketheoriginalseemludicrous,andmaylimittheeffectsofhateful
ordangerousspeech.Forexample,usershavetriedtoundermineperceptionsofthestrength
andpowerofISISbyalteringISISphotos,whichaboundonTwitter,withrubberduckheads
andaccoutrements(seeFigure14).

Figure14.SpoofingISIS.
Usersalsoembedanimatedvisualslikegifsandvideosincounterspeech.Forexample,many
tweetedESPNs#MoreThanMeanvideoofmenreadinghatefulanddangeroustweetssentto
femalesportsreporterswithacommentWeshouldberespectfulofothersinpersonor
online(seeFigure15).

27

Figure15.Videocounterspeech.
Humor
Humorseemstohavespecialpowerasacounterspeechstrategy.Itmayshiftthedynamicsof
communication,deescalateconflict,anddrawmuchmoreattentiontoamessagethanit
wouldotherwisegarner.Someperceivehumorasalinguisticorcommunicativeperformance,
whichputstheactofspeakingondisplayobjectifiesit,liftsittoadegreefromits
interactionalsettingandopensittoscrutinybyanaudience(Baumann&Briggs,1990,p.
73).Researchershavefoundthathumorinonlinesettingseaseshostility,offerssupportto
otheronlinespeakers,andencouragessocialcohesion.(Marone,2015,p.61).OnTwitter,
humoroftengetstheattentionofaverylargeanddiverseaudience,asillustratedbelow,and
isarelativelycommonstrategy.Itcomesinmanyforms,ofcourse,includingcaricatureand
sarcasm,andcanvaryimmenselyintone,fromconciliatorytoprovocativeoreven
aggressive.
Humorisoftensuccessfullyconveyedinimages,soitallowspeoplewhodonotsharea
languagetocounterspeaktogetherofteninlargenumbersandacrossculturalandnational
boundaries.OneexampleofthiscameinJuly2014,afterTurkishDeputyPrimeMinister
BlentArincsaid(givinghisideasforproperfemaleconduct)thatwomenshouldnotlaugh
loudlyinpublic.Withhoursofthespeech,#direnkahkaha(Turkishforresistandlaugh)
wasoneofthetoptentrendinghashtagsworldwide(Bacchi,2014).Inspired,outraged,and
amused,tensofthousandsofwomenandmenpostedlaughingselfies.Inmanycases,the
imageswerealsofunny,astheyshoweddogs,cats,goats,andhorsesapparentlylaughing.
AnotheramusingimagespreadlikewildfireonTwitterinApril2014,afterDaniAlves,a
Braziliansoccerplayer,wassubjectedtoanalltoofamiliarhumiliatingracistgesturewhena
spectatorthrewabananaathimonthefield.Alvespickedupthebanana,peeleditandateit.
Anotherplayer,NeymardaSilvaSantosJr.,quicklypostedanimageofhimselfeatinga
banana,onthehashtag#Somostodosmacacos,meaningweareallmonkeys.Thathashtag,
inEnglish,spreadquicklyaswell.Inthiscase,Neymarhadpreparedtheimageinadvance
withtheadviceofanadvertisingfirm,sincebothheandAlveshadbeenpeltedwithbananas
manytimesbeforeandthousandsofpeoplepostedbananaeatingselfiesinspontaneous
support.SeeFigure16.

28

Figure16.Tweetusingahashtagintwolanguages,tocounterracism.
Insomecases,humorisusedtosoftenamessagethatwouldotherwisebeblunt,andperhaps
tomakeitmorepersuasive.Apopularmeme,theItsTimeToStopPostingcatwasused
byoneusertoprotestSinnFinpresidentGerryAdamsuseofthenword.(seeFigure17).
Inresponsetothatcritiqueandothers,Adamsapologizedanddeletedhistweetquickly,but
saidhehadbeeninnocentlycomparingthesufferingofAfricanAmericanswiththatofIrish
nationalists.Whilethereareparallelsbetweenpeopleinstruggle,thetweetwas
inappropriate"hetweeted.

Figure17.Catmemeusedtosoftenademand
Usingcaricatures,bycontrast,isawaytocounterhatefulordangerousspeechbymockingit.
Forexample,severalcounterspeakerstookonauserwhotweetedracistisaderogatory
wordusedonlytodescribewhitepeoplewhoopposetheirgenocide.Antiracistiscodeword
forantiwhite.Diversityisacodewordforwhitegenocide.Oneuserrepliedwitha
caricatureoftheNFLteamWashingtonRedskinslogo,usingwhitestereotypes(e.g.striped
ties,golfclubs,blondhair,lanternjaw)tomocktheoriginaluser(seeFigure18).

29

Figure18.Imageofcaricatureofstereotypesbycounterspeaker
Humorouscounterspeechhasbeenusedeveninresponsetoviolentextremistcontenton
Twitter.Forexample,theISISrubberduckpicturesinFigure14aremeanttomake
intimidatingimagesandmessageslessfrighteningorpowerful.Similarly,whenthehuman
rightsactivistIyadElBaghdadiretweetedpartofacalltoactionfromISISin2015,Muslim
Twitterusersrespondedwithhumorousreasonswhytheycouldnotjoin(seeFigure19),
referringtopopularcultureandtotheQuran.LiketheISISrubberduckimages,thehumorin
thesecounterspeechtweetscanweakentheISIScontentbytrivializingit.

Figure19.HumorousresponsetoISISrecruitmentattempts
OftenTwitterhumorisalsosardonicorsarcastic,likethistweet:Heylet'strend
#KillAllMuslimsbutstillcallthemterroristbecauseitmakessomuchsense!!!!Another
counterspeakercommentedonDonaldTrumpsproposedbanonadmittingMuslimstothe
UnitedStates,@realDonaldTrumpcongratulations!You'rethenewHitler!#MuslimBan.
Users,aswellasresearchers(Maynard&Greenwood,2014),havedifficultyidentifying
sarcasmreliablyandsometimesgivethebenefitofthedoubt,Can'ttellifyou'rebeing
sarcasticornot.Doyoureallynotknowthatblackfaceisoffensiveandaracistact?On
theirownuseofsarcasm,somecounterspeakersleavenoroomfordoubt,taggingtheirtweets
withahashtagorsomeformofmodifiedcodelike/sarcasm(Liebrecht,Kunneman,&van
denBosch,2013).I'veneverhadasnowdayinmylife,evenwithtempsat46with
windchill,soI'mreallyfeelingsorryforyouall#sarcasm#fuckphyllis,orbysimply

30

announcingityesthatsrightyourhashtagsarentracistatall.Neitherarethephotosyou
haveposted.Justtomakesure,imbeingsarcastic.
Tone
Twittercounterspeechoccursalongaverywidetonalandemotionalspectrum,fromtweets
thatarejustasobscene,angry,andviciousasthetweetstowhichtheyrespond,tokindand
empathicmessages.Tweetsofdifferenttonehave,ofcourse,differenteffects.Hostiletweets
canpersuadeanoriginalspeakertodeletetweetsorevenawholeaccount,butareunlikelyto
eitherdeescalatetheconversationorpersuadetheoriginalspeakertorecantorapologize.
Forexample,onecounterspeakertweetedfucknotallmuslimsbelongtoISISinresponseto
anotherwhousedthehashtag#KillAllMuslims.Elsewhereonthespectrumaremany
varietiesoftone:empathic,kind,andpolite,orcivil,speech.
Ingeneral,expressionsofempathyhavebeenshowntoleadtopositiveattitudestowardsan
adversarywhentrustbetweenspeakersishigh,butnegativeeffectsiftrustislow(Nadler&
Liviatan,2006).Ithasalsobeenseenasaspeechaccommodationtactic,leadingto
convergencewherespeakersbegintoadopteachothersspeakingstyles(Street&Giles,
1982).Empathycanhelpaspeechpartnersavefaceindisclosure,helpingclosethe
interpersonaldistancebetweentwospeakerstoengendertrustandcredibilityofarguments.
Counterspeakersusedempathictonesintheirtweetsagainsthatefulhashtagsaswellas
towardsindividualuserswhoweretweetinghatefulcontent.Thehashtag#KillAllMuslims
yieldedmuchofthis,includingsuggestionsfornewhashtagstotrend.Forexample,oneuser
tweeted#KillAllMuslims#KillAllChristiansWhatatrocioushashtags.#LetLoveLiveand
#KillAllMuslimsHowabout#SaveHumanity??:")Anotheruserwasquitepolitein
denouncingthehashtag:#KillAllMuslimsIstronglydisagreewiththistag.ItisdisgustingI
believein#Love&#peaceit'sarealshame#ParisShooting.Evenwhenfacedwithexplicit
deathwishes,somerespondedwithempathy(seeFigure20).

Figure20.Empathicresponsetohatefulspeech.
InanextendedoneononeexchangewithauserintentontweetingracismonMartinLuther
KingDay2015,thecounterspeakerrepeatedlycounteredwithempathy,stillwishingyou
love,tryasyoumight,Imnevergoingtowishyouanythingbutlove,andIwouldnever

31

laughatyourmurder.Iwouldnevertauntyourgrieving.Iwouldnevermockyourfightfor
equality.Herempathetictweetsapparentlysucceededforadetaileddescriptionofthiscase,
seepage14above.
TypesofResponses
Wehaveidentifiedfivecounterspeechresponsetypes,whichmayserveasanimportant
startingpointforunderstandingwhencounterspeechissuccessful,withregardtoboththe
authorofahatefultweetanditsaudience.Theseresponsesare:anapologyorrecanting,the
deletionofthetweetoraccount,thecreationofmorehatefulspeech,asustainedandcivil
conversation,andelicitingnewcounterspeechfromtheaudience.
Apologyorrecanting
Theresponsetocounterspeechwhichmostclearlysuggestssuccessiswhentheoriginal
speakerrecantsorapologizes.Wefoundsuchresponsestoberare,butnotunheardof.We
observedthemwhenoriginalspeakersclaimedtobeunawarethattheirtweetswerehateful,
andwhentheywererebukedrepeatedly,orbyseveraldifferentaccounts.Forexample,inthe
MissAmericaincident,auseratfirstdeniedthathewasracist,andapologizedonlyafter
severalcounterspeakerscontradictedhim.Thissuggeststhatcontinuedcounterspeechas
originalspeakersslowlyretreatcaneventuallyleadtoanapology.Withoutinterviewing
thosewhorecantorapologize,however,wecannotknowwhytheyhavedoneso,orwhether
theirspeechremainslesshateful,orwhethertheirviewshavechanged.Sometimes,original
speakersmaybackdownorevenapologizesimplytogetcounterspeakerstoleavethem
alone,andmaystillspeakhatefullyinthefuture.
Deletion
Anothercommonresponsetocounterspeechisthattheoriginalspeakerdeletesthehateful
tweet,ortheentireaccountfromwhichitwassent.Weobservedthisasaresponseto
manytoonecounterspeechandinresponsetowarningsofofflineconsequences.As
mentionedabove,manycounterspeakersrespondedtothe#FuckPhyllishashtagbypointing
outtostudentsthatfutureemployerswouldbeabletoreadthosetweets.Inresponse,many
studentsdeletedtheirhatefultweets.
Measuringthesuccessofthistypeofresponserequiresacarefulconsiderationofitslong
termandshorttermeffects.Whilepushingtheoriginalspeakertoremovehatefulcontentisa
shorttermsuccess,itdoesnotensurethatthespeakerwillcommunicatelesshatefullyinthe
future.Evenifanaccountisdeleted,thespeakermaycontinueinthesameveinonanother
platformoronadifferentaccountonthesameplatform.Theoutcomethatmostindicates
longtermsuccessisameaningfulchangeintheoriginalspeakersfuturespeech,butthismay
alsoconstitutesilencing,atacticthatisoftenusedtocurbthespeechofwomen,minority
groups,anddissenters.Thishighlightsonereasonforourstrictdefinitionsofbothhateful
speechandcounterspeech:webelieveitisessentialnottosilencecontroversialviews.

32

Morehatefulspeech
SupportersofthesloganDontfeedthetrollsshouldbeunsurprisedtofindthatone
responsetocounterspeechthatweobservedismorehatefulspeech,notonlyfromtheoriginal
speaker,butsometimesalsofromothersonTwitterwhojoinintorefutethecounterspeaker.
Weobservedthisresponsecomingfromboththosewhoenjoyelicitingareaction,thosewho
seemtosincerelyholdhatefulbeliefs,andsomewhoareboth,e.g.racisttrolls.Itisareaction
similartothatofthosewithstronglyheldbeliefswhorespondtofactcheckingbybecoming
evenmoreentrenchedintheirviews.Thissortofresponseobviouslyincreaseshatefulspeech
online,butthatmay,inturn,inspiremorecounterspeakerstostepin.
Civilconversation
Wehavealsoobservedrepeatedexchangesbetweenthosewhodisagreestronglybutremain
civil.Successisuncertain,sincetheoriginalspeakeroftenrefusestoacknowledgeany
wrongdoing,butdiffusingasituationsothathatefulspeechisnolongerbeingproducedisat
leastapartialsuccess.Thecontacthypothesissuggeststhatprolongedcivilexchanges
betweenindividualswhosharedifferentviewsandnormallywouldnotinteractcanhelp
reducedistancebetweenthetwofactionsoropinions(Allport,1954Dekker,Belabas,&
Scholten,2015).Itisdifficulttoproveinthewildthatthecontacthypothesisisworking,
butthetheorysuggestssustainedcivilconversationsareasuccessfulformofcounterspeech.
Morecounterspeech
Thefinalresponsethatwehaveobservedisthecreationofmorecounterspeechbyother
Twitterusers,orcyberbystanders.Itisdifficulttoknowwhethersubsequent
counterspeakersweredrawnintoaconversationbytheinitialcounterspeech,butwehave
indicationsthatthisisthecaseandcansignalsuccessintheformoffavorableimpactonthe
audience.Duringthe#FuckPhyllisincident,oneUniversityofIllinoisalumnacredited
anothercounterspeakerwithinspiringherowncounterspeechagainstthehashtag.Shesaid,I
knew[redacted]wasanalumandsheinspiredmetousesocialmediaforgood,butIguess
myclassmatesdidntgetthememo.#FuckPhyllis.Similarly,dozensofTwitterusers
followedtheexampleofcomedianMichaelIanBlack,afterhediscoveredMegan
PhelpsRopersTwitteraccountandbegancounterspeakingtoit.Cyberbystanders,like
offlinebystanders,typicallychoosenottoactwhentheyperceivethatmanyothersare
presentandwatching(Dillon&Bushman,2015Dillon,2015),buttheyaremuchmore
likelytodosooncesomeoneelseacts(orcounterspeaks)(Markey,2000).
Thisresponsecouldbesuccessfulinthesensethatitmayhavesignificantimpactonthe
audience.Wespeculatethatasurgeincounterspeechmightacclimateotheruserstoit,sothat
someofthembegintopracticeit.Ifagroupofnewcounterspeakersbecomesvocal,evenif
stillintheminority,thatmaybeinfluentialinshiftingdiscoursenorms(Rosenberg,2011
Gladwell,2000).

33

Thistypeofresponseisnotwithoutitsdangers,however.Likecounterspeechthatsilences,
counterspeechwhichcreatesmorecounterspeechcanleadtodogpiling.Avoidingthis,like
silencing,requiresgoodjudgmentandselfcontrolonthepartofcounterspeakers.

ConclusionsandIdeasforFutureResearch
Wehopetohaveprovidedsomeusefulideasoncounterspeech,whichisworthfurther
explorationalthoughitissurelynot,byitself,asolutiontotheacuteproblemofvitrioland
extremismonline.
Wesuggestthatfutureresearchcontinueourfieldstudyapproach,gatheringlargerdatasets
andworkinginotherlanguagesandnationalcontexts,inordertocomparedatasets.Workon
thesefuturedatasetsshouldincludeacodebookorqualitativememoingsystem.
Itmayalsobeusefultoconductlaboratoryexperimentsonthecommunicativestrategiesfor
socialmediathatwevewitnessedonline.Itwouldbeespeciallyimportanttotesttheeffects
ofvariousformsofcounterspeechonanaudience,whichwouldprobablybemorefeasible
inalaboratoryexperimentthaninanobservationalstudylikeours.
Anotherlineofresearchwouldfocusontheuseofhashtagsashatefulspeechor
counterspeech.Itwouldbeusefultoexaminemultiplecasesinwhichhatefulhashtagsare
reclaimed,forexample,especiallywhentheytrendaftereventsthatignitestrongemotionsin
largepopulationsofpeople,suchasbombingsandshootings.
Otherimportantquestionstobetackledinclude:Whichsortsofcounterspeechworkbestfor
whatsortsofsubjects(people)?Arecertaintypesofcounterspeech(usingparticular
strategies)moreeffectiveinresponsetoparticularformsofhatefulordangerousspeechor
extremism(e.g.racismvs.misogyny,whitesupremacyvs.Islamicextremism)?Does
counterspeechdifferinnatureandeffectivenesswhenconversationsareamongstrangers
withnoties,peoplewhoknoweachother,orpeoplewithsomeofflinesharedidentityor
connection(e.g.studentsandalumnioftheUniversityofIllinois)?
Finally,itcouldbeofgreatinteresttointerviewindividualswhochangedtheirspeechor
attitudesafteronlineexchanges,tolearnmoreabouthowandwhysuchchangetakesplace,
andtointerviewcounterspeakerstolearnabouttheirreasoningandthoughtprocessasthey
debatepeoplewhoare,mostoften,perfectstrangers.

34

WorksCited
Allport,G.W.(1954).Thenatureofprejudice.Cambridge,MA:PerseusBooks.
AmichaiHamburger,Y.,&McKenna,K.Y.(2006).Thecontacthypothesisreconsidered:
InteractingviatheInternet.JournalofComputer MediatedCommunication,11(3),
825843.
Anderson,J.Q.,&Rainie,L.(2010).Thefutureofsocialrelations.PewResearchCenters
Internet&AmericanLifeProject,2.
AntiDefamationLeague.(2014).Bestpracticesforrespondingtocyberhate.
Barrett,R.(2014).TheIslamicState.TheSoufanGroup.
Bartlett,J.,&KrasodomskiJones,A.(2015).Counterspeech:Examiningcontentthat
challengesextremismonline.Demos.
Bauman,R.,&Briggs,C.L.(1990).Poeticsandperformanceascriticalperspectiveson
languageandsociallife.AnnualReviewofAnthropology,19,5988.
Baym,N.K.(1995).Theperformanceofhumorincomputermediatedcommunication.
JournalofComputer MediatedCommunication,1(2),00.
Bacchi,U.(2014,July30).#DirenKahkaha:Turkishwomeninsocialmedialaughprotest
againstErdogansdeputy.TheInternationalBusinessTimes.Retrievedat
http://www.ibtimes.co.uk/direnkahkahaturkishwomensocialmedialaughprotestag
ainsterdogansdeputy1459005.
Berger,J.M.,&Morgan,J.(2015).TheISISTwitterCensus:Defininganddescribingthe
populationofISISsupportersonTwitter.TheBrookingsProjectonUSRelationswith
theIslamicWorld,3(20).
Berger,J.M.,&Strathearn,B.(2013).Whomattersonline:Measuringinfluence,evaluating
contentandcounteringviolentextremisminonlinesocialnetworks.TheInternational
CentrefortheStudyofRadicalisationandPoliticalViolence.
Beauvois,J.L.,Joule,R.V.,&Brunetti,F.(1993).Cognitiverationalizationandact
rationalizationinanescalationofcommitment.BasicandAppliedSocialPsychology,
14,117.
Benesch,S.(2013).Proposedguidelinesondangerousspeech.TheDangerousSpeech
Project.Retrivedfromhttp://dangerousspeech.org/guidelines
Benesch,S.(2014).Defininganddiminishinghatespeech.InFreedomfromhate:Stateof
theworldsminoritiesandindigenouspeoples2014(pp.1825).London:Minority
GroupInternational.

35

BipartisanPolicyCenter.(2012).CounteringonlineradicalizationinAmerica.Washington,
DC:Neumann.
boyd,d.,Golder,S.,&Lotan,G.(2010,January).Tweet,tweet,retweet:Conversational
aspectsofretweetingonTwitter.InSystemSciences(HICSS),201043rdHawaii
InternationalConferenceon(pp.110).IEEE.
Brewer,M.B.,&Brown,R.J.(1998).Intergrouprelations.InD.T.Gilbert,S.T.Fiske,&
G.Lindzey(Eds.),Handbookofsocialpsychology(4ed.,Vol.2,pp.554594).New
York:McGrawHill
Briggs,R.,&Feve,S.(2013).Reviewofprogramstocounternarrativesofviolent
extremism.InstituteofStrategicDialogue.
Brown,R.(2016).Defusinghate:Astrategiccommunicationguidetocounteractdangerous
speech.UnitedStatesHolocaustMemorialMuseum.
Bruns,A.,&Burgess,J.E.(2011,August).TheuseofTwitterhashtagsintheformationof
adhocpublics.InProceedingsofthe6thEuropeanConsortiumforPolitical
Research(ECPR)GeneralConference2011.
Chen,A.(2015,November23).Unfollow:HowaprizeddaughteroftheWestboroBaptist
Churchcametoquestionitsbeliefs.NewYorker.Retrievedfrom
http://www.newyorker.com/magazine/2015/11/23/conversionviatwitterwestboroba
ptistchurchmeganphelpsroper
Citron,D.K.(2014).Hatecrimesincyberspace.Cambridge,MA:HarvardUniversityPress.
Citron,D.K.,&Norton,H.L.(2011).Intermediariesandhatespeech:Fosteringdigital
citizenshipforourinformationage.BostonUniversityLawReview,91,1435.
Conover,M.,Ratkiewicz,J.,Francisco,M.R.,Gonalves,B.,Menczer,F.,&Flammini,A.
(2011).Politicalpolarizationontwitter.ICWSM,133,8996.
Dekker,R.,Belabas,W.,&Scholten,P.(2015).InterethnicContactOnline:Contextualising
theImplicationsofSocialMediaUsebySecondGenerationMigrantYouth.Journal
ofInterculturalStudies,36(4),450467.
Dinakar,K.,Reichart,R.,&Lieberman,H.(2011).ModelingthedetectionofTextual
Cyberbullying.TheSocialMobileWeb,11,02.
Dillon,K.P.(2015,November).Helpmehelpyou:Victimcommunicationin
cyberbystanderinterventionincyberbullying,PaperpresentedattheNational
CommunicationAssociation101stAnnualConference,HumanCommunication&
TechnologyDivision,LasVegas,NV.

36

Dillon,K.P.,&Bushman,B.J.(2015).Unresponsiveorunnoticed?:Cyberbystander
interventioninanexperimentalcyberbullyingcontext.ComputersinHuman
Behavior,45,144150.
Frenett,R.,&Dow,M.(2015).Onetooneonlineinterventions:ApilotCVEmethodology.
InstituteforStrategicDialogue.
Foxman,A.H.,&Wolf,C.(2013).Viralhate:ContainingitsspreadontheInternet.New
York,NY:Macmillan.
Fraser,B.(1990).Perspectivesonpoliteness.JournalofPragmatics,14(2),219236.
Gladwell,M.(2006).Thetippingpoint:Howlittlethingscanmakeabigdifference.Boston:
Little,Brown.
Glaser,B.,&Strauss,A.(1967).Thediscoveryofgroundedtheory.Chicago:Aldine.
Herz,M.,&Molnar,P.(2012).Thecontentandcontextofhatespeech:Rethinking
regulationandresponses.NewYork,NY:CambridgeUniversityPress.
Hussain,G.,&Saltman,E.M.(2014).Jihadtrending:Acomprehensiveanalysisofonline
extremismandhowtocounterit.TheQuilliamFoundation.
Joffe,H.(2008).Thepowerofvisualmaterial:Persuasion,emotionandidentification.
Diogenes,55(1),8493.
Kane,A.A.,Argote,L.,&Levine,J.M.(2005).Knowledgetransferbetweengroupsvia
personnelrotation:Effectsofsocialidentityandknowledgequality.Organizational
BehaviorandHumanDecisionProcesses,96(1),5671.
Kuklinski,J.H.,Quirk,P.J.,Jerit,J.,Schweider,D.,&Rich,R.F.(2000).Misinformation
andthecurrencyofdemocraticcitizenship.TheJournalofPolitics,62(3),790816.
Kwok,I.,&Wang,Y.(2013).Locatethehate:Detectingtweetsagainstblacks.In
Proceedingsofthe27thAssociationfortheAdvancementofArtificialIntelligence
Conference,16211622.
LaGrandeur,K.(2003).DigitalimagesandclassicalpersuasioninMaryE.Hocks&
MichelleR.Kendrick(Eds.)Eloquentimages:Word&imageintheageofnew
media.Cambridge:MITPress.
Lewandowsky,S.,Ecker,U.K.H.,Seifert,C.M.,Schwarz,N.,&Cook,J.(2012).
Misinformationanditscorrection:Continuedinfluenceandsuccessfuldebiasing.
PsychologicalScienceinthePublicInterest,13(3),106131.
Liebrecht,C.C.,Kunneman,F.A.,Bosch,A.P.J.vanden(2013).Theperfectsolutionfor
detectingsarcasmintweets#not.InBalahur,A.Goot,E.vanderMontoyo,A.

37

(Eds.),Proceedingsofthe4thWorkshoponComputationalApproachesto
Subjectivity,SentimentandSocialMediaAnalysis,2937.
Markey,P.M.(2000).Bystanderinterventionincomputermediatedcommunication.
ComputersinHumanBehavior,16(2),183188.
Marone,V.(2015).Onlinehumourasacommunitybuildingcushioningglue.TheEuropean
JournalofHumourResearch,3(1),6183.
McDonald,S.N.(2014,December2).RacistsGettingFiredexposesweaknessesofInternet
vigilantism,nomatterhowwellintentioned.TheWashingtonPost.Retrievedfrom
https://www.washingtonpost.com/news/morningmix/wp/2014/12/02/racistsgettingfi
redexposesweaknessesofinternetvigilantismnomatterhowwellintentioned/
Mendel,T.,Herz,M.,&Molnar,P.(2012).Doesinternationallawprovideforconsistent
rulesonhatespeech?Thecontentandcontextofhatespeech:Rethinkingregulation
andresponses,417429.
Michael,K.,&Young,A.D.(2015,June11).FrenshipISDteacherapologizesafter
McKinneyrelatedsegregationpost,LubbockAvalancheJournal.Retrievedfrom
http://lubbockonline.com/education/20150610/frenshipisdteacherapologizesafter
mckinneyrelatedsegregationpost#.V1iqzasaLZo
MuslimAdvocates.(2014).Clickheretoendhate:AntiMuslimbigotryonline&howtotake
action.Retrievedfrom
http://www.muslimadvocates.org/wpcontent/uploads/ClickHeretoEndHate.pdf
Nadler,A.,&Liviatan,I.(2006).Intergroupreconciliation:Effectsofadversary'sexpressions
ofempathy,responsibility,andrecipients'trust.PersonalityandSocialPsychology
Bulletin,32(4),459470.
Nyhan,B.,&Reifler,J.(2010).Whencorrectionsfail:Thepersistenceofpolitical
misperceptions.PoliticalBehavior,32(2),303330.
Nyhan,B.,&Reifler,J.(2015).Doescorrectingmythsaboutthefluvaccinework?An
experimentalevaluationoftheeffectsofcorrectiveinformation.Vaccine,33(3),
459464.
Oring,E.(1992).Jokesandtheirrelations.Lexington:UniversityPressofKentucky.
Payne,S.(2015,January20).AnamazingwomanfieldsatrollonMLKDayanditwas
nothingshortofinspirational.[Blog].Retrievedfrom
https://www.dailykos.com/story/2015/1/20/1359055/Anamazingwomanfeedsatro
llonMLKDayanditwasnothingshortofinspirational

38

Prior,M.(2003).Liberatedviewers,polarizedvoters:Theimplicationsofincreasedmedia
choicefordemocraticpolitics.TheGoodSociety,11,1016.
Ronson,J.(2015,February12).HowonestupidtweetblewupJustineSaccoslife.TheNew
YorkTimes.Retrievedfrom
http://www.nytimes.com/2015/02/15/magazine/howonestupidtweetruinedjustines
accoslife.html
Rosenberg,T.(2011).Jointheclub:Howpeerpressurecantransformtheworld.NewYork,
NY:WWNorton&Company.
Rutkowski,G.K.,Gruder,C.L.,&Romer,D.(1983).Groupcohesiveness,socialnorms,and
bystanderintervention.JournalofPersonalityandSocialPsychology,44(3),545552.
Saleem,H.M.,Dillon,K.P.,Benesch,S.,Ruths,D.(2016).Awebofhate:Tacklinghateful
speechinonlinesocialspaces.EuropeanLanguageResourcesAssociation.
Saltman,E.M.,&Russell,J.(2014).TheroleofPreventincounteringonlineextremism.The
QuilliamFoundation.
Simeone,M.(2014,January29).Twitteroutrage,charted:Thepartialanatomyofthe
#FuckPhyllistrend,orwhyIdonttrustBuzzFeed.[Blog].Retrievedfrom
https://suffenus.wordpress.com/2014/01/29/twitteroutragechartedthepartialanatom
yofthefuckphyllistrendandwhyidonttrustbuzzfeed/
Stevens,M.(2013,May13).FAQ:Thegeographyofhate.[Blog].Retrievedfrom
http://www.floatingsheep.org/2013/05/faqgeographyofhate.html
Stevens,M.(2013,May13).Thegeographyofhate.[Blog].Retrievedfrom
http://www.floatingsheep.org/2013/05/hatemap.html
Street,R.L.,&Giles.H.(1982).Speechaccommodationtheory:Asocialcognitiveapproach
tolanguageandspeechbehavior.InM.Roloff&C.Berger(Eds.),Socialcognition
andcommunication(p.205255).BeverlyHills,CA:Sage.
Suler,J.(2004).Theonlinedisinhibitioneffect.Cyberpsychology&Behavior,7(3),321326.
Tajfel,H.(1974).Socialidentityandintergroupbehaviour.SocialScienceInformation,13,
6593.
Tanis,M.,&Postmes,T.(2003).SocialcuesandimpressionformationinCMC.Journalof
Communication,53(4),676693.
Walid,D.(2013,November24).Responsestomycallingoutthetermabeed.Retrieved
from
https://dawudwalid.wordpress.com/2013/11/24/responsestomycallingouttheterm
abeed/

39

Warner,W.&Hirschberg,J.(2012).Detectinghatespeechontheworldwideweb.In
ProceedingsoftheSecondWorkshoponLanguageinSocialMedia,1926.
AssociationforComputationalLinguistics.
Xiang,G.,Fan,B.,Wang,L.,Hong,J.,&Rose,C.(2012).Detectingoffensivetweetsvia
topicalfeaturediscoveryoveralargescaletwittercorpus.InProceedingsofthe21st
ACMinternationalconferenceonInformationandknowledgemanagement,
19801984.ACM.
Zuckerman,E.(2013).Digitalcosmopolitans:WhywethinktheInternetconnectsus,whyit
doesn't,andhowtorewireit.WWNorton&Company.

Anda mungkin juga menyukai